Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this study, Jeong and Choi examine neural correlates of behavior during a naturalistic foraging task in which rats must dynamically balance resource acquisition (foraging) with the risk of threat. Rats first learn to forage for sucrose reward from a spout, and when a threat is introduced (an attack-like movement from a "LobsterBot"), they adjust their behavior to continue foraging while balancing exposure to the threat, adopting anticipatory withdraw behaviors to avoid encounter with the LobsterBot. Using electrode recordings targeting the medial prefrontal cortex (PFC), they identify heterogenous encoding of task variables across prelimbic and infralimbic cortex neurons, including correlates of distance to the reward/threat zone and correlates of both anticipatory and reactionary avoidance behavior. Based on analysis of population responses, they show that prefrontal cortex switches between different regimes of population activity to process spatial information or behavioral responses to threat in a context-dependent manner. Characterization of the heterogenous coding scheme by which frontal cortex represents information in different goal states is an important contribution to our understanding of brain mechanisms underlying flexible behavior in ecological settings.

Strengths:

As many behavioral neuroscience studies employ highly controlled task designs, relatively less is generally known about how the brain organizes navigation and behavioral selection in naturalistic settings, where environment states and goals are more fluid. Here, the authors take advantage of a natural challenge faced by many animals - how to forage for resources in an unpredictable environment - to investigate neural correlates of behavior when goal states are dynamic. Related to his, they also investigate prefrontal cortex (PFC) activity is structured to support different functional "modes" (here, between a navigational mode and a threat-sensitive foraging mode) for flexible behavior. Overall, an important strength and real value of this study is the design of the behavioral experiment, which is trial-structured, permitting strong statistical methods for neural data analysis, yet still rich enough to encourage natural behavior structured by the animal's volitional goals. The experiment is also phased to measure behavioral changes as animals first encounter a threat, and then learn to adapt their foraging strategy to its presence. Characterization of this adaptation process is itself quite interesting and sets a foundation for further study of threat learning and risk management in the foraging context. Finally, the characterization of single-neuron and population dynamics in PFC in this naturalistic setting with fluid goal states is an important contribution to the field. Previous studies have identified neural correlates of spatial and behavioral variables in frontal cortex, but how these representations are structured, or how they are dynamically adjusted when animals shift their goals, has been less clear. The authors synthesize their main conclusions into a conceptual model for how PFC activity can support mode switching, which can be tested in future studies with other task designed and functional manipulations.

Weaknesses:

While the task design in this study is intentionally stimulus-rich and places minimal constraint on the animal to preserve naturalistic behavior, this also introduces confounds that limit interpretability of the neural analysis. For example, some variables which are the target of neural correlation analysis, such as spatial/proximity coding and coding of threat and threat-related behaviors, are naturally entwined. To their credit, the authors have included careful analyses and control conditions to disambiguate these variables and significantly improve clarity.

The authors also claim that the heterogenous coding of spatial and behavioral variables in PFC is structured in a particular way that depends on the animal's goals or context. As the authors themselves discuss, the different "zones" contain distinct behaviors and stimuli, and since some neurons are modulated by these events (e.g., licking sucrose water, withdrawing from the LobsterBot, etc.), differences in population activity may to some extent reflect behavior/event coding. The authors have included a control analysis, removing timepoints corresponding to salient events, to substantiate the claim that PFC neurons switch between different coding "modes." While this significantly strengthens evidence for their conclusion, this analysis still depends on relatively coarse labeling of only very salient events. Future experiment designs, which intentionally separate task contexts (e.g. navigation vs. foraging), could serve to further clarify the structure of coding across contexts and/or goal states.

Finally, while the study includes many careful, in-depth neural and behavioral analyses to support the notion that modal coding of task variables in PFC may play a role in organizing flexible, dynamic behavior, the study still lacks functional manipulations to establish any form of causality. This limitation is acknowledged in the text, and the report is careful not to over interpret suggestions of causal contribution, instead setting a foundation for future investigations.

Thank you for the positive comment. We also acknowledge the inherent drawbacks of studying naturalistic behavior. As you also mentioned in the second round of review, separating navigation and foraging tasks in a larger apparatus, such as the one illustrated below, could better distinguish neural activity patterns associated with these different task types. To address the limitations of the current study, we have revised the report to avoid overinterpretation or unwarranted assumptions, and we appreciate that you have recognized this effort.

Author response image 1.

Reviewer #2 (Public review):

Summary:

Jeong & Choi (2023) use a semi-naturalistic paradigm to tackle the question of how the activity of neurons in the mPFC might continuously encode different functions. They offer two possibilities: either there are separate dedicated populations encoding each function, or cells alter their activity dependent on the current goal of the animal. In a threat-avoidance task rats procurred sucrose in an area of a chamber where, after remaining there for some amount of time, a 'Lobsterbot' robot attacked. In order to initiate the next trial rats had to move through the arena to another area before returning to the robot encounter zone. Therefore the task has two key components: threat avoidance and navigating through space. Recordings in the IL and PL of the mPFC revealed encoding that depended on what stage of the task the animal was currently engaged in. When animals were navigating, neuronal ensembles in these regions encoded distance from the threat. However, whilst animals were directly engaged with the threat and simultaneously consuming reward, it was possible to decode from a subset of the population whether animals would evade the threat. Therefore the authors claim that neurons in the mPFC switched between two functional modes: representing allocentric spatial information, and representing egocentric information pertaining to the reward and threat. Finally, the authors propose a conceptual model based on these data whereby this switching of population encoding is driven by either bottom-up sensory information or top-down arbitration.

Strengths:

Whilst these multiple functions of activity in the mPFC have generally been observed in tasks dedicated to the study of a singular function, less work has been done in contexts where animals continuously switch between different modes of behaviour in a more natural way. Being able to assess whether previous findings of mPFC function apply in natural contexts is very valuable to the field, even outside of those interested in the mPFC directly. This also speaks to the novelty of the work; although mixed selectivity encoding of threat assessment and action selection has been demonstrated in some contexts (e.g. Grunfeld & Likhtik, 2018) understanding the way in which encoding changes on-the-fly in a self-paced task is valuable both for verifying whether current understanding holds true and for extending our models of functional coding in the mPFC.

The authors are also generally thoughtful in their analyses and use a variety of approaches to probe the information encoded in the recorded activity. In particular, they use relatively close analysis of behaviour as well as manipulating the task itself by removing the threat to verify their own results. The use of such a rich task also allows them to draw comparisons, e.g. in different zones of the arena or different types of responses to threat, that a more reduced task would not otherwise allow. Additional in-depth analyses in the updated version of the manuscript, particularly the feature importance analysis, as well as complimentary null findings (a lack of cohesive place cell encoding, and no difference in location coding dependent on direction of trajectory) further support the authors' conclusion that populations of cells in the mPFC are switching their functional coding based on task context rather than behaviour per se. Finally, the authors' updated model schematic proposes an intriguing and testable implementation of how this encoding switch may be manifested by looking at differentiable inputs to these populations.

Weaknesses:

The main existing weakness of this study is that its findings are correlational (as the authors highlight in the discussion). Future work might aim to verify and expand the authors' findings - for example, whether the elevated response of Type 2 neurons directly contributes to the decision-making process or just represents fear/anxiety motivation/threat level - through direct physiological manipulation. However, I appreciate the challenges of interpreting data even in the presence of such manipulations and some of the additional analyses of behaviour, for example the stability of animals' inter-lick intervals in the E-zone, go some way towards ruling out alternative behavioural explanations. Yet the most ideal version of this analysis is to use a pose estimation method such as DeepLabCut to more fully measure behavioural changes. This, in combination with direct physiological manipulation, would allow the authors to fully validate that the switching of encoding by this population of neurons in the mPFC has the functional attributes as claimed here.

I wanted to add a minor comment about interpreting the two possible accounts presented in fig. 8 to suggest a third possibility: that both bottom-up sensory and top-down arbitration mechanisms can occur simultaneously to influence whether the activity of the population switches. Indeed, a model where these inputs are balanced or pitted against each other, so to speak, to continuously modulate encoding in the mPFC seems both adaptive and likely. Further, some speculation on the source of the 'arbitrator' in the top-down account would make this model more tractable for future testing of its validity.

We thank the reviewer for highlighting this important perspective. We fully agree that an intricate and recurrent interaction between bottom-up and top-down modulations is a highly plausible account of how the mPFC changes its encoding mode. In line with this suggestion, we have incorporated this idea as a third possibility in the revised Discussion, alongside an updated version of Figure 8 that explicitly illustrates this competitive model.

Although we were unable to identify a definitive study directly measuring how the mPFC switches encoding modes across tasks, we did find relevant human EEG and fMRI studies addressing this issue. Based on these findings, we now propose the anterior cingulate cortex (ACC) as a potential hub for top-down arbitration. We have added a paragraph in the Discussion describing this possibility and its implications for future testing.

“Which brain region might act as this arbitrator? Evidence from human neuroimaging studies implicates the anterior cingulate cortex (ACC) as a central hub for switching cognitive modes. During task switching, the ACC shows increased activation (Hyafil et al., 2009), enhances connectivity with task-specific regions (Aben et al., 2020), correlates with multitask performance (Kondo et al., 2004), and monitors the reliability of competing decision systems (Lee et al., 2014). Collectively, these findings point to a pivotal role for the ACC in coordinating task assignment. Rodent studies also link the ACC to strategic mode switching (Tervo et al., 2014), suggesting that the rodent ACC could similarly arbitrate between strategies, determining which task-relevant variables are represented in the ventral mPFC, including the PL and IL. Future studies combining multi-context tasks with causal manipulations will be essential to determine whether these functional shifts are driven primarily by top-down arbitration or by bottom-up sensory inputs.”

Reviewer #3 (Public review):

Summary:

This study investigates how various behavioral features are represented in the medial prefrontal cortex (mPFC) of rats engaged in a naturalistic foraging task. The authors recorded electrophysiological responses of individual neurons as animals transitioned between navigation, reward consumption, avoidance, and escape behaviors. Employing a range of computational and statistical methods, including artificial neural networks, dimensionality reduction, hierarchical clustering, and Bayesian classifiers, the authors sought to predict from neural activity distinct task variables (such as distance from the reward zone and the success or failure of avoidance behavior). The findings suggest that mPFC neurons alternate between at least two distinct functional modes, namely spatial encoding and threat evaluation, contingent on the specific location.

Strengths:

This study attempt to address an important question: understanding the role of mPFC across multiple dynamic behaviors. The authors highlight the diverse roles attributed to mPFC in previous literature and seek to explain this apparent heterogeneity. They designed an ethologically relevant foraging task that facilitated the examination of complex dynamic behavior, collecting comprehensive behavioral and neural data. The analyses conducted are both sound and rigorous.

Weaknesses:

Because the study still lacks experimental manipulation, the findings remain correlational. The authors have appropriately tempered their claims regarding the functional role of the mPFC in the task. The nature of the switch between functional modes encoding distinct task variables (i.e., distance to reward, and threat-avoidance behavior type) is not established. Moreover, the evidence presented to dissociate movement from these task variables is not fully convincing, particularly without single-session video analysis of movement. Specifically, while the new analyses in Figure 7 are informative, they may not fully account for all potential confounding variables arising from changes in context or behavior.

Regarding the claim of highly stereotyped behavior, there are some inconsistencies. While the authors assert this, Figure 1F shows inter-animal variability, and the PETHs, representing averaged activity, may not fully capture the variability of the behavior across sessions and animals. To strengthen this aspect, a more detailed analysis that examines the relationship between behavior and neural activity on a trial-by-trial basis, or at minimum, per session, could help.

We thank the reviewer for this thoughtful recommendation and the opportunity to clarify our use of the term “stereotyped behavior.” By this, we were specifically referring to the animals’ consistent licking behavior in the E-zone, rather than to the latency of head withdrawal, which indeed varied across trials and animals. Because licking tempo and body posture during sucrose consumption were highly consistent, the decision to avoid or stay (AW vs. EW) could not be predicted from overt behavior alone. This consistency strengthens our conclusion that the significant predictive power of the Bayesian decoding analysis reflects intrinsic firing patterns of the mPFC neural network, rather than simple behavioral correlates of avoidance.

We also note that the Bayesian model was conducted on a trial-by-trial basis, and the reported prediction accuracy of 73% represents the average across all individual trials (Figure 6B, C). Thus, the analysis inherently captures variability across trials and animals, directly addressing the reviewer’s concern.

The reviewer is correct that the PETHs shown in Figure 5 are based on session-averaged activity aligned to head-entry and head-withdrawal events. The purpose of this analysis was to illustrate that certain modulation patterns could be grouped into 2–3 distinct categories. While averaged activity can provide insight into collective responses to external events, we agree that trial-based analyses provide a more rigorous demonstration of the link between neural ensemble activity and behavioral decisions. This is precisely why we complemented the PETH analysis with Bayesian decoding, which provides stronger evidence that mPFC ensemble activity is predictive of the animal’s choice to avoid or stay.

Similarly, the claim regarding the limited scope of extraneous behavior (beyond licking) requires further substantiation. It would be more convincing to quantify potential variations in licking vigor and to provide evidence for the absence of significant postural changes.

To address this concern, we quantified licking vigor using the inter-lick interval (ILI) as an indirect index. A lick was defined as the period from tongue contact with the IR beam (Lick-On) to withdrawal (Lick-Off), and the ILI was calculated as the time between a Lick-Off and the subsequent Lick-On. Across all animals, ILIs were clustered within a narrow range with a median of 0.155 s (see Author response image 4, left panel).

We analyzed licking vigor at two levels: within trials and within sessions. Because reduced vigor or satiation would lengthen ILIs, comparing the first half and the last half of ILIs within a trial or within a session provides a sensitive proxy for licking consistency.

Within trials: For each of 2,820 trials, we compared the mean ILI of the first half of licks to that of the second half. The average difference was only ~ 17 ms (middle panel). Across sessions: Trial-averaged ILIs were compared between the first and last halves of each session, yielding a mean difference ~ 1.7 ms per session (right panel).

These analyses demonstrate that rats maintained stable licking vigor whenever they entered the E-zone, regardless of avoidance outcome.

Author response image 2.

Concerning the ANN model, while I understand the choice of a 4-layer network for its performance, the study could have benefited from exploring simpler models. A model where weight corresponds directly to individual neurons could improve interpretability and facilitate the investigation of dynamic changes in neuronal 'modes' (i.e., weight adjustments) over time.

We fully agree with the reviewer on the importance of biologically interpretable models. While artificial neural networks (ANNs) share certain similarities with neural computation, they are not intended to capture biological realism. For example, the error correction mechanism used in ANNs, such as backpropagation has no direct counterpart in mammalian neural circuits. Although we considered approaches that would link each computational node more directly to the activity of individual neurons, building such a model would require temporally sensitive, mechanistic frameworks (e.g., leaky integrate-and-fire networks) and an extensive behavioral alignment effort, which is beyond the scope of the current study.

Our use of an ANN was intended solely as an analytical tool to uncover hidden patterns in multi-unit activity that may not be detectable with traditional methods. Among various machine-learning algorithms, we selected a four-layer ANN regressor because it achieved significantly lower decoding errors (Supplementary Figure S3) and showed robustness to hyperparameter variation (Glaser et al., 2020). To acknowledge the limitations of this approach and suggest future directions, we have revised the Results section to explicitly discuss these points.

“Among various machine learning algorithms, we selected a robust tool for decoding underlying patterns in the data, rather than to model the architecture of the mPFC. We implemented a four-layer artificial neural network regressor (ANN; see Materials and Methods for a detailed structure), as the ANN achieves significantly lower decoding errors (Supplementary Figure S3) and has robustness to hyperparameter changes (Glaser et al., 2020).”

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public review):

In the Late Triassic and Early Jurassic (around 230 to 180 Ma ago), southern Wales and adjacent parts of England were a karst landscape. The caves and crevices accumulated remains of small vertebrates. These fossil-rich fissure fills are being exposed in limestone quarrying. In 2022 (reference 13 of the article), a partial articulated skeleton and numerous isolated bones from one fissure fill of end-Triassic age (just over 200 Ma) were named Cryptovaranoides microlanius and described as the oldest known squamate - the oldest known animal, by some 20 to 30 Ma, that is more closely related to snakes and some extant lizards than to other extant lizards. This would have considerable consequences for our understanding of the evolution of squamates and their closest relatives, especially for their speed and absolute timing, and was supported in the same paper by phylogenetic analyses based on different datasets.

In 2023, the present authors published a rebuttal (reference 18) to the 2022 paper, challenging anatomical interpretations and the irreproducible referral of some of the isolated bones to Cryptovaranoides. Modifying the datasets accordingly, they found Cryptovaranoides outside Squamata and presented evidence that it is far outside. In 2024 (reference 19), the original authors defended most of their original interpretation and presented some new data, some of it from newly referred isolated bones. The present article discusses anatomical features and the referral of isolated bones in more detail, documents some clear misinterpretations, argues against the widespread but not justifiable practice of referring isolated bones to the same species as long as there is merely no known evidence to the contrary, further argues against comparing newly recognized fossils to lists of diagnostic characters from the literature as opposed to performing phylogenetic analyses and interpreting the results, and finds Cryptovaranoides outside Squamata again.

Although a few of the character discussions and the discussion of at least one of the isolated bones can probably still be improved (and two characters are addressed twice), I see no sign that the discussion is going in circles or otherwise becoming unproductive. I can even imagine that the present contribution will end it.

We appreciate the positive response from reviewer 1!

Reviewer #2 (Public review):

Congratulations on this thorough manuscript on the phylogenetic affinities of Cryptovaranoides.

Thank you.

Recent interpretations of this taxon, and perhaps some others, have greatly changed the field's understanding of reptile origins- for better and (likely) for worse.

We agree, and note that while it is possible for challenges to be worse than the original interpretations, both the original and subsequent challenges are essential aspects of what make science, science.

This manuscript offers a careful review of the features used to place Cryptovaranoides within Squamata and adequately demonstrates that this interpretation is misguided, and therefore reconciles morphological and molecular data, which is an important contribution to the field of paleontology. The presence of any crown squamate in the Permian or Triassic should be met with skepticism, the same sort of skepticism provided in this manuscript.

We agree and add that every testable hypothesis requires skepticism and testing.

I have outlined some comments addressing some weaknesses that I believe will further elevate the scientific quality of the work. A brief, fresh read‑through to refine a few phrases, particularly where the discussion references Whiteside et al. could also give the paper an even more collegial tone.

We have followed Reviewer 2’s recommendations closely (see below) and have justified in our responses if we do not fully follow a particular recommendation.

This manuscript can be largely improved by additional discussion and figures, where applicable. When I first read this manuscript, I was a bit surprised at how little discussion there was concerning both non-lepidosauromorph lepidosaurs as well as stem-reptiles more broadly. This paper makes it extremely clear that Cryptovaranoides is not a squamate, but would greatly benefit in explaining why many of the characters either suggested by former studies to be squamate in nature or were optimized as such in phylogenetic analyses are rather widespread plesiomorphies present in crownward sauropsids such as millerettids, younginids, or tangasaurids. I suggest citing this work where applicable and building some of the discussion for a greatly improved manuscript. In sum:

(1) The discussion of stem-reptiles should be improved. Nearly all of the supposed squamate features in Cryptovaranoides are present in various stem-reptile groups. I've noted a few, but this would be a fairly quick addition to this work. If this manuscript incorporates this advice, I believe arguments regarding the affinities of Cryptovaranoides (at least within Squamata) will be finished, and this manuscript will be better off for it.

(2) I was also surprised at how little discussion there was here of putative stem-squamates or lepidosauromorphs more broadly. A few targeted comparisons could really benefit the manuscript. It is currently unclear as to why Cryptovaranoides could not be a stem-lepidosaur, although I know that the lepidosaur total-group in these manuscripts lacks character sampling due to their scarcity.

We are responding to (1) and (2) together. We agree with the Reviewer that a thorough comparison of Cryptovaranoides to non-lepidosaurian reptiles is critical. This is precisely what we did in our previous study: Brownstein et al. (2023)— see main text and supplementary information therein. As addressed therein, there is a substantial convergence between early lepidosaurs and some groups of archosauromorphs (our inferred position for Cryptovaranoides). Many of those points are not addressed in detail here in order to avoid redundancy and are simply referenced back to Brownstein et al. (2023). Secondly, stem reptiles (i.e., non-lepidosauromorphs and non-archosauromorphs), such as suggested above (millerettids, younginids, or tangasaurids), are substantially more distantly related to Cryptovaranoides (following any of the published hypotheses). As such, they share fewer traits (either symplesiomorphies or homoplasies), and so, in our opinion, we would risk directing losing the squamate-focus of our study.

We thus respectfully decline to engage the full scope of the problem in this contribution, but do note that this level of detailed work would make for an excellent student dissertation research program.

(3) This manuscript can be improved by additional figures, such as the slice data of the humerus. The poor quality of the scan data for Cryptovaranoides is stated during this paper several times, yet the scan data is often used as evidence for the presence or absence of often minute features without discussion, leaving doubts as to what condition is true. Otherwise, several sections can be rephrased to acknowledge uncertainty, and probably change some character scorings to '?' in other studies.

We strongly agree with the reviewer. Unfortunately, the original publication (Whiteside et al., 2021) did not make available the raw CT scan data to make this possible. As noted below in the Responses to Recommendations Section, we only have access to the mesh files for each segmented element. While one of us has observed the specimens personally, we have not had the opportunity to CT scan the specimens ourselves.

Reviewer #3 (Public review):

Summary:

The study provides an interesting contribution to our understanding of Cryptovaranoides relationships, which is a matter of intensive debate among researchers. My main concerns are in regard to the wording of some statements, but generally, the discussion and data are well prepared. I would recommend moderate revisions.

Strengths:

(1) Detailed analysis of the discussed characters.

(2) Illustrations of some comparative materials.

Thank you for noting the strengths inherent to our study.

Weaknesses:

Some parts of the manuscript require clarification and rewording.

One of the main points of criticism of Whiteside et al. is using characters for phylogenetic considerations that are not included in the phylogenetic analyses therein. The authors call it a "non-trivial substantive methodological flaw" (page 19, line 531). I would step down from such a statement for the reasons listed below:

(1) Comparative anatomy is not about making phylogenetic analyses. Comparative anatomy is about comparing different taxa in search of characters that are unique and characters that are shared between taxa. This creates an opportunity to assess the level of similarity between the taxa and create preliminary hypotheses about homology. Therefore, comparative anatomy can provide some phylogenetic inferences.

That does not mean that tests of congruence are not needed. Such comparisons are the first step that allows creating phylogenetic matrices for analysis, which is the next step of phylogenetic inference. That does not mean that all the papers with new morphological comparisons should end with a new or expanded phylogenetic matrix. Instead, such papers serve as a rationale for future papers that focus on building phylogenetic matrices.

We agree completely. We would also add that not every study presenting comparative anatomical work need be concluded with a phylogenetic analysis.

Our criticism of Whiteside et al. (2022) and (2024) is that these studies provided many unsubstantiated claims of having recovered synapomorphies between Cryptovaranoides and crown squamates without actually having done so through the standard empirical means (i.e., phylogenetic analysis and ancestral state reconstruction). Both Whiteside et al. (2022) and (2024) indicate characters presented as ‘shared with squamates’ along with 10 characters presented as synapomorphies (10). However, their actual phylogenetically recovered synapomorphies were few in number (only 3) and these were not discussed.

Furthermore, Whiteside et al. (2022) and (2024) comparative anatomy was restricted to comparing †Cryptovaranoides to crown squamates., based on the assumption that †Cryptovaranoides was a crown squamate and thus only needed to be compared to crown squamates.

In conclusion, we respectfully, we maintain such efforts are “non-trivial substantive methodological flaw(s)”.

(2) Phylogenetic matrices are never complete, both in terms of morphological disparity and taxonomic diversity. I don't know if it is even possible to have a complete one, but at least we can say that we are far from that. Criticising a work that did not include all the possibly relevant characters in the phylogenetic analysis is simply unfair. The authors should know that creating/expanding a phylogenetic matrix is a never-ending work, beyond the scope of any paper presenting a new fossil.

Respectfully, we did not criticize previous studies for including an incomplete phylogeny. Instead, we criticized the methodology behind the homology statements made in Whiteside et al. (2022) and Whiteside et al. (2024).

(3) Each additional taxon has the possibility of inducing a rethinking of characters. That includes new characters, new character states, character state reordering, etc. As I said above, it is usually beyond the scope of a paper with a new fossil to accommodate that into the phylogenetic matrix, as it requires not only scoring the newly described taxon but also many that are already scored. Since the digitalization of fossils is still rare, it requires a lot of collection visits that are costly in terms of time.

We agree on all points, but we are unsure of what the Reviewer is asking us to do relative to this study.

(4) If I were to search for a true flaw in the Whiteside et al. paper, I would check if there is a confirmation bias. The mentioned paper should not only search for characters that support Cryptovaranoides affinities with Anguimorpha but also characters that deny that. I am not sure if Whiteside et al. did such an exercise. Anyway, the test of congruence would not solve this issue because by adding only characters that support one hypothesis, we are biasing the results of such a test.

We would refer the Reviewer to their section (1) on comparative anatomy. As we and the Reviewer have pointed out, Whiteside et al. did not perform comparative anatomical statements outside of crown Squamata in their original study. More specifically, Whiteside et al. (2022, Fig. 8) presented a phylogeny where Cryptovaranoides formed a clade with Xenosaurus within the crown of Anguimorpha or what they termed “Anguiformes”, and made comparisons to the anatomies of the legless anguids, Pseudopus and Ophisaurus. Whiteside et al. (2024), abandoned “Anguiformes”, maintained comparisons to Pseudopus and emphasized affinities with Anguimorpha (but almost all of their phylogenies as published, they do not recover a monophyletic Angumimorpha unless amphisbaenians and snakes are considered to be anguimorphans. Thus, we agree that confirmation bias was inherent in their studies.

To sum up, there is nothing wrong with proposing some hypotheses about character homology between different taxa that can be tested in future papers that will include a test of congruence. Lack of such a test makes the whole argumentation weaker in Whiteside et al., but not unacceptable, as the manuscript might suggest. My advice is to step down from such strong statements like "methodological flaw" and "empirical problems" and replace them with "limitations", which I think better describes the situation.

We agree with the first sentence in this paragraph – there is nothing wrong with proposing character homologies between different taxa based on comparative anatomical studies. However, that is not what Whiteside et al. (2022) and (2024) did. Instead, they claimed that an ad hoc comparison of Cryptovaranoides to crown Squamata confirmed that Cryptovaranoides is in fact a crown squamate and likely a member of Anguimorpha. Their study did not recognize limitations, but rather, concluded that their new taxon pushed the age of crown Squamata into the Triassic.

As noted by Reviewer 2, such a claim, and the ‘data’ upon which it is based, should be treated with skepticism. We have elected to apply strong skepticism and stringent tests of falsification to our critique.

Reviewer #1 (Recommendations for the authors):

(1) Lines 596-598 promise the following: "we provide a long[-]form review of these and other features in Cryptovaranoides that compare favorably with non-squamate reptiles in Supplementary Material." You have kindly informed me that all this material has been moved into the main text; please amend this passage.

This has been deleted.

(2) Comments on science

41: I would rather say "an additional role".

This has been edited accordingly.

43: Reconstructing the tree entirely from extant organisms and adding fossils later is how Hennig imagined it, because he was an entomologist, and fossil insects are, on average,e extremely rare and usually very incomplete (showing a body outline and/or wing venation and little or nothing else). He was wrong, indeed wrong-headed. As a historical matter, phylogenetic hypotheses were routinely built on fossils by the mid-1860s, pretty much as soon as the paleontologists had finished reading On the Origin of Species, and this practice has never declined, let alone been interrupted. As a theoretical matter, including as many extinct taxa as possible in a phylogenetic analysis is desirable because it breaks up long branches (as most recently and dramatically shown by Mongiardino Koch & Parry 2020), and while some methods and some kinds of data are less susceptible to long-branch attraction and long-branch repulsion than others, none are immune; and while missing data (on average more common in fossils) can actively mislead parametric methods, this is not the case with parsimony, and even in Bayesian inference the problem is characters with missing data, not taxa with missing data. Some of you have, moreover, published tip-dated phylogenetic analyses. As a practical matter, molecular data are almost never available from fossils, so it is, of course, true that analyses which only use molecular data can almost never include fossils; but in the very rare exceptions, there is no reason to treat fossil evidence as an afterthought.

We agree and have changed “have become” to “is.”

49-50, 59: The ages of individual fissure fills can be determined by biostratigraphy; as far as I understand, all specimens ever referred to Cryptovaranoides [13, 19] come from a single fill that is "Rhaetian, probably late Rhaetian (equivalent of Cotham Member, Lilstock Formation)" [13: pp. 2, 15].

We appreciate this comment; the recent literature, however, suggests that variable ages are implied by the biostratigraphy at the English Fissure Fills, so we have chosen to keep this as is. Also note that several isolated bones were not recovered with the holotype but were discussed by Whiteside et al. (2024). The provenance of these bones was not clearly discussed in that paper.

59-60: Why "putative"? Just to express your disagreement? I would do that in a less misleading way, for example: "and found this taxon as a crown-group squamate (squamate hereafter) in their phylogenetic analyses." - plural because [19] presented four different analyses of two matrices just in the main paper.

We have removed this word.

121-124: The entepicondylar foramen is homologous all the way down the tree to Eusthenopteron and beyond. It has been lost a quite small number of times. The ectepicondylar foramen - i.e., the "supinator" (brachioradialis) process growing distally to meet the ectepicondyle, fusing with it and thereby enclosing the foramen - goes a bit beyond Neodiapsida and also occurs in a few other amniote clades (...as well as, funnily enough, Eusthenopteron in later ontogeny, but that's independent).

We agree. However, the important note here is that the features on the humerus of Cryptovaranoides are not comparable (differ in location and morphology) to the ent- and ectepondylar foramina in other reptiles, as we discuss at length. As such, we have kept this sentence as is.

153: Yes, but you [18] mistakenly wrote "strong anterior emargination of the maxillary nasal process, which is [...] a hallmark feature of archosauromorphs" in the main text (p. 14) - and you make the same mistake again here in lines 200-206! Also, the fact [19: Figure 2a-c] remains that Cryptovaranoides did not have an antorbital fenestra, let alone an antorbital fossa surrounding it (a fossa without a fenestra only occurs in some cases of secondary loss of the fenestra, e.g., in certain ornithischian dinosaurs). Unsurprisingly, therefore, Cryptovaranoides also does not have an orbital-as-opposed-to-nasal process on its maxilla [19: Figure 2a-c].

Line 243-249 (in original manuscript) deal with the emargination of maxillary nasal process (but this does not imply a full antorbital fenestra).  We explicitly state that this feature alone "has limited utility" for supporting archosauromorph affinity.

158-173: The problem here is not that the capitellum is not preserved; from amniotes and "microsaurs" to lissamphibians and temnospondyls, capitella ossify late, and larger capitella attach to proportionately larger concave surfaces, so there is nothing wrong with "the cavity in which it sat clearly indicates a substantial condyle in life". Instead, the problem is a lack of quantification (...as has also been the case in the use of the exact same character in the debate on the origin of lissamphibians); your following sentence (lines 173-175) stands. The rest of the paragraph should be drastically shortened.

We appreciate this comment. We note that the ontogenetic variation of this feature is in part the issue with the interpretation provided by Whiteside et al. (2024). The issue is the lack of consistency on the morphology of the capitellum in that study. We are unclear on what the reviewer means by ‘quantification,’ as the character in question is binary. 

250-252: It's not going to matter here, but in any different phylogenetic context, "sphenoid" would be confusing given the sphenethmoid, orbitosphenoid, pleurosphenoid, and laterosphenoid. I actually recommend "parabasisphenoid" as used in the literature on early amniotes (fusion of the dermal parasphenoid and the endochondral basisphenoid is standard for amniotes).

We have added "(=parabasisphenoid)" on first use but retain use of sphenoid because in the squamate and archosauromorph literature, sphenoid (or basisphenoid) is used more frequently.

314-315: Vomerine teeth are, of course, standard for sarcopterygians. Practically all extant amphibians have a vomerine toothrow, for example. A shagreen of denticles on the vomer is not as widespread but still reaches into the Devonian (Tulerpeton).

We agree, but vomerine teeth are rare in lepidosaurs and archosaurs and occur only in very recent clades e.g. anguids and one stem scincoid. Their presence in amphibians is not directly relevant to the phylogenetic placement of Cryptovaranoides among reptiles.

372: Fusion was not scored as present in [13], but as unknown (as "partial" uncertainty between states 0 and 1 [19:8]), and seemingly all three options were explored in [19].

We politely disagree with the reviewer; state 1 is scored in Whiteside et al. (2024).

377-383: Together with the partially fused NHMUK PV R37378 [13: Figure 4B, C; 19: 8], this is actually an argument that Cryptovaranoides is outside but close to Unidentata. The components of the astragalus fuse so early in extant amniotes that there is just a single ossification center in the already fused cartilage, but there are Carboniferous and Permian examples of astragali with sutures in the expected places; all of the animals in question (Diadectes, Hylonomus, captorhinids) seem to be close to but outside Amniota. (And yet, the astragalus has come undone in chamaeleons, indicating the components have not been lost.) - Also, if NHMUK PV R37378 doesn't belong to a squamate close to Unidentata, what does it belong to? Except in toothless beaks, premaxillary fusion is really rare; only molgin newts come to mind (and age, tooth size, and tooth number of NHMUK PV R37378 are wholly incompatible with a salamandrid).

The relevance of the astragalus is to the current discussion is unclear as we do not mention this element in our manuscript.  We discuss the fusion in the premaxillae in response to previous comment. 

471-474: That thing is concave. (The photo is good enough that you can enlarge it to 800% before it becomes too pixelated.) It could be a foramen filled with matrix; it does not look like a grain sticking to the outside of the bone. Also, spell out that you're talking about "suc.fo" in Figure 3j.

We are also a bit confused about this comment, as we state:

“Finally, we note here that Whiteside et al. [19] appear to have labeled a small piece of matrix attached to a coracoid that they refer to †C. microlanius as the supracoroacoid [sic] foramen in their figure 3, although this labeling is inferred because only “suc, supracoroacoid [sic]” is present in their figure 3 caption.” (L. 519-522, P. 17). We cannot verify that this structure is concave, as so we keep this text as is.

476-489: [19] conceded in their section 4.1 (pp. 11-12) that the atlas pleurocentrum, though fused to the dorsal surface of the axis intercentrum as usual for amniotes and diadectomorphs, was not fused to the axis pleurocentrum.

This is correct, as we note in the MS. The issue is whether these elements are clearly identifiable.

506-510: [19:12] did identify what they considered a possible ulnar patella, illustrated it (Figure 4d), scored it as unknown, and devoted the entire section 4.4 to it.<br /> 512-523: What I find most striking is that Whiteside et al., having just discovered a new taxon, feel so certain that this is the last one and any further material from that fissure must be referable to one of the species now known from there.

We agree with these points and believe we have devoted adequate text to addressing them. Note that the reviewer does not recommend any revisions to these sections.

553: Not that it matters, but I'm surprised you didn't use TNT 1.6; it came out in 2023 and is free like all earlier versions.

We have kept this as is following the reviewer comment, and because we were interested in replicating the analyses in the previous publications that have contributed to the debate about the identity of this taxon.  For the present simple analyses both versions should perform identically, as the search algorithms for discrete characters are identical across these versions.

562: Is "01" a typo, or do you mean "0 or 1"? In that case, rather write "0/1" or "{01}".

This has been corrected to {01}

(3) Comments on nomenclature and terminology

55, 56: Delete both "...".

This has been corrected.

100: "ent- and ectepicondylar"

For clarity, we have kept the full words.

107-108: I understand that "high" is proximal and "low" is distal, but what is "the distal surface" if it is not the articular surface in the elbow joint?

This has been corrected.

120: "stem pan-lepidosaurs, and stem pan-squamates"; Lepidosauria and Squamata are crown groups that don't contain their stems

This has been corrected.

122, 123: Italics for Claudiosaurus and Delorhynchus.

This has been corrected.

130: Insert a space before "Tianyusaurus" (it's there in the original), and I recommend de-italicizing the two genus names to keep the contrast (as you did in line 162).

This has been corrected.

130, 131: Replace both "..." by "[...]", though you can just delete the second one.

This has been corrected.

174: Not a capitulum, but a grammatically even smaller (double diminutive) capitellum.

This has been corrected.

209, 224, Table 1: Both teams have consistently been doing this wrong. It's "recessus scalae tympani". The scala tympani ("ladder/staircase of the [ear]drum") isn't the recess, it's what the recess is for; therefore, the recess is named "recess of the scala tympani", and because there was no word for "of" in Classical Latin ("de" meant "off" and "about"), the genitive case was the only option. (For the same reason, the term contains "tympani", the genitive of "tympanum".)

This has been corrected.

415-425: This is a terminological nightmare. Ribs can have (and I'm not sure this is exhaustive): a) two separate processes (capitulum, tuberculum) that each bear an articulating facet, and a notch in between; b) the same, but with a non-articulating web of bone connecting the processes; c) a single uninterrupted elongate (even angled) articulating facet that articulates with the sutured or fused dia- and parapophysis; d) a single round articulating facet. Certainly, a) is bicapitate and d) is unicapitate, but for b) and c) all bets are off as to how any particular researcher is going to call them. This is a known source of chaos in phylogenetic analyses. I recommend writing a sentence or three on how the terms "unicapitate" & "bicapitate" lack fixed meanings and have caused confusion throughout tetrapod phylogenetics, and that the condition seen in Cryptovaranoides is nonetheless identical to that in archosauromorphs.

This has been added: “This confusion in part stems from the lack of a fixed meaning for uni- and bicapitate rib heads; in any case, †C. microlanius possesses a condition identical to archosauromorphs as we have shown.”  (L.475-477, P.16).

439-440: Other than in archosaurs, some squamates and Mesosaurus, in which sauropsids are dorsal intercentra absent?

We are unclear about the relevance of the question to this section. The issue at hand is that some squamate lineages possess dorsal intercentra, so the absence of dorsal intercentra cannot be considered a squamate synapomorphy without the optimization of this feature along a phylogeny (which was not accomplished by Whiteside et al.).

458: prezygapophyses.

This has been corrected.

516: "[...]".

This has been corrected.

566: synapomorphies.

This has been corrected.

587: Macrocnemus.

This has been corrected.

585: I strongly recommend either taking off and nuking the name Reptilia from orbit (like Pisces) or using it the way it is defined in Phylonyms, namely as the crown group (a subset of Neodiapsida). Either would mean replacing "neodiapsid reptiles" with "neodiapsids".

This has been corrected to “neodiapsids.”

625: Replace "inclusive clades" by "included clades", "component clades", "subclades", or "parts," for example.

This has been kept as is because “inclusive clades” is common terminology and is used extensively in, for example, the PhyloCode. 

659: Please update.

References are updated.

Fig. 8: Typo in Puercosuchus.

This has been corrected.

(4) Comments on style and spelling

You inconsistently use the past and the present tense to describe [13, 19], sometimes both in the same sentence (e.g., lines 323 vs. 325). I recommend speaking of published papers in the past tense to avoid ascribing past views and acts to people in their present state.

This has been corrected to be more consistent throughout the manuscript.

48: Remove the second comma.

This has been corrected.

91: Replace "[13] and WEA24" by "[13, 19]".

This has been corrected.

100: Commas on both sides of "in fact" or on neither

This has been corrected.

117: I recommend "the interpretation in [19]". I have nothing against the abbreviation "WEA24", but you haven't defined it, and it seems like a remnant of incomplete editing. - That said, eLife does not impose a format on such things. If you prefer, you can just bring citation by author & year back; in that case, this kind of abbreviation would make perfect sense (though it should still be explicitly defined).<br /> 129, 145: Likewise.

We have modified this [13] and [19] where necessary.

192-198: Surely this should be made part of the paragraph in lines 158-175, which has the exact same headline?

This has been corrected.

200-206: Surely this should be made part of the paragraph in lines 148-156, which has the exact same headline?

These sections deal with different issues pertaining to the analyses of Whiteside et al. (2024) and so we have kept to organization as is.

214: Delete "that".

This has been deleted.

312: "Vomer" isn't an adjective; I'd write "main vomer body" or "vomer's main body" or "main body of the vomer".

This has been corrected.

350: "figured"

This has been corrected.

400: Rather, "rearticulated" or "worked to rearticulate"? - And why "several"? Just write "two". "Several" implies larger numbers.

These issues have been corrected.

448, 500: As which? As what kind of feature? I'm aware that "as such" is fairly widely used for "therefore", but it still confuses me every time, and I have to suspect I'm not the only one. I recommend "therefore" or "for this reason" if that is what you mean.

“As such” has been deleted.

452: Adobe Reader doesn't let me check, but I think you have two spaces after "of".

This has been corrected.

514, 539, 546, 552, 588, Fig. 3, 5, 6, Table 1: "WEA24" strikes again.

This has been corrected.

515: Remove the parentheses.

This has been corrected.

531: Insert a space after the period.

This has been corrected.

532: Remove both commas and the second "that".

This has been corrected.

538: Remove the comma.

This has been kept as is because changing it would render the sentence grammatically incorrect.

545: "[...]" or, better, nothing.

This has been corrected.

547: Spaces on both sides of the dash or on neither (as in line 553).

This has been corrected.

552: Rather, "conducted a parsimony analysis".

This has been corrected.

556: Space after "[19]".

This has been corrected.

560: Comma after "narrow".

This has been corrected.

600: Comma after "above" to match the one in the preceding line - there's an insertion in the sentence that must be flanked by commas on both sides.

This has been corrected.

603: Compound adjectives like "alpha-taxonomic" need a hyphen to avoid tripping readers up.

This has been corrected.

612: Similarly, "ancestral-state reconstruction" needs one to make immediately clear it isn't a state reconstruction that is ancestral but a reconstruction of ancestral states.

This has been corrected.

613: If you want to keep this comma, you need to match it with another after "Cryptovaranoides" in line 611.

We have kept this as is, because removing this comma would render the sentence grammatically incorrect.

615: Likewise, you need a comma after "and" because "except for a few features" is an insertion. The other comma is actually optional; it depends on how much emphasis you want to place on what comes after it.

this has been added.

622: Comma after "[48, 49]".

this has been added.

672: Missing italics and two missing spaces.

This has been corrected.

678, 680-681, 693, 700-701, 734, 742, 747, 788, 797, 799, 803, 808, 810-811, 814, 817, 820, 823, 828, 841, 843: Missing italics.

This has been corrected.

683, 689: These are book chapters. Cite them accordingly.

This has been corrected.

737: Missing DOI.

No DOI is available.

793: Missing Bolosaurus major; and I'd rather cite it as "2024" than "in press", and "online early" instead of "n/a".

This has been corrected.

835: Hoffstetter, RJ?

This has been corrected.

836: Is there something missing?

This has been corrected.

839: This is the same reference as number 20 (lines 683-684), and it is miscited in a different way...!

This has been corrected.

Reviewer #2 (Recommendations for the authors):

(1) There is a brief mention of a phylogenetic analysis being re-run, but it is unclear if any modifications (changes in scoring) based on the very observations were made. Please state this explicitly.

This is explained from lines 600-622, P.20-21, in the section “Apomorphic characters not empirically obtained.”  "In order to check the characters listed by Whiteside et al. [19] (p.19) as “two diagnostic characters” and “eight synapomorphies” in support of a squamate identity for †Cryptovaranoides, we conducted a parsimony analysis of the revised version of the dataset [32] provided by Whiteside et al. [19] in TNT v 1.5 [91]. We used Whiteside et al.’s [19] own data version"

(2) Line 20: There is almost no discussion of non‑lepidosaur lepidosauromorphs. I suggest including this, as the archosauromorph‑like features reported in Cryptovaranoides appear rather plastic. Furthermore, diagnostic features of Archosauromorpha in other datasets (e.g., Ezcurra 2016 or the works of Spiekman) are notably absent (and unsampled) in Cryptovaranoides. Expanding this comparison would greatly strengthen the manuscript.

The brief discussion (although not absent) of non-lepidosaur lepidosauromorphs is largely a function of the poor fossil record of this grade. But where necessary, we do discuss these taxa. Also see our previous study (Brownstein et al. 2023) for an extensive discussion of characters relevant to archosauromorphs.

(3) Line 38: I suggest removing "Archosauromorpha" from the keywords. The authors make a compelling case that Cryptovaranoides is not a squamate, yet they do not fully test its placement within Archosauromorpha (as they acknowledge). Perhaps use "Reptilia" instead?

We have removed this keyword.

(4) Line 99: The authors' points here are well made and largely valid. The presence of the ent‑ and ectepicondylar foramina is indeed an amniote plesiomorphy and cannot confirm a squamate identity. Their absence, however, can be informative - although it is unclear whether the CT scans of the humerus are of sufficient resolution, and Figure 4 of Brownstein et al. looks hastily reconstructed (perhaps owing to limited resolution). Moreover, the foramina illustrated by Whiteside do resemble those of other reptiles, albeit possibly over‑prepared and exaggerated.

The issue with the noted figure is indeed due to poor resolution from the scans. Although we agree with the reviewer, we hesitate to talk about absence in this taxon being phylogenetically informative given the confounding influence of ontogeny.

(5) I encourage the authors to provide slice data to support the claim that the foramina are absent (which could certainly be correct!); otherwise, the assertion remains unsubstantiated.

We only have access to the mesh files of segmented bones, not the raw (reconstructed slice) data.

(6) PLEASE NOTE - because the specimen is juvenile, the apparent absence of the ectepicondylar foramen is equivocal: the supinator process develops through ontogeny and encloses this foramen (see Buffa et al. 2025 on Thadeosaurus, for example).

See above.

(7) Line 122: Italicize 'Delorhynchus'

This has been corrected.

(8) Lines 131‑132: I'd suggest deleting the final sentence; it feels a little condescending, and your argument is already persuasive.

This has been corrected.

(9) Line 129: Please note that owenettid "parareptiles" also lack this process, as do several other stem‑saurians. Its absence is therefore not diagnostic of Squamata.<br /> Also: Such plasticity is common outside the crown. Milleropsis and Younginidae develop this process during ontogeny, even though a lower temporal bar never fully forms.

We appreciate this point. See discussion later in the manuscript.

(11) Line 172: Consider adding ontogeny alongside taphonomy and preservation. A juvenile would likely have a poorly developed radial condyle, if any. Acknowledging this possibility will add some needed nuance.

This sentence has been modified, but we have not added in discussion of ontogeny here because it is not immediately relevant to refuting the argument about inference of the presence of this feature when it is not preserved.

(12) Line 177: The "septomaxilla" in Whiteside et al. (2024, Figure 1C) resembles the contralateral premaxilla in dorsal view, with the maxillary process on the left and the palatal (or vomerine) process on the right (the dorsal process appears eroded). The foramen looks like a prepalatal foramen, common to many stem and crown reptiles. Consequently, scoring the septomaxilla as absent may be premature; this bone often ossifies late. In my experience with stem‑reptile aggregations, only one of several articulated individuals may ossify this element.

We agree that presence of a late-ossifying septomaxilla cannot be ruled out, but our point remains (and in agreement with Referee) that scoring the septomaxilla as present based on the amorphous fragments is premature.

(13) Line 200: Tomography data should be shown before citing it. The posterior margin of the maxilla appears rather straight, and the maxilla itself is tall for an archosauromorph. It would be more convincing to score this feature as present only after illustrating the relevant slices - and, as you note, the trait is widespread among non‑archosauromorphs.

See above and Brownstein et al. (2023).

(14) Line 208: Well argued: how could Whiteside et al. confidently assign a disarticulated element? Their "vagus" foramen actually resembles a standard hypoglossal foramen - identical to that seen in many stem reptiles, which often have one large and one small opening.

Thank you!

(15) Line 248: Again, please illustrate this region. One cannot argue for absence without showing the slice data. Note that millerettids and procolophonians - contemporaneous with Cryptovaranoides - possess an enclosed vidian canal, so the feature is broadly distributed.

See above.

(16) Line 258: The choanal fossa is intriguing: originally created for squamate matrices, yet present (to varying degrees) in nearly every reptile I have examined. It is strongly developed in millerettids (see Jenkins et al. 2025 on Milleropsis and Milleretta) and younginids, much like in squamates - Tiago appropriately scores it as present. Thus, it may be more of a "Neodiapsida + millerettids" character. In any case, the feature likely forms an ordered cline rather than a simple binary state.

We agree and look forward to future study of this feature.

(17) Line 283: Bolosaurids are not diapsids and, per Simões, myself, and others, "Diapsida" is probably invalid, at least how it is used here. Better to say "neodiapsids" for choristoderes and "stem‑reptiles" or "sauropsids" for bolosaurids. Jenkins et al.'s placement is largely a function of misidentifying the bolosaurid stapes as the opisthotic.

We are not entirely clear on this point since bolosaurids are not mentioned in this section.

(18) Line 298: Here, you note that the CT scans are rather coarse, which makes some earlier statements about absence/presence less certain (e.g., humeral foramina). It may strengthen the paper to make fewer definitive claims where resolution limits interpretation.

We appreciate this point. However, in the case of the humeral foramina the coarseness of the scans is one reason why we question Whiteside et al. scoring of the presence of these features.

(19) Line 314: Multiple rows of vomerine teeth are standard for amniotes; lepidosauromorphs such as Paliguana and Megachirella also exhibit them (though they may not have been segmented in the latter's description). Only a few groups (e.g., varanopids, some millerettids) have a single medial row.

We appreciate this point and have added in those citations into the following added sentence: “Multiple rows of vomerine teeth are common in reptiles outside of Squamata [76]; the presence of only one row is restricted to a handful of clades, including millerettids [77,78], †Tanystropheus [49], and some [79], but not all [71,80] choristoderes.” (L. 360-363, P. 12).

(20) Line 317: This is likely a reptile plesiomorphy - present in all millerettids (e.g., Milleropsis and Milleretta per Jenkins et al.). Citing these examples would clarify that it is not uniquely squamate. Could it be secondarily lost in archosauromorphs?

We appreciate this point and have cited Jenkins et al. here. It is out of the scope of this discussion to discuss the polarity of this feature relative to Archosauromorpha.

(21) Line 336: Unfortunately, a distinct quadratojugal facet is usually absent in Neodiapsids and millerettids; where present, the quadratojugal is reduced and simply overlaps the quadrate.

We appreciate this point but feel that reviewing the distribution of this feature across all reptiles is not relevant to the text noted.

(22) Line 357: Pterygoid‑quadrate overlap is likely a tetrapod plesiomorphy. Whiteside et al. do not define its functional or phylogenetic significance, and the overlap length is highly variable even among sister taxa.

We agree, but in any case this feature is impossible to assess in Cryptovaranoides.

(23) Line 365: Another well‑written section - clear and persuasive.

Thank you!

(24) Line 385: The cephalic condyle is widespread among neodiapsids, so it is not uniquely squamate.

We agree.

(25) Character 391: Note that the frontal underlapping the parietal is widespread, appearing in both millerettids and neodiapsids such as Youngina.

We appreciate this point, but the point here deals with the fact that this feature is not observable in the holotype of Cryptovaranoides.

(26) Line 415: The "anterior process" is actually common among crown reptiles, including sauropterygians, so it cannot by itself place Cryptovaranoides within Archosauromorpha.

We agree but also note that we do not claim this feature unambiguously unites Cryptovaranoides with Archosauromorpha.

(28) Line 460: Yes - Whiteside et al. appear to have relabeled the standard amniote coracoid foramen. Excellent discussion.

Thank you!

(29) Line 496: While mirroring Whiteside's structure, discussing this mandibular character earlier, before the postcrania, might aid readability.

We have chosen to keep this structure as is.

(30) Lines 486-588: This section oversimplifies the quadrate articulation.

We are unclear how this is an oversimplification.

(31) Both Prolacerta and Macrocnemus possess a cephalic condyle and some mobility (though less than many squamates). In Prolacerta (Miedema et al. 2020, Figure 4), the squamosal posteroventral process loosely overlaps the quadrate head.

We assume this comment refers to the section "Peg-in-notch articulation of quadrate head"; we appreciate clarification that this feature occurs in variable extent outside squamates, but this does not affect our statement that the material of Cryptovaranoides is too poorly preserved to confirm its presence.

(32) Where is this process in Cryptovaranoides? It is not evident in Whiteside's segmentation of the slender squamosal - please illustrate.

We are unclear as to which section this comment refers.

(33) Additionally, the quadrate "conch" of Cryptovaranoides is well developed, bearing lateral and medial tympanic crests; the lateral crest is absent in the cited archosauromorphs.

We note that no vertebrate has a medial tympanic crest (it is always laterally placed for the tympanic membrane, when present). If this is what the reviewer refers to, this is a feature commonly found across all tetrapods bearing a tympanum attached to the quadrate (e.g., most reptiles), and so it is not very relevant phylogenetically. Regarding its presence in Cryptovaranoides, the lateral margin of the quadrate is broken (Brownstein et al., 2023), so it cannot be determined. This incomplete preservation also makes an interpretation of a quadrate conch very hard to determine. But as currently preserved, there is no evidence whatsoever for this feature.

(34) Line 591: The cervical vertebrae of Cryptovaranoides are not archosauromorph‑like. Archosauromorph cervicals are elongate, parallelogram‑shaped, and carry long cervical ribs-none of which apply here. As the manuscript lacks a phylogenetic analysis, including these features seems unnecessary. Should they be added to other datasets, I suspect Cryptovaranoides would align along the lepidosaur stem (though that remains to be tested).

We politely disagree. The reviewer here mentions that the cervical vertebrae of archosauromorphs are generally shaped differently from those in Cryptovaranoides. The description provided (“elongate, parallelogram‑shaped, and carry long cervical ribs-none”) is basically limited to protorosaurians (e.g., tanystropheids, Macrocnemus) and early archosauriforms. We note that archosauromorph cervicals are notoriously variable in shape, especially in the crown, but also among early archosauromorphs. Further, the cervical ribs, are notoriously similar among early archosauromorphs (including protorosaurians) and Cryptovaranoides, as discussed and illustrated in Brownstein et al., 2023 (Figs. 2 and 3), especially concerning the presence of the anterior process.

Further, we do include a phylogenetic analysis of the matrix provided in Whiteside et al. (2024) as noted in our results section. In any case, we direct the reviewer to our previous study (Brownstein et al., 2023), in which we conduct phylogenetic analyses that included characters relevant to this note.

Reviewer #3 (Recommendations for the authors):

(1) The authors should use specimen numbers all over the text because we are talking about multiple individuals, and the authors contest the previous affinity of some of them. For example, on page 16, line 447, they mention an isolated vertebra but without any number. The specimen can be identified in the referenced article, but it would be much easier for the reader if the number were also provided here

Agreed and added.

(2) Abstract: "Our team questioned this identification and instead suggested Cryptovaranoides had unclear affinities to living reptiles."

That is very imprecise. The team suggested that it could be an archosauromorph or an indeterminate neodiapsid. Please change accordingly.

We politely disagree. We stated in our 2023 study that whereas our phylogenetic analyses place this taxon in Archosauromorpha, it remains unclear where it would belong within the latter. This is compatible with “unclear affinities to living reptiles”.

(3) Page 7, line 172: "Taphonomy and poor preservation cannot be used to infer the presence of an anatomical feature that is absent." Unfortunate wording. Taphonomy always has to be used to infer the presence or absence of anatomical features. Sometimes the feature is not preserved, but it leaves imprints/chemical traces or other taphonomic indicators that it was present in the organism. Please remove or rewrite the sentence.

We agree and have modified the sentence to read: “Taphonomy and poor preservation cannot be used alone to justify the inference that an anatomical feature was present when it is not preserved and there is no evidence of postmortem damage. In a situation when the absence of a feature is potentially ascribable to preservation, its presence should be considered ambiguous.” (L. 141-145, P.5).

(4) Page 4, line 91, please explain "WEA24" here, though it is unclear why this abbreviation is used instead of citation in the manuscript.

This has been corrected to Whiteside et al. [19].

(5) Page 6, line 144: "Together, these observations suggest that the presence of a jugal posterior process was incorrectly scored in the datasets used by WEA24 (type (ii) error)." That sentence is unclear. Why did the authors use "suggest"? Does it mean that they did not have access to the original data matrix to check it? If so, it should be clearly stated at the beginning of the manuscript.

See earlier; this has been modified and “suggest” has been removed.

(6) Page 7, line 174: "Finally, even in the case of the isolated humerus with a preserved capitulum, the condyle illustrated by Whiteside et al. [19] is fairly small compared to even the earliest known pan-squamates, such as Megachirella wachtleri (Figure 4)." Figure 4 does not show any humeri. Please correct.

The reference to figure 4 has been removed.

(7) Page 8, line 195-198: "This is not the condition specified in either of the morphological character sets that they cite [18,38], the presence of a distinct condyle that is expanded and is by their own description not homologous to the condition in other squamates." This is a bit unclear. Could the authors explain it a little bit further? How is the condition that is specified in the referred papers different compared to the Whiteside et al. description?

We appreciate this comment and have broken this sentence up into three sentences to clarify what we mean:

“The projection of the radial condyle above the adjacent region of the distal anterior extremity is not the condition specified in either of the morphological character sets that Whiteside et al. [19] cite [18,32]. The condition specified in those studies is the presence of a distinct condyle that is expanded. The feature described in Whiteside et al. [19] does not correspond to the character scored in the phylogenetic datasets.” (L.220-225, P.8).

(8) Page 16, line 446: "they observed in isolated vertebrae that they again refer to C. microlanius without justification". That is not true. The referred paper explains the attribution of these vertebrae to Cryptovaranoides (see section 5.3 therein). The authors do not have to agree with that justification, but they cannot claim that no justification was made. Please correct it here and throughout the text.

We have modified this sentence but note that the justification in Whiteside et al. (2024) lacked rigor. Whiteside et al. (2024) state: “Brownstein et al. [5] contested the affinities of three vertebrae, cervical vertebra NHMUK PV R37276, dorsal vertebra NHMUK PV R37277 and sacral vertebra NHMUK PV R37275. While all three are amphicoelous and not notochordal, the first two can be directly compared to the holotype. Cervical vertebra NHMUK PV R37276 is of the same form as the holotype CV3 with matching neural spine, ventral keel (=crest) and the posterior lateral ridges or lamina (figure 3c,d) shown by Brownstein et al. [5, fig. 1a]. The difference is that NHMUK PV R37276 has a fused neural arch to the pleurocentrum and a synapophysis rather than separate diapophysis and parapophysis of the juvenile holotype (figure 3c). Neurocentral fusion of the neural arch and centrum can occur late in modern squamates, ‘up to 82% of the species maximum size’ [28].

The dorsal surface of dorsal vertebra NHMUK PV R37277 (figure 3e) can be matched to the mid-dorsal vertebra in the †Cryptovaranoides holotype (figure 4d, dor.ve) and has the same morphology of wide, dorsally and outwardly directed, prezygapophyses, downwardly directed postzygapophyses and similar neural spine. It is also of similar proportions to the holotype when viewed dorsally (figures 3e and 4d), both being about 1.2 times longer anteroposteriorly than they are wide, measured across the posterior margin. The image in figure 4d demonstrates that the posterior vertebrae are part of the same spinal column as the truncated proximal region but the spinal column between the two parts is missing, probably lost in quarrying or fossil collection.”

This justification is based on pointing out the presence of supposed shared features between these isolated vertebrae and those in the holotype of Cryptovaranoides, even though none of these features are diagnostic for that taxon. We have changed the sentence in our manuscript to read:

“Whiteside et al. [19] concur with Brownstein et al. [18] that the diapophyses and parapophyses are unfused in the anterior dorsals of the holotype of †Cryptovaranoides microlanius, and restate that fusion of these structures is based on the condition they observed in isolated vertebrae that they refer to †C. microlanius based on general morphological similarity and without reference to diagnostic characters of †C. microlanius” (L. 502-507, P. 17).

(9) Figure 2. The figure caption lacks some explanations. Please provide information about affinity (e.g., squamate/gekkotan), ag,e and locality of the taxa presented. Are these left or right palatines? The second one seems to be incomplete, and maybe it is worth replacing it with something else?

The figure caption has been modified:

“Figure 2. Comparison of palatine morphologies. Blue shading indicates choanal fossa. Top image of †Cryptovaranoides referred left palatine is from Whiteside et al. [19]. Middle is the left palatine of †Helioscopos dickersonae (Squamata: Pan-Gekkota) from the Late Jurassic Morrison Formation [62]. Bottom is the right palatine of †Eoscincus ornatus (Squamata: Pan-Scincoidea) from the Late Jurassic Morrison Formation [31].”

(10) Figure 8. The abbreviations are not explained in the figure caption.

These have been added.

Orlowitz points out that Wikipedia is already widely used by people across all professions, including experts. Instead of telling users to avoid it completely, he argues that educators should focus on teaching critical and informed use of information. The key message is that digital literacy—not prohibition—is what makes Wikipedia useful and reliable.

LiDA101

Author response:

The following is the authors’ response to the original reviews

We would like to thank all reviewers for their constructive and in-depth reviews. Thanks to your feedback, we realized that the main objective of the paper was not presented clearly enough, and that our use of the same “modality-agnostic” terminology for both decoders and representations caused confusion. We addressed these two major points as outlined in the following. 

In the revised manuscript, we highlight that the main contribution of this paper is to introduce modality-agnostic decoders. Apart from introducing this new decoder type, we put forward their advantages in comparison to modality-specific decoders in terms of decoding performance and analyze the modality-invariant representations (cf. updated terminology in the following paragraph) that these decoders rely on. The dataset that these analyses are based on is released as part of this paper, in the spirit of open science (but this dataset is only a secondary contribution for our paper). 

Regarding the terminology, we clearly define modality-agnostic decoders as decoders that are trained on brain imaging data from subjects exposed to stimuli in multiple modalities. The decoder is not given any information on which modality a stimulus was presented in, and is therefore trained to operate in a modality-agnostic way. In contrast, modality-specific decoders are trained only on data from a single stimulus modality. These terms are explained in Figure 2. While these terms describe different ways of how decoders can be trained, there are also different ways to evaluate them afterwards (see also Figure 3); but obviously, this test-time evaluation does not change the nature of the decoder, i.e., there is no contradiction in applying a modality-specific decoder to brain data from a different modality.

Further, we identify representations that are relevant for modality-agnostic decoders using the searchlight analysis. We realized that our choice of using the same “modality-agnostic” term to describe these brain representations created unnecessary debate and confusion. In order to not conflate the terminology, in the updated manuscript we call these representations modality-invariant (and the opposite modality-dependent). Our methodology does not allow us to distinguish whether certain representations merely share representational structure to a certain degree, or are truly representations that abstract away from any modality-dependent information. However, in order to be useful for modality-agnostic decoding, a significant degree of shared representational structure is sufficient, and it is this property of brain representations that we now define as “modality-invariant”. 

We updated the manuscript in line with this new terminology and focus: in particular, the first Related Work section on Modality-invariant brain representations, as well as the Introduction and Discussion.

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors introduce a densely-sampled dataset where 6 participants viewed images and sentence descriptions derived from the MS Coco database over the course of 10 scanning sessions. The authors further showcase how image and sentence decoders can be used to predict which images or descriptions were seen, using pairwise decoding across a set of 120 test images. The authors find decodable information widely distributed across the brain, with a left-lateralized focus. The results further showed that modality-agnostic models generally outperformed modality-specific models, and that data based on captions was not explained better by caption-based models but by modality-agnostic models. Finally, the authors decoded imagined scenes.

Strengths:

(1) The dataset presents a potentially very valuable resource for investigating visual and semantic representations and their interplay.

(2) The introduction and discussion are very well written in the context of trying to understand the nature of multimodal representations and present a comprehensive and very useful review of the current literature on the topic.

Weaknesses:

(1) The paper is framed as presenting a dataset, yet most of it revolves around the presentation of findings in relation to what the authors call modality-agnostic representations, and in part around mental imagery. This makes it very difficult to assess the manuscript, whether the authors have achieved their aims, and whether the results support the conclusions.

Thanks for this insightful remark. The dataset release is only a secondary contribution of our study; this was not clear enough in the previous version. We updated the manuscript to make the main objective of the paper more clear, as outlined in our general response to the reviews (see above).

(2) While the authors have presented a potential use case for such a dataset, there is currently far too little detail regarding data quality metrics expected from the introduction of similar datasets, including the absence of head-motion estimates, quality of intersession alignment, or noise ceilings of all individuals.

As already mentioned in the general response, the main focus of the paper is to introduce modality-agnostic decoders. The dataset is released in addition, this is why we did not focus on reporting extensive quality metrics in the original manuscript. To respond to your request, we updated the appendix of the manuscript to include a range of data quality metrics. 

The updated appendix includes head motion estimates in the form of realignment parameters and framewise displacement, as well as a metric to assess the quality of intersession alignment. More detailed descriptions can be found in Appendix 1 of the updated manuscript.

Estimating noise ceilings based on repeated presentations of stimuli (as for example done in Allen et al. (2022)) requires multiple betas for each stimulus. All training stimuli were only presented once, so this could only be done for the test stimuli which were presented repeatedly. However, during our preprocessing procedure we directly calculated stimulus-specific betas based on data from all sessions using one single GLM, which means that we did not obtain separate betas for repeated presentations of the same stimulus. We will however share the raw data publicly, so that such noise ceilings can be calculated using an adapted preprocessing procedure if required.

Allen, E. J., St-Yves, G., Wu, Y., Breedlove, J. L., Prince, J. S., Dowdle, L. T., Nau, M., Caron, B., Pestilli, F., Charest, I., Hutchinson, J. B., Naselaris, T., & Kay, K. (2022). A massive 7T fMRI dataset to bridge cognitive neuroscience and artificial intelligence. Nature Neuroscience, 25(1), 116–126. https://doi.org/10.1038/s41593-021-00962-x

(3) The exact methods and statistical analyses used are still opaque, making it hard for a reader to understand how the authors achieved their results. More detail in the manuscript would be helpful, specifically regarding the exact statistical procedures, what tests were performed across, or how data were pooled across participants.

In the updated manuscript, we improved the level of detail for the descriptions of statistical analyses wherever possible (see also our response to your “Recommendations for the authors”, Point 6).

Regarding data pooling across participants: 

Figure 8 shows averaged results across all subjects (as indicated in the caption)

Regarding data pooling for the estimation of the significance threshold of the searchlight analysis for modality-invariant regions: We updated the manuscript to clarify that we performed a permutation test, combined with a bootstrapping procedure to estimate a group-level null distribution: “For each subject, we evaluated the decoders 100 times with shuffled labels to create per-subject chance-level results. Then, we randomly selected one of the 100 chance-level results for each of the 6 subjects and calculated group-level statistics (TFCE values) the exact same way as described in the preceding paragraph. We repeated this procedure 10,000 times resulting in 10,000 permuted group-level results.”

Additionally, we indicated that the same permutation testing methods were applied to assess the significance threshold for the imagery decoding searchlight maps (Figure 10). 

(4) Many findings (e.g., Figure 6) are still qualitative but could be supported by quantitative measures.

The Figures 6 and 7 are intentionally qualitative results to support the quantitative decoding results presented in Figures 4 and 5. (see also Reviewer 2 Comment 2)

Figures 4 and 5 show pairwise decoding accuracy as a quantitative measure for evaluation of the decoders. This metric is the main metric we used to compare different decoder types and features. Based on the finding that modality-agnostic decoders using imagebind features achieve the best score on this metric, we performed the additional qualitative analysis presented in Figures 6 and 7. (Note that we expanded the candidate set for the qualitative analysis in order to have a larger and more diverse set of images.)

(5) Results are significant in regions that typically lack responses to visual stimuli, indicating potential bias in the classifier. This is relevant for the interpretation of the findings. A classification approach less sensitive to outliers (e.g., 70-way classification) could avoid this issue. Given the extreme collinearity of the experimental design, regressors in close temporal proximity will be highly similar, which could lead to leakage effects.

It is true that our searchlight analysis revealed significant activity in regions outside of the visual cortex. However, it is assumed that the processing of visual information does not stop at the border of the visual cortex. The integration of information such as the semantics of the image is progressively processed in other higher-level regions of the brain. Recent studies have shown that activity in large areas of the cortex (including many outside of the visual cortex) can be related to visual stimulation (Solomon et al. 2024; Raugel et al. 2025). Our work confirms this finding and we therefore do not see reason to believe that this is due to a bias in our decoders.

Further, you are suggesting that we could replace our regression approach with a 70-way classification. However, this is difficult using our fMRI data as we do not see a straightforward way to assign the training and testing stimuli with class labels (the two datasets consist of non-overlapping sets of naturalistic images).

To address your concerns regarding the collinearity of the experimental design and possible leakage effects, we trained and evaluated a decoder for one subject after running a “null-hypothesis” adapted preprocessing. More specifically, for all sessions, we shifted the functional data of all runs by one run (moving the data of the last run to the very front), but leaving the design matrices in place. Thereby, we destroyed the relationship of stimuli and brain activity but kept the original data and design with its collinearity (and possible biases). We preprocessed this adapted data for subject 1, and ran a whole-brain decoding using Imagebind features and verified that the decoding performance was at chance level:  Pairwise accuracy (captions): 0.43 | Pairwise accuracy (images): 0.47 | Pairwise accuracy (imagery): 0.50. This result provides evidence against the notion that potential collinearity or biases in our experimental design or evaluation procedure could have led to inflated results.

Raugel, J., Szafraniec, M., Vo, H.V., Couprie, C., Labatut, P., Bojanowski, P., Wyart, V. and King, J.R. (2025). Disentangling the Factors of Convergence between Brains and Computer Vision Models. arXiv preprint arXiv:2508.18226.

Solomon, S. H., Kay, K., & Schapiro, A. C. (2024). Semantic plasticity across timescales in the human brain. bioRxiv, 2024-02.

(6) The manuscript currently lacks a limitations section, specifically regarding the design of the experiment. This involves the use of the overly homogenous dataset Coco, which invites overfitting, the mixing of sentence descriptions and visual images, which invites imagery of previously seen content, and the use of a 1-back task, which can lead to carry-over effects to the subsequent trial.

Regarding the dataset CoCo: We agree that CoCo is somewhat homogenous, it is however much more diverse and naturalistic than the smaller datasets used in previous fMRI experiments with multimodal stimuli. Additionally, CoCo has been widely adopted as a benchmark dataset in the Machine Learning community, and features rich annotations for each image (e.g. object labels, segmentations, additional captions, people’s keypoints) facilitating many more future analyses based on our data.

Regarding the mixing of sentence descriptions and images: Subjects were not asked to visualize sentences and different techniques for the one-back tasks might have been used. Generally, we do not see it as problematic if subjects are performing visual imagery to some degree while reading sentences, and this might even be the case during normal reading as well. A more targeted experiment comparing reading with and without interleaved visual stimulation in the form of images and a one-back task would be required to assess this, but this was not the focus of our study. For now, it is true that we can not be sure that our results generalize to cases in which subjects are just reading and are less incentivized to perform mental imagery.

Regarding the use of a 1-back task: It was necessary to make some design choices in order to realize this large-scale data collection with approximately 10 hours of recording per subject. Specifically, the 1-back task was included in the experimental setup in order to assure continuous engagement of the participant during the rather long sessions of 1 hour. The subjects did indeed need to remember the previous stimulus to succeed at the 1-back task, which means that some brain activity during the presentation of a stimulus is likely to be related to the previous stimulus. We aimed to account for this confound during the preprocessing stage when fitting the GLM, which was fit to capture only the response to the presented image/caption, not the preceding one. Still, it might have picked up on some of the activity from preceding stimuli, causing some decrease of the final decoding performance.

We added a limitations section to the updated manuscript to discuss these important issues.

(7) I would urge the authors to clarify whether the primary aim is the introduction of a dataset and showing the use of it, or whether it is the set of results presented. This includes the title of this manuscript. While the decoding approach is very interesting and potentially very valuable, I believe that the results in the current form are rather descriptive, and I'm wondering what specifically they add beyond what is known from other related work. This includes imagery-related results. This is completely fine! It just highlights that a stronger framing as a dataset is probably advantageous for improving the significance of this work.

Thanks a lot for pointing this out. Based on this comment and feedback from the other reviewers we restructured the abstract, introduction and discussion section of the paper to better reflect the primary aim. (cf. general response above).

You further mention that it is not clear what our results add beyond what is known from related work. We list the main contributions here:

A single modality-agnostic decoder can decode the semantics of visual and linguistic stimuli irrespective of the presentation modality with a performance that is not lagging behind modality-specific decoders.

Modality-agnostic decoders outperform modality-specific decoders for decoding captions and mental imagery.

Modality-invariant representations are widespread across the cortex (a range of previous work has suggested they were much more localized (Bright et al. 2004; Jung et al. 2018; Man et al. 2012; Simanova et al. 2014).

Regions that are useful for imagery are largely overlapping with modality-invariant regions

Bright, P., Moss, H., & Tyler, L. K. (2004). Unitary vs multiple semantics: PET studies of word and picture processing. Brain and language, 89(3), 417-432.

Jung, Y., Larsen, B., & Walther, D. B. (2018). Modality-Independent Coding of Scene Categories in Prefrontal Cortex. Journal of Neuroscience, 38(26), 5969–5981.

Liuzzi, A. G., Bruffaerts, R., Peeters, R., Adamczuk, K., Keuleers, E., De Deyne, S., Storms, G., Dupont, P., & Vandenberghe, R. (2017). Cross-modal representation of spoken and written word meaning in left pars triangularis. NeuroImage, 150, 292–307. https://doi.org/10.1016/j.neuroimage.2017.02.032

Man, K., Kaplan, J. T., Damasio, A., & Meyer, K. (2012). Sight and Sound Converge to Form Modality-Invariant Representations in Temporoparietal Cortex. Journal of Neuroscience, 32(47), 16629–16636.

Simanova, I., Hagoort, P., Oostenveld, R., & van Gerven, M. A. J. (2014). Modality-Independent Decoding of Semantic Information from the Human Brain. Cerebral Cortex, 24(2), 426–434.

Reviewer #2 (Public review):

Summary:

This study introduces SemReps-8K, a large multimodal fMRI dataset collected while subjects viewed natural images and matched captions, and performed mental imagery based on textual cues. The authors aim to train modality-agnostic decoders--models that can predict neural representations independently of the input modality - and use these models to identify brain regions containing modality-agnostic information. They find that such decoders perform comparably or better than modality-specific decoders and generalize to imagery trials.

Strengths:

(1) The dataset is a substantial and well-controlled contribution, with >8,000 image-caption trials per subject and careful matching of stimuli across modalities - an essential resource for testing theories of abstract and amodal representation.

(2) The authors systematically compare unimodal, multimodal, and cross-modal decoders using a wide range of deep learning models, demonstrating thoughtful experimental design and thorough benchmarking.

(3) Their decoding pipeline is rigorous, with informative performance metrics and whole-brain searchlight analyses, offering valuable insights into the cortical distribution of shared representations.

(4) Extension to mental imagery decoding is a strong addition, aligning with theoretical predictions about the overlap between perception and imagery.

Weaknesses:

While the decoding results are robust, several critical limitations prevent the current findings from conclusively demonstrating truly modality-agnostic representations:

(1) Shared decoding ≠ abstraction: Successful decoding across modalities does not necessarily imply abstraction or modality-agnostic coding. Participants may engage in modality-specific processes (e.g., visual imagery when reading, inner speech when viewing images) that produce overlapping neural patterns. The analyses do not clearly disambiguate shared representational structure from genuinely modality-independent representations. Furthermore, in Figure 5, the modality-agnostic encoder did not perform better than the modality-specific decoder trained on images (in decoding images), but outperformed the modality-specific decoder trained on captions (in decoding captions). This asymmetry contradicts the premise of a truly "modality-agnostic" encoder. Additionally, given the similar performance between modality-agnostic decoders based on multimodal versus unimodal features, it remains unclear why neural representations did not preferentially align with multimodal features if they were truly modality-independent.

We agree that successful modality-agnostic and cross-modal decoding does not necessarily imply that abstract patterns were decoded. In the updated manuscript, we therefore refer to these representations as modality-invariant (see also the updated terminology explained in the general response above).

If participants are performing mental imagery when reading, and this is allowing us to perform cross-decoding, then this means that modality-invariant representations are formed during this mental imagery process, i.e. that the representations formed during this form of mental imagery are compatible with representations during visual perception (or, in your words, produce overlapping neural patterns). While we can not know to what extent people were performing mental imagery while reading (or having inner speech while viewing images), our results demonstrate that their brain activity allows for decoding across modalities, which implies that modality-invariant representations are present.

It is true that our current analyses can not disambiguate modality-invariant representations (or, in your words, shared representational structure) from abstract representations (in your words, genuinely modality-independent representations). As the main goal of the paper was to build modality-agnostic decoders, and these only require what we call “modality-invariant” representations (see our updated terminology in the general reviewer response above), we leave this question open for future work. We do however discuss this important limitation in the Discussion section of the updated manuscript.

Regarding the asymmetry of decoding results when comparing modality-agnostic decoders with the two respective modality-specific decoders for captions and images: We do not believe that this asymmetry contradicts the premise of a modality-agnostic decoder. Multiple explanations for this result are possible: (1) The modality-specific decoder for images might benefit from the more readily decodable lower-level modality-dependent neural activity patterns in response to images, which are less useful for the modality-agnostic decoder because they are not useful for decoding caption trials. The modality-specific decoders for captions might not be able to pick up on low-level modality-dependent neural activity patterns as these might be less easily decodable. 

The signal-to-noise ratio for caption trials might be lower than for image trials (cf. generally lower caption decoding performance), therefore the addition of training data (even if it is from another modality) improves the decoding performance for captions, but not for images (which might be at ceiling already).

Regarding the similar performance between modality-agnostic decoders based on multimodal versus unimodal features: Unimodal features are based on rather high-level features of the respective modality (e.g. last-layer features of a model trained for semantic image classification), which can be already modality-invariant to some degree. Additionally, as already mentioned before, in the updated manuscript we only require representations to be modality-invariant and not necessarily abstract.

(2) The current analysis cannot definitively conclude that the decoder itself is modality-agnostic, making "Qualitative Decoding Results" difficult to interpret in this context. This section currently provides illustrative examples, but lacks systematic quantitative analyses.

The qualitative decoding results in Figures 6 and 7 present exemplary qualitative results for the quantitative results presented in Figures 4 and 5 (see also Reviewer 1 Comment 4).

Figures 4 and 5 show pairwise decoding accuracy as a quantitative measure for evaluation of the decoders. This metric is the main metric we used to compare different decoder types and features. Based on the finding that modality-agnostic decoders using imagebind features achieve the best score on this metric, we performed the additional qualitative analysis presented in Figures 6 and 7. (Note that we expanded the candidate set for the qualitative analysis in order to have a larger and more diverse set of images.)

(3) The use of mental imagery as evidence for modality-agnostic decoding is problematic.

Imagery involves subjective, variable experiences and likely draws on semantic and perceptual networks in flexible ways. Strong decoding in imagery trials could reflect semantic overlap or task strategies rather than evidence of abstraction.

It is true that mental imagery does not necessarily rely on modality-agnostic representations. In the updated manuscript we revised our terminology and refer to the analyzed representations as modality-invariant, which we define as “representations that significantly overlap between modalities”. 

The manuscript presents a methodologically sophisticated and timely investigation into shared neural representations across modalities. However, the current evidence does not clearly distinguish between shared semantics, overlapping unimodal processes, and true modality-independent representations. A more cautious interpretation is warranted.

Nonetheless, the dataset and methodological framework represent a valuable resource for the field.

We fully agree with these observations, and updated our terminology as outlined in the general response.

Reviewer #3 (Public review):

Summary:

The authors recorded brain responses while participants viewed images and captions. The images and captions were taken from the COCO dataset, so each image has a corresponding caption, and each caption has a corresponding image. This enabled the authors to extract features from either the presented stimulus or the corresponding stimulus in the other modality.

The authors trained linear decoders to take brain responses and predict stimulus features.

"Modality-specific" decoders were trained on brain responses to either images or captions, while "modality-agnostic" decoders were trained on brain responses to both stimulus modalities. The decoders were evaluated on brain responses while the participants viewed and imagined new stimuli, and prediction performance was quantified using pairwise accuracy. The authors reported the following results:

(1) Decoders trained on brain responses to both images and captions can predict new brain responses to either modality.

(2) Decoders trained on brain responses to both images and captions outperform decoders trained on brain responses to a single modality.

(3) Many cortical regions represent the same concepts in vision and language.

(4) Decoders trained on brain responses to both images and captions can decode brain responses to imagined scenes.

Strengths:

This is an interesting study that addresses important questions about modality-agnostic representations. Previous work has shown that decoders trained on brain responses to one modality can be used to decode brain responses to another modality. The authors build on these findings by collecting a new multimodal dataset and training decoders on brain responses to both modalities.

To my knowledge, SemReps-8K is the first dataset of brain responses to vision and language where each stimulus item has a corresponding stimulus item in the other modality. This means that brain responses to a stimulus item can be modeled using visual features of the image, linguistic features of the caption, or multimodal features derived from both the image and the caption. The authors also employed a multimodal one-back matching task, which forces the participants to activate modality-agnostic representations. Overall, SemReps-8K is a valuable resource that will help researchers answer more questions about modality-agnostic representations.

The analyses are also very comprehensive. The authors trained decoders on brain responses to images, captions, and both modalities, and they tested the decoders on brain responses to images, captions, and imagined scenes. They extracted stimulus features using a range of visual, linguistic, and multimodal models. The modeling framework appears rigorous, and the results offer new insights into the relationship between vision, language, and imagery. In particular, the authors found that decoders trained on brain responses to both images and captions were more effective at decoding brain responses to imagined scenes than decoders trained on brain responses to either modality in isolation. The authors also found that imagined scenes can be decoded from a broad network of cortical regions.

Weaknesses:

The characterization of "modality-agnostic" and "modality-specific" decoders seems a bit contradictory. There are three major choices when fitting a decoder: the modality of the training stimuli, the modality of the testing stimuli, and the model used to extract stimulus features. However, the authors characterize their decoders based on only the first choice-"modality-specific" decoders were trained on brain responses to either images or captions, while "modality-agnostic" decoders were trained on brain responses to both stimulus modalities. I think that this leads to some instances where the conclusions are inconsistent with the methods and results.

In our analysis setup, a decoder is entirely determined by two factors: (1) the modality of the stimuli that the subject was exposed to, and (2) the machine learning model used to extract stimulus features.

The modality of the testing stimuli defines whether we are evaluating the decoder in a within-modality or cross-modality setting, but is not an inherent characteristic of a trained decoder

First, the authors suggest that "modality-specific decoders are not explicitly encouraged to pick up on modality-agnostic features during training" (line 137) while "modality-agnostic decoders may be more likely to leverage representations that are modality-agnostic" (line 140). However, whether a decoder is required to learn modality-agnostic representations depends on both the training responses and the stimulus features. Consider the case where the stimuli are represented using linguistic features of the captions. When you train a "modality-specific" decoder on image responses, the decoder is forced to rely on modality-agnostic information that is shared between the image responses and the caption features. On the other hand, when you train a "modality-agnostic" decoder on both image responses and caption responses, the decoder has access to the modality-specific information that is shared by the caption responses and the caption features, so it is not explicitly required to learn modality-agnostic features. As a result, while the authors show that "modality-agnostic" decoders outperform "modality-specific" decoders in most conditions, I am not convinced that this is because they are forced to learn more modality-agnostic features.

It is true that for example a modality-specific decoder trained on fmri data from images with stimulus features extracted from captions might also rely on modality-invariant features. We still call this decoder modality-specific, as it has been trained to decode brain activity recorded from a specific stimulus modality. In the updated manuscript we corrected the statement that “modality-specific decoders are not explicitly encouraged to pick up on modality-invariant features during training” to include the case of decoders trained on features from the other modality which might also rely on modality-invariant features.

It is true that a modality-agnostic decoder can also have access to modality-dependent information for captions and images. However, as it is trained jointly with both modalities and the modality-dependent features are not compatible, it is encouraged to rely on modality-invariant features. The result that modality-agnostic decoders are outperforming modality-specific decoders trained on captions for decoding captions confirms this, because if the decoder was only relying on modality-dependent features the addition of additional training data from another stimulus modality could not increase the performance. (Also, the lack of a performance drop compared to modality-specific decoders trained on images is only possible thanks to the reliance on modality-invariant features. If the decoder only relied on modality-dependent features the addition of data from another modality would equal an addition of noise to the training data which must result in a performance drop at test time.). We can not exclude the possibility that modality-agnostic decoders are also relying on modality-dependent features, but our results suggest that they are relying at least to some degree on modality-invariant features.

Second, the authors claim that "modality-specific decoders can be applied only in the modality that they were trained on, while "modality-agnostic decoders can be applied to decode stimuli from multiple modalities, even without knowing a priori the modality the stimulus was presented in" (line 47). While "modality-agnostic" decoders do outperform "modality-specific" decoders in the cross-modality conditions, it is important to note that "modality-specific" decoders still perform better than expected by chance (figure 5). It is also important to note that knowing about the input modality still improves decoding performance even for "modality-agnostic" decoders, since it determines the optimal feature space-it is better to decode brain responses to images using decoders trained on image features, and it is better to decode brain responses to captions using decoders trained on caption features.

Thanks for this important remark. We corrected this statement and now say that “modality-specific decoders that are trained to be applied only in the modality that they were trained on”, highlighting that their training process optimizes them for decoding in a specific modality. They can indeed be applied to the other modality at test time, this however results in a substantial performance drop.

It is true that knowing the input modality can improve performance even for modality-agnostic decoders. This can most likely be explained by the fact that in that case the decoder can leverage both, modality-invariant and modality-dependent features. We will not further focus on this result however as the main motivation to build modality-agnostic decoders is to be able to decode stimuli without knowing the stimulus modality a priori. 

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

I will list additional recommendations below in no specific order:

(1) I find the term "modality agnostic" quite unusual, and I believe I haven't seen it used outside of the ML community. I would urge the authors to change the terminology to be more common, or at least very early explain why the term is much better suited than the range of existing terms. A modality agnostic representation implies that it is not committed to a specific modality, but it seems that a representation cannot be committed to something.

In the updated manuscript we now refer to the identified brain patterns as modality-invariant, which has previously been used in the literature (Man et al. 2012; Devereux et al. 2013; Patterson et al. 2016; Deniz et al. 2019, Nakai et al. 2021) (see also the general response on top and the Introduction and Related Work sections of the updated manuscript).

We continue to refer to the decoders as modality-agnostic, as this is a new type of decoder, and describes the fact that they are trained in a way that abstracts away from the modality of the stimuli. We chose this term as we are not aware of any work in which brain decoders were trained jointly on multiple stimulus modalities and in order not to risk contradictions/confusions with other definitions.

Deniz, F., Nunez-Elizalde, A. O., Huth, A. G., & Gallant, J. L. (2019). The Representation of Semantic Information Across Human Cerebral Cortex During Listening Versus Reading Is Invariant to Stimulus Modality. Journal of Neuroscience, 39(39), 7722–7736. https://doi.org/10.1523/JNEUROSCI.0675-19.2019

Devereux, B. J., Clarke, A., Marouchos, A., & Tyler, L. K. (2013). Representational Similarity Analysis Reveals Commonalities and Differences in the Semantic Processing of Words and Objects. The Journal of Neuroscience, 33(48).

Nakai, T., Yamaguchi, H. Q., & Nishimoto, S. (2021). Convergence of Modality Invariance and Attention Selectivity in the Cortical Semantic Circuit. Cerebral Cortex, 31(10), 4825–4839. https://doi.org/10.1093/cercor/bhab125

Man, K., Kaplan, J. T., Damasio, A., & Meyer, K. (2012). Sight and Sound Converge to Form Modality-Invariant Representations in Temporoparietal Cortex. Journal of Neuroscience, 32(47), 16629–16636.

Patterson, K., & Lambon Ralph, M. A. (2016). The Hub-and-Spoke Hypothesis of Semantic Memory. In Neurobiology of Language (pp. 765–775). Elsevier. https://doi.org/10.1016/B978-0-12-407794-2.00061-4

(2) The table in Figure 1B would benefit from also highlighting the number of stimuli that have overlapping captions and images.

The number of overlapping stimuli is rather small (153-211 stimuli depending on the subject). We added this information to Table 1B. 

(3) The authors wrote that training stimuli were presented only once, yet they used a one-back task. Did the authors also exclude the first presentation of these stimuli?

Thanks for pointing this out. It is indeed true that some training stimuli were presented more than once, but only for the case of one-back target trials. In these cases the second presentation of the stimulus was excluded, but not the first. As the subject can not be aware of the fact that the upcoming presentation is going to be a one-back target, the first presentation can not be affected by the presence of the subsequent repeated presentation. We updated the manuscript to clarify this issue.

(4) Coco has roughly 80-90 categories, so many image captions will be extremely similar (e.g., "a giraffe walking", "a surfer on a wave", etc.). How can people keep these apart?

It is true that some captions and images are highly similar even though they are not matching in the dataset. This might result in several false button presses because the subjects identified an image-caption pair as matching when in fact it wasn't intended to. However, as there was no feedback given on the task performance, this issue should not have had a major influence on the brain activity of the participants.

(5) Footnotes for statistics are quite unusual - could the authors integrate statistics into the text?

Thanks for this remark, in the updated manuscript all statistics are part of the main text.

(6) It may be difficult to achieve the assumptions of a permutation test - exchangeability, which may bias statistical results. It is not uncommon for densely sampled datasets to use bootstrap sampling on the predictions of the test data to identify if a given percentile of that distribution crosses 0. The lowest p-value is given by the number of bootstrap samples (e.g., if all 10,000 bootstrap samples are above chance, then p < 0.0001). This may turn out to be more effective.

Thanks for this comment. Our statistical procedure was in fact involving a bootstrapping procedure to generate a null distribution on the group-level. We updated the manuscript to describe this method in more detail. Here is the updated paragraph: “To estimate the statistical significance of the resulting clusters we performed a permutation test, combined with a bootstrapping procedure to estimate a group-level null distribution see also Stelzer et al., 2013). For each subject, we evaluated the decoders 100 times with shuffled labels to create per-subject chance-level results. Then, we randomly selected one of the 100 chance-level results for each of the 6 subjects and calculated group-level statistics (TFCE values) the exact same way as described in the preceding paragraph. We repeated this procedure 10,000 times resulting in 10,000 permuted group-level results. We ensured that every permutation was unique, i.e. no two permutations were based on the same combination of selected chance-level results. Based on this null distribution, we calculated p-values for each vertex by calculating the proportion of sampled permutations where the TFCE value was greater than the observed TFCE value. To control for multiple comparisons across space, we always considered the maximum TFCE score across vertices for each group-level permutation (Smith and Nichols, 2009).”

(7) The authors present no statistical evidence for some of their claims (e.g., lines 335-337). It would be good if they could complement this in their description. Further, the visualization in Figure 4 is rather opaque. It would help if the authors could add a separate bar for the average modality-specific and modality-agnostic decoders or present results in a scatter plot, showing modality-specific on the x-axis and modality-agnostic on the y-axis and color-code the modality (i.e., making it two scatter colors, one for images, one for captions). All points will end up above the diagonal.

We updated the manuscript and added statistical evidence for the claims made:

We now report results for the claim that when considering the average decoding performance for images and captions, modality-agnostic decoders perform better than modality-specific decoders, irrespective of the features that the decoders were trained on.

Additionally, we report the average modality-agnostic and modality-specific decoding accuracies corresponding to Figure 4. For modality-agnostic decoders the average value is 81.86\%, for modality-specific decoders trained on images 78.15\%, and for modality-specific decoders trained on captions 72.52\%. We did not add a separate bar to Figure 4 as this would add additional information to a Figure which is already very dense in its information content (cf. Reviewers 2’s recommendations for the authors). We therefore believe it is more useful to report the average values in the text and provide results for a statistical test comparing the decoder types. A scatter plot would make it difficult to include detailed information on the features, which we believe is crucial.

We further provide statistical evidence for the observation regarding the directionality of cross-modal decoding.

Reviewer #2 (Recommendations for the authors):

For achieving more evidence to support modality-agnostic representations in the brain, I suggest more thorough analyses, for example:

(1) Traditional searchlight RSA using different deep learning models. Through this approach, it might identify different brain areas that are sensitive to different formats of information (visual, text, multimodal); subsequently, compare the decoding performance using these ROIs.

(2) Build more dissociable decoders for information of different modality formats, if possible. While I do not have a concrete proposal, more targeted decoder designs might better dissociate representational formats (i.e., unimodal vs. modality-agnostic).

(3) A more detailed exploration of the "qualitative decoding results"--for example, quantitatively examining error types produced by modality-agnostic versus modality-specific decoders--would be informative for clarifying what specific content the decoder captures, potentially providing stronger evidence for modality-agnostic representations.

Thanks for these suggestions. As the main goal of the paper is to introduce modality-agnostic decoders (which should be more clear from the updated manuscript, see also the general response to reviews), we did not include alternative methods for identifying modality-invariant regions. Nonetheless, we agree that in order to obtain more in-depth insight into the nature of representations that were recorded, performing analyses with additional methods such as RSA, comparisons with more targeted decoder designs in terms of their target features will be indispensable, as well as more in-depth error type analyses. We leave these analyses as promising directions for future work.

The writing could be further improved in the introduction and, accordingly, the discussion. The authors listed a series of theories about conceptual representations; however, they did not systematically explain the relationships and controversies between them, and it seems that they did not aim to address the issues raised by these theories anyway. Thus, the extraction of core ideas is suggested. The difference between "modality-agnostic" and terms like "modality-independent," "modality-invariant," "abstract," "amodal," or "supramodal," and the necessity for a novel term should be articulated.

The updated manuscript includes an improved introduction and discussion section that highlight the main focus and contributions of the study.

We believe that a systematic comparison of theories on conceptual representations involving their relationships and controversies would require a dedicated review paper. Here, we focused on the aspects that are relevant for the study at hand (modality-invariant representations), for which we find that none of the considered theories can be rejected based on our results.

Regarding the terminology (modality-agnostic vs. modality-invariant, ..) please refer to the general response.

The figures also have room to improve. For example, Figures 4 and 5 present dense bar plots comparing multiple decoding settings (e.g., modality-specific vs. modality-agnostic decoders, feature space, within-modal vs. cross-modal, etc.); while comprehensive, they would benefit from clearer labels or separated subplots to aid interpretation. All figures are recommended to be optimized for greater clarity and directness in future revisions.

Thanks for this remark. We agree that the figures are quite dense in information. However, splitting them up into subplots (e.g. separate subplots for different decoder types) would make it much less straightforward to compare the accuracy scores between conditions. As the main goal of these figures is to compare features and decoder types, we believe that it is useful to keep all information in the same plot. 

You are also suggesting to improve the clarity of the labels. It is true that the top left legend of Figures 4 and 5 was mixing information about decoder type and broad classes of features  (vision/language/multimodal). To improve clarity, we updated the figures and clearly separated information on decoder type (the hue of different bars) and features (x-axis labels).  The broad classes of features (vision/language/multimodal) are distinguished by alternating light gray background colors and additional labels at the very bottom of the plots.

The new plots allow for easy performance comparison of the different decoder types and additionally provide information on confidence intervals for the performance of modality-specific decoders, which was not available in the previous figures.

Reviewer #3 (Recommendations for the authors):

(1) As discussed in the Public Review, I think the paper would greatly benefit from clearer terminology. Instead of describing the decoders as "modality-agnostic" and "modality-specific", perhaps the authors could describe the decoding conditions based on the train and test modalities (e.g., "image-to-image", "caption-to-image", "multimodal-to-image") or using the terminology from Figure 3 (e.g., "within-modality", "cross-modality", "modality-agnostic").

We updated our terminology to be clearer and more accurate, as outlined in the general response. The terms modality-agnostic and modality-specific refer to the training conditions, and the test conditions are described in Figure 3 and are used throughout the paper.

(2) Line 244: I think the multimodal one-back task is an important aspect of the dataset that is worth highlighting. It seems to be a relatively novel paradigm, and it might help ensure that the participants are activating modality-agnostic representations.

It is true that the multimodal one-back task could play an important role for the activation of modality-invariant representations. Future work could investigate to what degree the presence of widespread modality-invariant representations is dependent on such a paradigm.

(3) Line 253: Could the authors elaborate on why they chose a random set of training stimuli for each participant? Is it to make the searchlight analyses more robust?

A random set of training stimuli was chosen in order to maximize the diversity of the training sets, i.e. to avoid bias based on a specific subsample of the CoCo dataset. Between-subject comparisons can still be made based on the test set which was shared for all subjects, with the limitation that performance differences due to individual differences or to the different training sets can not be disentangled. However, the main goal of the data collection was not to make between-subject comparisons based on common training sets, but rather to make group-level analyses based on a large and maximally diverse dataset. 

(4) Figure 4: Could the authors comment more on the patterns of decoding performance in Figure 5? For instance, it is interesting that ResNet is a better target than ViT, and BERT-base is a better target than BERT-large.

A multitude of factors influence the decoding performance, such as features dimensionality, model architecture, training data, and training objective(s) (Conwell et al. 2023; Raugel et al. 2025). Bert-base might be better than bert-large because the extracted features are of lower dimension. Resnet might be better than ViT because of its architecture (CNN vs. Transformer). To dive deeper into these differences further controlled analysis would be necessary, but this is not the focus of this paper. The main objective of the feature comparison was to provide a broad overview over visual/linguistic/multimodal feature spaces and to identify the most suitable features for modality-agnostic decoding.

Conwell, C., Prince, J. S., Kay, K. N., Alvarez, G. A., & Konkle, T. (2023). What can 1.8 billion regressions tell us about the pressures shaping high-level visual representation in brains and machines? (p. 2022.03.28.485868). bioRxiv. https://doi.org/10.1101/2022.03.28.485868

Raugel, J., Szafraniec, M., Vo, H.V., Couprie, C., Labatut, P., Bojanowski, P., Wyart, V. and King, J.R. (2025). Disentangling the Factors of Convergence between Brains and Computer Vision Models. arXiv preprint arXiv:2508.18226.

(5) Figure 7: It is interesting that the modality-agnostic decoder predictions mostly appear traffic-related. Is there a possibility that the model always produces traffic-related predictions, making it trivially correct for the presented stimuli that are actually traffic-related? It could be helpful to include some examples where the decoder produces other types of predictions to dispel this concern.

The presented qualitative examples were randomly selected. To make sure that the decoder is not always predicting traffic-related content, we included 5 additional randomly selected examples in Figures 6 and 7 of the updated manuscript. In only one of the 5 new examples the decoder was predicting traffic-related content, and in this case the stimulus had actually been traffic-related (a bus).

Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

Learn more at Review Commons

Reply to the reviewers

Manuscript number: RC-2025-03195R

Point-by-Point Response to Reviewers

We thank the reviewers for their thoughtful and constructive evaluations, which have helped us substantially improve the clarity, rigor, and balance of our manuscript. We are grateful for their recognition that our integrated ATAC-seq and RNA-seq analyses provide a valuable and technically sound contribution to understanding soxB1-2 function and regenerative neurogenesis in planarians.

We have carefully addressed the reviewers' major points as follows:

1. Direct versus indirect regulation by SoxB1-2:____ In the revision, we explicitly acknowledge the limitations of inferring direct regulation from our current datasets and have revised statements throughout the Results and Discussion to emphasize that our findings are correlative.
2. Evidence for pioneer activity:____ Although the pioneer role of SoxB1 transcription factors in well established in other systems, we agree that additional binding or motif data would be required to formally demonstrate SoxB1-2 pioneer function. Accordingly, we performed motif analysis and revised the text throughout to frame SoxB1-2's proposed role as consistent with, rather than demonstrating transcriptional activator activity.
3. Motif enrichment and downstream regulatory interactions:____ In response to Reviewer #1's suggestion, we have included a new motif enrichment analysis in the supplement to contextualize possible co-regulators within the SoxB1-2 network.
4. Data reproducibility and peak-calling consistency:____ We have included sample correlations ____and peak overlaps for ATAC-seq samples in the revision, providing a clearer assessment of reproducibility.
5. Clarification of co-expression and downstream targets:____ We included co-expression plots for soxB1-2 with mecom and castor in the supplemental materials. These plots were generated from previously published scRNA-seq data and demonstrate that cells expressing soxB1-2 also express mecom and __ __We appreciate the reviewers' recognition that our methods are rigorous and our data accessible. We have incorporated all major revisions suggested and believe have strengthened the manuscript's precision, interpretations, and conclusions. Below, we respond to each comment in detail.

Reviewer #1 (Evidence, reproducibility and clarity (Required)):

Summary

The authors of this interesting study take the approach of combining RNAi, RNA-seq and ATAC-seq to try to build a regulatory network surrounding the function of a planarian SoxB1 ortholog, broadly required for neural specification during planarian regeneration. They find a number of chromatin regions that differentially accessible (measured by ATAC-seq), associate these with potential genes by proximity to the TSS. They then compare this set of genes with those that are differentially regulated (using RNA-seq), after SoxB1 RNAi mediated knockdown. This allows them the authors some focus on potential directly regulated targets of the planarian SoxB1. Two of these downstream targets, the mecom and castor transcription factors are then studied in greater detail.

Major Comments

I have no suggestions for new experiments that fit sensibly with the scope of the current work. There are other analyses that could be appropriate with the ATAC-seq data, but may not make sense in the content of SoxB1 acting as pioneer factor.

I would like to see motif enrichment analysis under the set of peaks to see if SoxB1 is opening chromatin for a restricted set of other transcription factors to then bind. Much of this could be taken from Neiro et al, eLife 2022 (which also used ATAC-seq) and matched planarians TF families to likely binding motifs. This could add some breadth to the regulatory network. It could be revealing for example if downstream TF also help regulate other targets that SoxB1 makes available, this is pattern often seen for cell specification (as I am sure the authors are aware). Alternatively, it may reveal other candidate regulators.

Thank you for this suggestion. We agree with the reviewers that this analysis should be done. We ran the motif enrichment analysis using the same methods as outlined in Neiro et al. eLife, 2022. We have included a new motif enrichment analysis in the supplement to contextualize possible co-regulators within the SoxB1-2 network.

Overall peak calling consistency with ATAC-sample would be useful to report as well, to give readers an idea of noise in the data. What was the correlation between samples?

__Excellent point. In response to this comment, we ran a Pearson correlation test on replicates within gfp and soxB1-2 RNAi replicates to get an idea of overall correlation between replicates. Additionally, we calculated percent overlap of peaks for biological replicates and between treatment groups. __

While it is logical to focus on downregulated genes, it would also be interesting to look at upregulated genes in some detail. In simple terms would we expect to see the representation of an alternate set of fate decisions being made by neoblast progeny?

This is also an important point that we considered but initially did not pursue it due to the lack of tools to test upregulated gene function. However, the reviewer is correct that this is straightforward to perform computationally. Thus, we have performed Gene Ontology analysis on the upregulated genes in all RNA-seq datasets (soxB1-2 RNAi, mecom RNAi, and castor RNAi). Both mecom and castor datasets did not reveal enrichment within the upregulated portion of the dataset. Genes upregulated after soxB1-2 RNAi were enriched for metabolic, xenobiotic detoxification, potassium homeostasis, and endocytic programs. Rather than indicating a shift toward alternative lineages, including non-ectodermal fates, these signatures are consistent with stress-responsive and homeostatic programs activated following loss of soxB1-2. We did not detect enrichment patterns strongly associated with alternative cell fates. We conclude that this analysis does not formally exclude potential shifts in lineage-specific transcriptional programs, but does support our hypothesis that soxB1-2 functions as a transcriptional activator.

Can the authors be explicit about whether they have evidence for co-expression of SoxB1/castor and SoxB1/mecom? I could find this clearly and it would be important to be clear whether this basic piece of evidence is in place or not at this stage.

We included co-expression plots for soxB1-2 with mecom and castor in the supplemental material. These plots were generated from previously published scRNA-seq data and demonstrate that cells expressing soxB1-2 also express mecom and castor. We have not done experiments showing co-expression via in situ at this time.

Minor comments

Formally loss of castor and mecom expression does mean these cells are absent, strictly the cell absence needs an independent method. It might be useful to clarify this with the evidence of be clear that cells are "very probably" not produced.

We agree that loss of castor and mecom expression does not formally demonstrate the physical absence of these cells, and that independent methods would be required to definitively confirm their loss. In response, we have revised our wording to indicate that castor- and mecom-expressing cells are very likely not being produced, rather than stating that they are absent.

Reviewer #1 (Significance (Required)):

Significance

Strengths and limitations.

The precise exploitation of the planarian system to identify potential targets, and therefore regulatory mechanisms, mediated by SoxB1 is an interesting contribution to the fi eld. We know almost nothing about the regulatory mechanisms that allow regeneration and how these might have evolved, and this work is well-executed step in that direction.

Advance

The paper makes a clear advance in our understanding of an important process in animals (neural specification) and how this happens in the context in the context during an example of animal regeneration. The methods are state-of-the-art with respect to what is possible in the planarian system.

Audience

This will be of wide interest to developmental biologists, particularly those studying regeneration in planarians and other regenerative systems,and those who study comparative neurodevelopment.

Expertise

I have expertise in functional genomics in the context of stem cells and regeneration, particularly in the planarian model system

Reviewer #2 (Evidence, reproducibility and clarity (Required)):

Review - Cathell, et al (RC-2025-03195)

Summary and Significance:

Understanding regenerative neurogenesis has been difficult due to the limited amount of neurogenesis that occurs after injury in most animal species. Planarians, with their adult neurogenesis and robust post-injury response, allow us to get a glimpse into regenerative neurogenesis. The Zayas laboratory previously revealed a key role for SoxB1-2 in maintenance and regeneration of a broad set of sensory and peripheral neurons in the planarian body. SoxB1-2 also has a role in many epidermal fates. Their previous work left open the tempting possibility that SoxB1-2 acts as a very upstream regulator of epidermal and neuronal fates, potentially acting as a pioneer transcription factor within these lineages. In the manuscript currently under review, Cathell and colleagues use ATAC-Seq and RNA-Seq to investigate chromatin changes after SoxB1-2(RNAi). With the experimental limitations in planarians, this is a strong first step toward testing their hypothesis that SoxB1-2acts as a pioneer within a set of planarian lineages. Beyond these cell types, this work is also important because planarian cell fates often rely on a suite of transcription factors, but the nature of transcription factor cooperation has been much less well understood. Indeed, the authors do show that loss of SoxB1-2 by RNAi causes changes in a number of accessible regions of the genome; many of these chromatin changes correspond to changes in gene expression of genes nearby these peaks. The authors also examine in more detail two genes that have genomic and transcriptomic changes after SoxB1-2(RNAi), mecom and castor. The authors completed RNA-Seq on mecom(RNAi) and castor(RNAi) animals, identifying genes downregulated after loss of either factor that are also seen in SoxB1-2(RNAi). The results in this paper are rigorous and very well presented. I will share two major limitations of the study and some suggestions for addressing them, but this work may also be acceptable without those changes at some journals.

Limitation 1:

The paper aims to test the hypothesis that SoxB1-2 is a pioneer transcription factor. Observation that SoxB1-2(RNAi) leads to loss of many accessible regions in the chromatin supports the hypothesis. However, an alternate possibility is that SoxB1-2 leads to transcription of another factor that is a pioneer factor or a chromatin remodeling enzyme; in either of these cases, the accessibility peak changes may not be due to SoxB1-2 directly but due to another protein that SoxB1-2 promotes. The authors describe how they can address this limitation in the future; in the meantime, is it known what the likely binding for SoxB1-2 would be (experimentally or based on homology)? If so, could the authors examine the relative abundance of SoxB1-2 binding sites in peaks that change after SoxB1-2(RNAi)? This could be compared to the abundance of the same binding sequence in non-changing peaks. Enrichment of SoxB1-2 binding sites in ATAC peaks that change after its RNAi would support the argument that chromatin changes are directly due to SoxB1-2.

We appreciate the feedback and agree that distinguishing between direct SoxB1-2 pioneer activity and indirect effects mediated through downstream regulators is an important consideration. While we did not perform a direct abundance analysis of potential chromatin-remodeling cofactors, we conducted a motif enrichment analysis following the approach of Neiro et al. (eLife, 2022), comparing control and soxB1-2(RNAi) peak sets. This analysis revealed that Sox-family motifs, particularly SoxB1-like motifs, were among the most enriched in regions that remain accessible in control animals relative to soxB1-2(RNAi) animals, consistent with a model in which SoxB1-2 directly contributes to establishing or maintaining accessibility at these loci. We have now included this analysis in the supplemental materials to further contextualize potential co-regulators and transcriptional partners within the SoxB1-2 regulatory network. We agree and acknowledge in the report that future studies assessing chromatin remodeling factor expression and abundance will be valuable to definitively separate direct and indirect pioneer activity.

Limitation 2:

The characterization of mecom and castor is somewhat preliminary relative to the deep work in the rest of the paper. I think this could be addressed with a few experiments. The authors could validate RNA-seq findings with ISH to show that cells are lost after reduction of either TF (this would support the model figure). The authors could also try to define whether loss of either TF causes behavioral phenotypes that might be similar to SoxB1-2(RNAi); this would be a second line of evidence that the TFs are downstream of key events in the SoxB1-2

pathway.

Thank you for this suggestion. We agree that additional validation of the mecom and castor RNA-seq results and further phenotypic characterization would strengthen this section. We are currently conducting in situ hybridization experiments to validate transcriptional changes in mecom and castor using the same experimental framework applied to soxB1-2 downstream candidates. We anticipate completing these studies within the next three months and will incorporate the results into future work.

Regarding behavioral phenotypes, we performed preliminary screening for robust behavioral responses, including mechanosensory responses, but did not observe overt defects. However, the lack of established, standardized behavioral assays in planarians presents a current limitation; such assays need to be developed de novo, and predicting specific behavioral phenotypes in advance remains challenging. We fully agree that functional behavioral assays represent an important next step and are actively exploring strategies to systematically develop and implement them going forward.

Other questions or comments for the authors:

Is it known how other Sox factors work as pioneer TFs? Are key binding partners known? I wondered if it would be possible to show that SoxB1-2 is co-expressed with the genes that encode these partners and/or if RNAi of these factors would phenocopy SoxB1-2. This is likely beyond the scope of this paper, but if the authors wanted to further support their argument about SoxB1-2 acting as a pioneer in planarians, this might be an additional way to do it.

In other systems, Sox pioneer factors often act together with POU family transcription factors (for example, Oct4 and Brn2) and PAX family members such as Pax6. In planarians, a POU homolog (pou-p1) is expressed in neoblasts and may represent an interesting candidate co-factor for future investigation in the context of SoxB1-2 pioneer activity. We have also previously examined the relationship between SoxB1-2 and the POU family transcription factors pou4-1 and pou4-2. Although RNAi of these factors does not fully phenocopy soxB1-2 knockdown, pou4-2(RNAi) results in loss of mechanosensation, suggesting that downstream POU factors may contribute to aspects of neural function regulated by SoxB1-2 (McCubbin et al. eLife 2025). We agree that co-expression and functional interaction studies with these candidates would be highly informative, and we view this as an exciting future direction beyond the scope of the current manuscript.

This paper is one of few to use ATAC-Seq in planarians. First, I think the authors should make a bigger deal of their generation of a dataset with this tool! Second, it would be great to know whether the ATAC-Seq data (controls and/or RNAi) will be browsable in any planarian databases or in a new website for other scientists. I believe that in addition to the data being used to test hypotheses about planarians, the data could also be a huge hypothesis generating resource in the planarian community, so I would encourage the authors to both self-promote their contribution and make plans to share it as widely and usably as possible.

Thank you very much for this encouraging feedback. We appreciate the suggestion and have strengthened the text to emphasize the significance of generating this ATAC-seq resource for the planarian field. We agree that these datasets represent a valuable community resource and are committed to making all control and soxB1-2(RNAi) ATAC-seq data publicly accessible.

Reviewer #2 (Significance (Required)):

This paper's strengths are that it addresses an important problem in regenerative biology in a rigorous manner. The writing and presentation of the data are excellent. The paper also provides excellent datasets that will be very useful to other researchers in the fi eld. Finally, the work is one of, if not the first to examine how the action of one transcription factor in planarians leads to changes in the cellular and chromatin environment that could then be acted upon by subsequent factors. This is an important contribution to the planarian fi eld, but also one that will be useful for other developmental neuroscientists and regenerative biologists.

I described a couple of limitations in the review above, but the strengths outweigh the weaknesses.

Reviewer #3 (Evidence, reproducibility and clarity (Required)):

The authors investigated the role of soxB1-2 in planarian neural and epidermal lineage specification. Using ATAC-seq and RNA-seq from head fragments after soxB1-2 RNAi, they identified regions of decreased chromatin accessibility and reduced gene expression, demonstrating that soxB1-2 induces neural and sensory programs. Integration of the datasets yielded 31 overlapping candidate targets correlating ATAC-seq and RNA-seq. Downstream analyses of transcription factors that had either/or differentially accessible regulatory region or showed differential expression (castor and mecom) implicated these transcription factors in mechanosensory and ciliary modules. The authors combined additional techniques, such as in situ hybridization to support the observations based on the ATACseq/RNAseq data. The manuscript is clearly written as well as data presentation in the main and supplementary figures. The major claim of the manuscript is that SoxB1-2 is likely a pioneer transcription factor that alters the accessibility of the chromatin, which if true, would be one of the first demonstrations of direct transcriptional regulation in planarians. As described below, I am not certain that this interpretation of the data is more valid than alternative interpretations.

Major comments

1. Direct vs. indirect regulation. The current analysis does not distinguish between direct and indirect soxB1-2 targets, therefore, this analysis cannot indicate whether soxB1-2 functions as a pioneer transcription. ATAC-seq and RNA-seq, as performed here, do not determine whether reduced accessibility or downregulation of gene expression represents a change within existing cells or a reduction in the proportion of specific cell types in the libraries produced. This limitation should be explicitly recognized where causal statements are made. In fact, several pieces of information strongly suggest that indirect effects are abundant in the data: (1) the observed loss of accessibility and gene expression in late epidermal progenitors likely represent indirect effects, indicating that within the timeframe of the experiment, it is impossible (using these techniques) to distinguish between the scenarios. (2) The finding that castor knockdown reduces soxB1-2 expression likely reflects population loss rather than direct regulation, given overlapping expression domains. This further illustrates the difficulty in inferring directionality from such datasets. In order to provide evidence for a more direct association between soxB1-2 and the differentially accessible chromatin regions, a sequence(e.g., motif) analysis would be required. Other approaches to infer direct regulation would have been useful, but they are not available in planarians to the best of my knowledge.

We agree that distinguishing between direct SoxB1-2 pioneer activity and indirect chromatin changes mediated by downstream factors is an important consideration. As suggested, examining the enrichment of SoxB1-2 binding motifs in regions that lose accessibility following soxB1-2(RNAi) can provide supporting evidence for direct regulation.

While we did not conduct a direct abundance analysis of all potential chromatin-remodeling cofactors, we performed a motif enrichment analysis following the methodology of Neiro et al. (eLife, 2022), comparing control-specific and soxB1-2(RNAi)-specific accessible peak sets. Consistent with a direct role for SoxB1-2 in chromatin regulation, Sox-family motifs, particularly SoxB1-like motifs, were among the most significantly enriched in regions that maintain accessibility in control animals relative to soxB1-2(RNAi) animals.

Evidence for pioneer activity. The authors correctly acknowledge that they do not present direct evidence of soxB1-2 binding or chromatin opening. However, the section title in the Discussion could be interpreted as implying otherwise. The claim of pioneer activity should remain explicitly tentative until supported (at least) by motif or binding data.

We have performed suggested motif analysis and changed the language in this section to better fit the data.

Replication and dataset comparability. Both ATAC-seq and soxB1-2 RNA-seq were performed on head fragments, but the number of replicates differ between assays (ATAC-seq n=2 per group, RNA-seq n=4-6). This is of course acceptable, but when interpreting the results, it should be taken into consideration that the statistical power is different when using data collected using different techniques and having a varied number of replicates.

Thank you for raising this important point regarding replication and comparability across datasets. We agree that the differing number of biological replicates between the ATAC-seq and RNA-seq experiments results in different statistical power across assays. We have now clarified this consideration in the manuscript text.

Minor comments

"Thousands of accessible chromatin sites". Please state the number of peaks and the thresholds for calling them. Ensure consistency between text (264 DA peaks) and Figure 1 legend (269 DA peaks).

__We have clarified specific peak numbers and will include the calling parameters in the methods section. Additionally, we will fix the discrepancies between differential peaks. __

Specify the y-axis normalization units in all coverage plots.

We have specified this across plots.

Clarify replicate numbers consistently in the text and figure legends.

We have identified and corrected discrepancies in the figure legends vs text and correct them and ensured they are included consistently across datasets.

Referees cross commenting

The reviews are highly consistent. They recognize the value of the work, and raise similar points. The main shared view is that the current data do not distinguish direct from indirect effects, and claims about pioneer activity should be softened, and further analysis of the differentially accessible peaks could strengthen the link between SoxB1-2 and the chromatin changes.

-I don't think that it's necessary to further characterize experimentally mecom or castor (as suggested), but of course that it could have value.

We thank all three reviewers for their positive assessment of the value of our work aiming to elucidate mechanisms by which SoxB1-2 programs planarian stem cells. In the revision, we have improved the presentation and carefully edited conclusions about the function of SoxB1-2. Performing motif analysis and GO annotation of upregulated genes has strengthened our observation that SoxB1-2 acts as an activator and has revealed putative binding sites.

The preliminary revision does not yet include further characterization of mecom and castor downstream genes. In response to Reviewer #2, we appreciate that additional validation of the mecom and castor RNA-seq results and further phenotypic characterization would strengthen this section. Although we are currently conducting in situ hybridization experiments to validate transcriptional changes in mecom and castor using the same experimental framework applied to soxB1-2 downstream candidates, we also reconsidered, as we did in our first revision, whether this is necessary or better suited for future investigations.

In the revision, we noted that our Discussion points were not balanced and that we emphasized the mecom and castor results in a manner that distracted from the major focus of the work, likely contributing to the impression that additional experimental evidence was required. Therefore, we have revised the section accordingly and streamlined the Discussion to avoid repetitive statements and to focus on the insights gained into the mechanism of SoxB1-2 function in planarian neurogenesis. We remain open to including these additional experiments if the reviewers or handling editors consider them essential; however, we agree that their inclusion is not absolutely necessary.

Reviewer #3 (Significance (Required)):

General assessment. The study offers valuable observations by combining chromatin and transcriptional analysis of planarian neural differentiation. The integration with in situ validation convincingly demonstrates effects on neural tissues and provides a solid resource for future functional work. However, mechanistic interpretation remains limited, partly because of technical limitations of the system. The data support an important role for soxB1-2 in neural and epidermal lineage regulation, but not direct binding or chromatin-opening activity. The authors have previously published analysis of soxB1-2 in planarians, so the addition of ATAC-seq data contributes to solving another piece of the puzzle.

__Advance. __

This is one of the first studies to couple ATAC-seq and RNA-seq in planarian tissue to dissect regulatory logic during regeneration. It identifies new candidate regulators of sensory and epidermal differentiation and identifies soxB1-2 as a likely upstream factor in ectodermal lineage networks. The work extends previous studies on soxB1-2 activity and neural cell production by integrating chromatin and transcriptional layers. In that respect the results are very solid, although the study remains correlative at the mechanistic level.

Audience.

This work will potentially interest researchers interested in regeneration and transcriptional networks. The datasets and gene lists will be valuable references for follow-up studies on planarian ectodermal lineages, and therefore will appeal to this community.

Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

Learn more at Review Commons

Referee #3

Evidence, reproducibility and clarity

The authors investigated the role of soxB1-2 in planarian neural and epidermal lineage specification. Using ATAC-seq and RNA-seq from head fragments after soxB1-2 RNAi, they identified regions of decreased chromatin accessibility and reduced gene expression, demonstrating that soxB1-2 induces neural and sensory programs. Integration of the datasets yielded 31 overlapping candidate targets correlating ATAC-seq and RNA-seq. Downstream analyses of transcription factors that had either/or differentially accessible regulatory region or showed differential expression (castor and mecom) implicated these transcription factors in mechanosensory and ciliary modules. The authors combined additional techniques, such as in situ hybridization to support the observations based on the ATACseq/RNAseq data. The manuscript is clearly written as well as data presentation in the main and supplementary figures. The major claim of the manuscript is that SoxB1-2 is likely a pioneer transcription factor that alters the accessibility of the chromatin, which if true, would be one of the first demonstrations of direct transcriptional regulation in planarians. As described below, I am not certain that this interpretation of the data is more valid than alternative interpretations.

Major comments

1. Direct vs. indirect regulation. The current analysis does not distinguish between direct and indirect soxB1-2 targets, therefore, this analysis cannot indicate whether soxB1-2 functions as a pioneer transcription. ATAC-seq and RNA-seq, as performed here, do not determine whether reduced accessibility or downregulation of gene expression represents a change within existing cells or a reduction in the proportion of specific cell types in the libraries produced. This limitation should be explicitly recognized where causal statements are made. In fact, several pieces of information strongly suggest that indirect effects are abundant in the data: (1) the observed loss of accessibility and gene expression in late epidermal progenitors likely represent indirect effects, indicating that within the timeframe of the experiment, it is impossible (using these techniques) to distinguish between the scenarios. (2) The finding that castor knockdown reduces soxB1-2 expression likely reflects population loss rather than direct regulation, given overlapping expression domains. This further illustrates the difficulty in inferring directionality from such datasets. In order to provide evidence for a more direct association between soxB1-2 and the differentially accessible chromatin regions, a sequence (e.g., motif) analysis would be required. Other approaches to infer direct regulation would have been useful, but they are not available in planarians to the best of my knowledge.
2. Evidence for pioneer activity. The authors correctly acknowledge that they do not present direct evidence of soxB1-2 binding or chromatin opening. However, the section title in the Discussion could be interpreted as implying otherwise. The claim of pioneer activity should remain explicitly tentative until supported (at least) by motif or binding data.
3. Replication and dataset comparability. Both ATAC-seq and soxB1-2 RNA-seq were performed on head fragments, but the number of replicates differ between assays (ATAC-seq n=2 per group, RNA-seq n=4-6). This is of course acceptable, but when interpreting the results, it should be taken into consideration that the statistical power is different when using data collected using different techniques and having a varied number of replicates.

Minor comments

"Thousands of accessible chromatin sites". Please state the number of peaks and the thresholds for calling them. Ensure consistency between text (264 DA peaks) and Figure 1 legend (269 DA peaks). Specify the y-axis normalization units in all coverage plots. Clarify replicate numbers consistently in the text and figure legends.

Referees cross commenting

The reviews are highly consistent. They recognize the value of the work, and raise similar points. The main shared view is that the current data do not distinguish direct from indirect effects, and claims about pioneer activity should be softened, and further analysis of the differentially accessible peaks could strengthen the link between SoxB1-2 and the chromatin changes.

• I don't think that it's necessary to further characterize experimentally mecom or castor (as suggested), but of course that it could have value.

Significance

General assessment. The study offers valuable observations by combining chromatin and transcriptional analysis of planarian neural differentiation. The integration with in situ validation convincingly demonstrates effects on neural tissues and provides a solid resource for future functional work. However, mechanistic interpretation remains limited, partly because of technical limitations of the system. The data support an important role for soxB1-2 in neural and epidermal lineage regulation, but not direct binding or chromatin-opening activity. The authors have previously published analysis of soxB1-2 in planarians, so the addition of ATAC-seq data contributes to solving another piece of the puzzle.

Advance. This is one of the first studies to couple ATAC-seq and RNA-seq in planarian tissue to dissect regulatory logic during regeneration. It identifies new candidate regulators of sensory and epidermal differentiation and identifies soxB1-2 as a likely upstream factor in ectodermal lineage networks. The work extends previous studies on soxB1-2 activity and neural cell production by integrating chromatin and transcriptional layers. In that respect the results are very solid, although the study remains correlative at the mechanistic level.

Audience. This work will potentially interest researchers interested in regeneration and transcriptional networks. The datasets and gene lists will be valuable references for follow-up studies on planarian ectodermal lineages, and therefore will appeal to this community.

Core Themes: Key idea; monarchy as national symbol

Contributes to the terrorist attack narrative, highlighting the fact that this was a Tesla in front of a Trump hotel, where the two work together, and the article shows that.

[INT] On the third day of Bartleby's employment, the lawyer needs him to revise a document in terms of naturalized, capitalist demands: fast, so he. as the boss, could proceed with other (more meaningful) tasks "without the least delay."

Recurring theme shown with:<br /> Loopt<br /> Y Combinator<br /> Elon Musk

Reviewer #2 (Public review):

Summary:

The authors report describes a novel vaccine platform derived from a newly discovered organelle called a migrasome. First, the authors address a technical hurdle for using migrasomes as a vaccine platform. Natural migrasome formation occurs at low levels and is labor intensive, however, by understanding the molecular underpinning of migrasome formation, the authors have designed a method to make engineered migrasomes from cultures cells at higher yields utilizing a robust process. These engineered migrasomes behave like natural migrasomes. Next, the authors immunized mice with migrasomes that either expressed a model peptide or the SARS-CoV-2 spike protein. Antibodies against the spike protein were raised that could be boosted by a 2nd vaccination and these antibodies were functional as assessed by an in vitro pseudoviral assay. This new vaccine platform has the potential to overcome obstacles such as cold chain issues for vaccines like messenger RNA that require very stringent storage conditions.

Strengths:

The authors present very robust studies detailing the biology behind migrasome formation and this fundamental understanding was used to from engineered migrasomes, which makes it possible to utilize migrasomes as a vaccine platform. The characterization of engineered migrasomes is thorough and establishes comparability with naturally occurring migrasomes. The biophysical characterization of the migrasomes is well done, including thermal stability and characterization of the particle size (important characterizations for a good vaccine).

Weaknesses:

With a new vaccine platform technology, it would be nice to compare them head-to-head against a proven technology. The authors would improve the manuscript if they made some comparisons to other vaccine platforms such as a SARS-CoV-2 mRNA vaccine or even an adjuvanted recombinant spike protein. This would demonstrate a migrasome based vaccine could elicit responses comparable to a proven vaccine technology. Additionally, understanding the integrity of the antigens expressed in their migrasomes could be useful. This could be done by looking at functional monoclonal antibody binding to their migrasomes in a confocal microscopy experiment.

Updates after revision:

The revised manuscript has additional experiments that I believe improve the strength of evidence presented in the manuscript and address the weaknesses of the first draft. First, they provide a comparison to the antibody responses induced by their migrasome based platform to recombinant protein formulated in an adjuvant and show the response is comparable. Second, they provide evidence that the spike protein incorporated into their migrasomes retains structural integrity by preserving binding to monoclonal antibodies. Together, these results strengthen the paper significantly and support the claims that the novel migrasome based vaccine platform could be a useful in the vaccine development field.

Johnston keeps up a strict rhyme scheme throughout the poem - aabbccbb, etc. However, the speaker's somewhat fourth wall breaking exclamation here tips that rhyme scheme on its head. Blin', in the Scots dialect, is a near rhyme with the word him, though that rhyme is lost some in other accents. Beyond the loose rhyme here, the outburst also changes the otherwise even flow of the rhythm through the poem.

The thought of her child falling to the floor forces the speaker out of her careful patterns, highlighting the mother's love and care for her children.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review):

This study presents an exploration of PPGL tumour bulk transcriptomics and identifies three clusters of samples (labeled as subtypes C1-C3). Each subtype is then investigated for the presence of somatic mutations, metabolism-associated pathways and inflammation correlates, and disease progression. The proposed subtype descriptions are presented as an exploratory study. The proposed potential biomarkers from this subtype are suitably caveated and will require further validation in PPGL cohorts together with a mechanistic study.  

The first section uses WGCNA (a method to identify clusters of samples based on gene expression correlations) to discover three transcriptome-based clusters of PPGL tumours. The second section inspects a previously published snRNAseq dataset, and labels some of the published cells as subtypes C1, C2, C3 (Methods could be clarified here), among other cells labelled as immune cell types. Further details about how the previously reported single-nuclei were assigned to the newly described subtypes C1-C3 require clarification.

Thank you for your valuable suggestion. In response to the reviewer’s request for further clarification on “how previously published single-nuclei data were assigned to the newly defined C1-C3 subtypes,” we have provided additional methodological details in the revised manuscript (lines 103-109). Specifically, we aggregated the single-nucleus RNA-seq data to the sample level by summing gene counts across nuclei to generate pseudo-bulk expression profiles. These profiles were then normalized for library size, log-transformed (log1p), and z-scaled across samples. Using genesets scores derived from our earlier WGCNA analysis of PPGLs, we defined transcriptional subtypes within the Magnus cohort (Supplementary Figure. 1C). We further analyzed the single-nucleus data by classifying malignant (chromaffin) nuclei as C1, C2, or C3 based on their subtype scores, while non-malignant nuclei (including immune, stromal, endothelial, and others) were annotated using canonical cell-type markers (Figure. 4A). 

The tumour samples are obtained from multiple locations in the body (Figure 1A). It will be important to see further investigation of how the sample origin is distributed among the C1C3 clusters, and whether there is a sample-origin association with mutational drivers and disease progression.

Thank you for your valuable suggestion. In the revised manuscript (lines 74-79), Figure. 1A, Table S1 and Supplementary Figure. 1A, we harmonized anatomic site annotations from our PPGL cohort and the TCGA cohort and analyzed the distribution of tumor origin (adrenal vs extra-adrenal) across subtypes. The site composition is essentially uniform across C1-C3— approximately 75% pheochromocytoma (PC) and 25% paraganglioma (PG)—with only minimal variation. Notably, the proportion of extra-adrenal origin (paraganglioma origin) is slightly higher in the C1 subtype (see Supplementary Figure 1A), which aligns with the biological characteristics of tumors from this anatomical site, which typically exhibit more aggressive behavior.

Reviewer #2 (Public Review):

A study that furthers the molecular definition of PPGL (where prognosis is variable) and provides a wide range of sub-experiments to back up the findings. One of the key premises of the study is that identification of driver mutations in PPGL is incomplete and that compromises characterisation for prognostic purposes. This is a reasonable starting point on which to base some characterisation based on different methods. The cohort is a reasonable size, and a useful validation cohort in the form of TCGA is used. Whilst it would be resource-intensive (though plausible given the rarity of the tumour type) to perform RNA-seq on all PPGL samples in clinical practice, some potential proxies are proposed.

We sincerely thank the reviewer for their positive assessment of our study’s rationale. We fully agree that RNA sequencing for all PPGL samples remains resource-intensive in current clinical practice, and its widespread application still faces feasibility challenges. It is precisely for this reason that, after defining transcriptional subtypes, we further focused on identifying and validating practical molecular markers and exploring their detectability at the protein level.

In this study, we validated key markers such as ANGPT2, PCSK1N, and GPX3 using immunohistochemistry (IHC), demonstrating their ability to effectively distinguish among molecular subtypes (see Figure. 5). This provides a potential tool for the clinical translation of transcriptional subtyping, similar to the transcription factor-based subtyping in small cell lung cancer where IHC enables low-cost and rapid molecular classification.

It should be noted that the subtyping performance of these markers has so far been preliminarily validated only in our internal cohort of 87 PPGL samples. We agree with the reviewer that largerscale, multi-center prospective studies are needed in the future to further establish the reliability and prognostic value of these markers in clinical practice.

The performance of some of the proxy markers for transcriptional subtype is not presented.

We agree with your comment regarding the need to further evaluate the performance of proxy markers for transcriptional subtyping. In our study, we have in fact taken this point into full consideration. To translate the transcriptional subtypes into a clinically applicable classification tool, we employed a linear regression model to compare the effect values (β values) of candidate marker genes across subtypes (Supplementary Figure. 1D-F). Genes with the most significant β values and statistical differences were selected as representative markers for each subtype.

Ultimately, we identified ANGPT2, PCSK1N, and GPX3—each significantly overexpressed in subtypes C1, C2, and C3, respectively, and exhibiting the most pronounced β values—as robust marker genes for these subtypes (Figure. 5A and Supplementary Figure. 1D-F). These results support the utility of these markers in subtype classification and have been thoroughly validated in our analysis.

There is limited prognostic information available.

Thank you for your valuable suggestion. In this exploratory revision, we present the available prognostic signal in Figure. 5C. Given the current event numbers and follow-up time, we intentionally limited inference. We are continuing longitudinal follow-up of the PPGL cohort and will periodically update and report mature time-to-event analyses in subsequent work.

Reviewer #1 (Recommendations for the authors):

There is no deposition reference for the RNAseq transcriptomics data. Have the data been deposited in a suitable data repository?

Thank you for your valuable suggestion. We have updated the Data availability section (lines 508–511) to clarify that the bulk-tissue RNA-seq datasets generated in this study are available from the corresponding author upon reasonable request.

In the snRNAseq analysis of existing published data, clarify how cells were labelled as "C1", "C2", "C3", alongside cells labelled by cell type (the latter is described briefly in the Methods).

Thank you for your valuable suggestion. In response to the reviewer’s request for further clarification on “how previously published single-nuclei data were assigned to the newly defined C1-C3 subtypes,” we have provided additional methodological details in the revised manuscript (lines 103-109). Specifically, we aggregated the single-nucleus RNA-seq data to the sample level by summing gene counts across nuclei to generate pseudo-bulk expression profiles. These profiles were then normalized for library size, log-transformed (log1p), and z-scaled across samples. Using genesets scores derived from our earlier WGCNA analysis of PPGLs, we defined transcriptional subtypes within the Magnus cohort (Supplementary Figure. 1C). We further analyzed the single-nucleus data by classifying malignant (chromaffin) nuclei as C1, C2, or C3 based on their subtype scores, while non-malignant nuclei (including immune, stromal, endothelial, and others) were annotated using canonical cell-type markers (Figure. 4A).

Package versions should be included (e.g., CellChat, monocle2).

We greatly appreciate your comments and have now added a dedicated “Software and versions” subsection in Methods. Specifically, we report Seurat (v4.4.0), sctransform (v0.4.2), CellChat (v2.2.0), monocle (v2.36.0; monocle2), pheatmap (v1.0.13), clusterProfiler (v4.16.0), survival (v3.8.3), and ggplot2 (v3.5.2) (lines 514-516). We also corrected a typographical error (“mafools” → “maftools”) (lines 463).

Reviewer #2 (Recommendations for the authors):

It would be helpful to provide a little more detail on the clinical composition of the cohort (e.g., phaeo vs paraganglioma, age, etc.) in the text, acknowledging that this is done in Figure 1.

Thank you for your valuable suggestion. In the revision, we added Table S1 that provides a detailed summary of the clinical composition of the PPGL cohort. Specifically, we report the numbers and proportions (Supplementary Figure. 1A) of pheochromocytoma (PC) versus paraganglioma (PG), further subclassifying PG into head and neck (HN-PG), retroperitoneal (RPPG), and bladder (BC-PG).

How many of each transcriptional subtype had driver mutations (germline or somatic)? This is included in the figures but would be worth mentioning in the text. Presumably, some of these may be present but not detected (e.g., non-coding variants), and this should be commented on. It is feasible that if methods to detect all the relevant genomic markers were improved, then the rate of tumours without driver mutations would be less and their prognostic utility would be more comprehensive.

Thank you for your valuable suggestion. In the revision (lines 113–116), we now report the prevalence of driver mutations (germline or somatic) overall and by transcriptional subtype. We analyzed variant data across 84 PPGL-relevant genes from 179 tumors in the TCGA cohort and 30 tumors in Magnus’s cohort (Fig. 2A; Table S2). High-frequency genes were consistent with known biology—C1 enriched for [e.g., VHL/SDHB], C2 for [e.g., RET/HRAS], and C3 for [e.g., SDHA/SDHD]. We also note that a subset of tumors lacked an identifiable driver, which likely reflects current assay limitations (e.g., non-coding or structural variants, subclonality, and purity effects). Broader genomic profiling (deep WGS/long-read, RNA fusion, methylation) would be expected to reduce the “driver-negative” fraction and further enhance the prognostic utility of these classifiers.

ANGPT2 provides a reasonable predictive capacity for the C1 subtype as defined by the ROC AUC. What was the performance of the PCSK1N and GPX3 as markers of the other subtypes?

We agree with your comment regarding the need to further evaluate the performance of proxy markers for transcriptional subtyping, and we have supplemented the analysis with ROC and AUC values for two additional parameters (Author response image 1 , see below). Furthermore, in our study, we have in fact taken this point into full consideration. To translate the transcriptional subtypes into a clinically applicable classification tool, we employed a linear regression model to compare the effect values (β values) of candidate marker genes across subtypes (Supplementary Figure. 1D-F). Genes with the most significant β values and statistical differences were selected as representative markers for each subtype.

Ultimately, we identified ANGPT2, PCSK1N, and GPX3—each significantly overexpressed in subtypes C1, C2, and C3, respectively, and exhibiting the most pronounced β values—as robust marker genes for these subtypes (Figure. 5A and Supplementary Figure. 1D-F). These results support the utility of these markers in subtype classification and have been thoroughly validated in our analysis.

Author response image 1.

Extended Data Figure A-B. (A) The ROC curve illustrates the diagnostic ability to distinguish PCSK1N expression in PPGLs, specifically differentiating subtype C2 from non-C2 subtypes. The red dot indicates the point with the highest sensitivity (93.1%) and specificity (82.8%). AUC, the area under the curve. (B) The ROC curve illustrates the diagnostic ability to distinguish GPX3 expression in PPGLs, specifically differentiating subtype C3 from non-C3 subtypes. The red dot indicates the point with the highest sensitivity (83.0%) and specificity (58.8%). AUC, the area under the curve.

In the discussion, I think it would be valuable to summarise existing clinical/molecular predictors in PPGL and, acknowledging that their performance may be limited, compare them to the potential of these novel classifiers.

Thank you for your valuable suggestion. We have added a concise overview of established clinical and molecular predictors in PPGL and compared them with the potential of our transcriptional classifiers. The new paragraph (Discussion, lines 315–338) now reads:

“Compared to existing clinical and molecular predictors, risk assessment in PPGL has long relied on the following indicators: clinicopathological features (e.g., tumor size, non-adrenal origin, specific secretory phenotype, Ki-67 index), histopathological scoring systems (such as PASS/GAPP), and certain genetic alterations (including high-risk markers like SDHB inactivation mutations, as well as susceptibility gene mutations in ATRX, TERT promoter, MAML3, VHL, NF1, among others). Although these metrics are highly actionable in clinical practice, they exhibit several limitations: first, current molecular markers only cover a subset of patients, and technical constraints hinder the detection of many potentially significant variants (e.g., non-coding mutations), thereby compromising the comprehensiveness of prognostic evaluation; second, histopathological scoring is susceptible to interobserver variability; furthermore, the lack of standardized detection and evaluation protocols across institutions limits the comparability and generalizability of results. Our transcriptomic classification system—comprising C1 (pseudohypoxic/angiogenic signature), C2 (kinase-signaling signature), and C3 (SDHx-related signature)—provides a complementary approach to PPGL risk assessment. These subtypes reflect distinct biological backgrounds tied to specific genetic alterations and can be approximated by measuring the expression of individual genes (e.g., ANGPT2, PCSK1N, or GPX3). This study demonstrates that the classifier offers three major advantages: first, it accurately distinguishes subtypes with coherent biological features; second, it retains significant predictive value even after adjusting for clinical covariates; third, it can be implemented using readily available assays such as immunohistochemistry. These findings suggest that integrating transcriptomic subtyping with conventional clinical markers may offer a more comprehensive and generalizable risk stratification framework. However, this strategy would require validation through multi-center prospective studies and standardization of detection protocols.”

A little more explanation of the principles behind WGCNA would be useful in the methods.

We are grateful for your comments. We have expanded the Methods to briefly explain the principles of WGCNA (lines 426-454). In short, WGCNA constructs a weighted coexpression network from normalized gene expression, identifies modules of tightly co-expressed genes, summarizes each module by its eigengene (the first principal component), and then correlates module eigengenes with phenotypes (e.g., transcriptional subtypes) to highlight biologically meaningful gene sets and candidate hub genes. We now specify our preprocessing, choice of softthresholding power to approximate scale-free topology, module detection/merging criteria, and the statistics used for module–trait association and downstream gene-set scoring. 

On line 234, I think the figure should be 5C?

We greatly appreciate your comments and Correct to Figure 5C.

Guenevere's accusation suggest that Sir Gauwaine cannot claim moral superiority, as his own family history is fraught with similar transgression. This highlights the recurring theme of hypocrisy and flawed virtue among Arthurian knights. Furthermore, Guenevere is referring to the affair of Sir Gauwaine's mother, Morgause. She was killed by her son, Gaheris, when he discovered her relationship with Sir Lamorak. https://kingarthursknights.com/arthurian-characters/morgause/

Reviewer #2 (Public review):

Summary:

In this manuscript, Malfatti et al. study the role of Chrna2 Martinotti cells (Mα2 cells), a subset of SST interneurons, for motor learning and motor cortex activity. The authors trained mice on a forelimb prehension task while recording neuronal activity of pyramidal cells using calcium imaging with a head-mounted miniscope. While chemogenetically increasing Mα2 cell activity did not affect motor learning, it changed pyramidal cell activity such that activity peaks became sharper and differently timed than in control mice. Moreover, co-active neuronal assemblies become more stable with a smaller spatial distribution. Increasing Mα2 cell activity in previously trained mice did increase performance on the prehension task and led to increased theta and gamma band activity in the motor cortex. On the other hand, genetic ablation of Mα2 cells affected fine motor movements on a pasta handling task while not affecting the prehension task.

Strengths:

The proposed question of how Chrna2-expressing SST interneurons affect motor learning and motor cortex activity is important and timely. The study employs sophisticated approaches to record neuronal activity and manipulate the activity of a specific neuronal population in behaving mice over the course of motor learning. The authors analyze a variety of neuronal activity parameters, comparing different behavior trials, stages of learning, and the effects of Mα2 cell activation. The analysis of neuronal assembly activity and stability over the course of learning by tracking individual neurons throughout the imaging sessions is notable, since technically challenging, and yielded the interesting result that neuronal assemblies are more stable when activating Mα2 cells.

Overall, the study provides compelling evidence that Mα2 cells regulate certain aspects of motor behaviors, likely by shaping circuit activity in the motor cortex.

Weaknesses:

The main limitation of the study lies in its small sample sizes and the absence of key control experiments, which substantially weaken the strength of the conclusions.

Core findings of this paper, such as the lack of effect of Mα2 cell activation on motor learning, as well as the altered neuronal activity, rely ona sample size of n=3 mice per condition, which is likely underpowered to detect differences in behavior and contributes to the somewhat disconnected results on calcium activity, activity timing, and neuronal assembly activity.

More comprehensive analyses and data presentation are also needed to substantiate the results. For example, examining calcium activity and behavioral performance on a trial-by-trial basis could clarify whether closely spaced reaching attempts influence baseline signals and skew interpretation.

The study uses cre-negative mice as controls for hM3Dq-mediated activation, which does not account for potential effects of Cre-dependent viral expression that occur only in Cre-positive mice.

This important control would be necessary to substantiate the conclusion that it is increased Mα2 cell activity that drives the observed changes in behavior and cortical activity.

The deliberate refrain of "young" nature and the emphasized double "young, young children" point out the irony and tragedy of how life shouldn't be for these children. While nature frolic and play, the human children are weeping bitterly. In fact, some poems in the Victorian period use this juxtaposition of the free natural world versus the state of the oppressive poor. Thomas Hood's "Song of the Shirt" has these lines: "Oh! but to breathe the breath Of the cowslip and primrose sweet--- With the sky above my head, And the grass beneath my feet". Gerald Massey wrote in "Cry of the Unemployed": "Heaven droppeth down with manna still in many a golden shower, And feeds the leaves with fragrant breath, with silver dew, the flower; There's honeyed fruit for bee and bird, with bloom laughs out the tree". Nature is plentiful, beautiful, and free while humans suffer from hunger and fetters of their working class.

The emphasis on "first US leader" puts an emphasis on the unusualness of the matter and proximity for American readers.

This just popped into my head: If a classroom is discussing the concept of "souls," dont some (unnamed) cultures strictly believe women don't have souls? How do we justify changing what is the linguistic standard without also logically opening up the interpretation for everyone? And do we REALLY want to validate a movement that preaches women should have no autonomy or mention of a soul?

I mean, I would LOVE it. But some might not.

This is a reach of an argument, but whatever.

cesca the kitten

the voice is cesca and i am watching. her

meeeee

This implies a sense of pride and confidence in what one was doing, which reinforces that the mind is without fear, because the body is visibly showing that. This also serves to highlight that the body and mind are acting in unison.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public review):

Lack of Sensitivity Analyses for some Key Methodological Decisions: Certain methodological choices in this manuscript diverge from approaches used in previous works. In these cases, I recommend the following: (i) The authors could provide a clear and detailed justification for these deviations from established methods, and (ii) supplementary sensitivity analyses could be included to ensure the robustness of the findings, demonstrating that the results are not driven primarily by these methodological changes. Below, I outline the main areas where such evaluations are needed:

This detailed guidance is incredibly valuable, and we are grateful. Work of this kind is in its relative infancy, and there are so many design choices depending on the data available, questions being addressed, and so on. Help us navigate that has been extremely useful. In our revised manuscript we are very happy to add additional justification for design choices made, and wherever possible test the impact of those choices. It is certainly the case that different approaches have been used across the handful of papers published in this space, and, unlike in other areas of systems neuroscience, we have yet to reach the point where any of these approaches are established. We agree with the reviewer that wherever possible these design choices should be tested. 

Use of Communicability Matrices for Structural Connectivity Gradients: The authors chose to construct structural connectivity gradients using communicability matrices, arguing that diffusion map embedding "requires a smooth, fully connected matrix." However, by definition, the creation of the affinity matrix already involves smoothing and ensures full connectedness. I recommend that the authors include an analysis of what happens when the communicability matrix step is omitted. This sensitivity test is crucial, as it would help determine whether the main findings hold under a simpler construction of the affinity matrix. If the results significantly change, it could indicate that the observations are sensitive to this design choice, thereby raising concerns about the robustness of the conclusions. Additionally, if the concern is related to the large range of weights in the raw structural connectivity (SC) matrix, a more conventional approach is to apply a log-transformation to the SC weights (e.g., log(1+𝑆𝐶_𝑖𝑗)), which may yield a more reliable affinity matrix without the need for communicability measures.

The reason we used communicability is indeed partly because we wanted to guarantee a smooth fully connected matrix, but also because our end goal for this project was to explore structure-function coupling in these low-dimensional manifolds.  Structural communicability – like standard metrics of functional connectivity – includes both direct and indirect pathways, whereas streamline counts only capture direct communication. In essence we wanted to capture not only how information might be routed from one location to another, but also the more likely situation in which information propagates through the system. 

In the revised manuscript we have given a clearer justification for why we wanted to use communicability as our structural measure (Page 4, Line 179):

“To capture both direct and indirect paths of connectivity and communication, we generated weighted communicability matrices using SIFT2-weighted fibre bundle capacity (FBC). These communicability matrices reflect a graph theory measure of information transfer previously shown to maximally predict functional connectivity (Esfahlani et al., 2022; Seguin et al., 2022). This also foreshadowed our structure-function coupling analyses, whereby network communication models have been shown to increase coupling strength relative to streamline counts (Seguin et al., 2020)”.

We have also referred the reader to a new section of the Results that includes the structural gradients based on the streamline counts (Page 7, line 316):

“Finally, as a sensitivity analysis, to determine the effect of communicability on the gradients, we derived affinity matrices for both datasets using a simpler measure: the log of raw streamline counts. The first 3 components derived from streamline counts compared to communicability were highly consistent across both NKI  (r_s = 0.791, r_s = 0.866, r_s = 0.761) and the referred subset of CALM (r_s = 0.951, r_s = 0.809, r_s = 0.861), suggesting that in practice the organisational gradients are highly similar regardless of the SC metric used to construct the affinity matrices”. 

Methodological ambiguity/lack of clarity in the description of certain evaluation steps: Some aspects of the manuscript’s methodological description are ambiguous, making it challenging for future readers to fully reproduce the analyses based on the information provided. I believe the following sections would benefit from additional detail and clarification:

Computation of Manifold Eccentricity: The description of how eccentricity was computed (both in the results and methods sections) is unclear and may be problematic. The main ambiguity lies in how the group manifold origin was defined or computed. (1) In the results section, it appears that separate manifold origins were calculated for the NKI and CALM groups, suggesting a dataset-specific approach. (2) Conversely, the methods section implies that a single manifold origin was obtained by somehow combining the group origins across the three datasets, which seems contradictory. Moreover, including neurodivergent individuals in defining the central group manifold origin in conceptually problematic. Given that neurodivergent participants might exhibit atypical brain organization, as suggested by Figure 1, this inclusion could skew the definition of what should represent a typical or normative brain manifold. A more appropriate approach might involve constructing the group manifold origin using only the neurotypical participants from both the NKI and CALM datasets. Given the reported similarity between group-level manifolds of neurotypical individuals in CALM and NKI, it would be reasonable to expect that this combined origin should be close to the origin computed within neurotypical samples of either NKI or CALM. As a sanity check, I recommend reporting the distance of the combined neurotypical manifold origin to the centres of the neurotypical manifolds in each dataset. Moreover, if the manifold origin was constructed while utilizing all samples (including neurodivergent samples) I think this needs to be reconsidered. 

This is a great point, and we are very happy to clarify. Separate manifolds were calculated for the NKI and CALM participants, hence a dataset-specific approach. Indeed, in the long-run our goal was to explore individual differences in these manifolds, relative to the respective group-level origins, and their intersection across modalities, so manifold eccentricity was calculated at an individual level for subsequent analyses. At the group level, for each modality, we computed 3 manifold origins: one for NKI, one for the referred subset of CALM, and another for the neurotypical portion of CALM. Crucially, because the manifolds are always normal, in each case the manifold origin point is near-zero (extremely near-zero, to the 6^th or 7^th decimal place). In other words, we do indeed calculate the origin separately each time we calculate the gradients, but the origin is zero in every case. As a result, differences in the origin point cannot be the source of any differences we observe in manifold eccentricity between groups or individuals. We have updated the Methods section with the manifold origin points for each dataset and clarified our rationale (Page 16, Line 1296):

“Note that we used a dataset-specific approach when we computed manifold eccentricity for each of the three groups relative to their group-level origin: neurotypical CALM (SC origin = -7.698 x 10^-7, FC origin = 6.724 x 10^-7), neurodivergent CALM (SC origin = -6.422 x 10 , FC origin = 1.363 x 10 ), and NKI (SC origin = -7.434 x 10 , FC origin = 4.308 x 10^-6). Eccentricity is a relative measure and thus normalised relative to the origin. Because of this normalisation, each time gradients are constructed the manifold origin is necessarily near-zero, meaning that differences in manifold eccentricity of individual nodes, either between groups or individuals, are stem from the eccentricity of that node rather than a difference in origin point”. 

We clarified the computation of the respective manifold origins within the Results section, and referred the reader to the relevant Methods section (Page 9, line 446):

“For each modality (2 levels: SC and FC) and dataset (3 levels: neurotypical CALM, neurodivergent CALM, and NKI), we computed the group manifold origin as the mean of their respective first three gradients. Because of the normal nature of the manifolds this necessarily means that these origin points will be very near-zero, but we include the exact values in the ‘Manifold Eccentricity’ methodology sub-section”. 

Individual-Level Gradients vs. Group-Level Gradients: Unlike previous studies that examined alterations in principal gradients (e.g., Xia et al., 2022; Dong et al., 2021), this manuscript focuses on gradients derived directly from individual-level data. In contrast, earlier works have typically computed gradients based on grouped data, such as using a moving window of individuals based on age (Xia et al.) or evaluating two distinct age groups (Dong et al.). I believe it is essential to assess the sensitivity of the findings to this methodological choice. Such an evaluation could clarify whether the observed discrepancies with previous reports are due to true biological differences or simply a result of different analytical strategies.

This is a brilliant point. The central purpose of our project was to test how individual differences in these gradients, and their intersection across modalities, related to differences in phenotype (e.g. cognitive difficulties). These necessitated calculating gradients at the level of individuals and building a pipeline to do so, given that we could find no other examples. Nonetheless, despite this different goal and thus approach, we had expected to replicate a couple of other key findings, most prominently the ‘swapping’ of gradients shown by Dong et al. (2021). We were also surprised that we did not find this changing in order. The reviewer is right and there could be several design features that produce the difference, and in the revised manuscript we test several of them. We have added the following text to the manuscript as a sensitivity analysis for the Results sub-section titled “Stability of individual-level gradients across developmental time” (Page 7, Line 344 onwards):

“One possibility is that our observation of gradient stability – rather than a swapping of the order for the first two gradients (Dong et al., 2021) – is because we calculated them at an individual level. To test this, we created subgroups and contrasted the first two group-level structural and functional gradients derived from children (younger than 12 years old) versus those from adolescents (12 years old and above), using the same age groupings as prior work (Dong et al., 2021). If our use of individually calculated gradients produces the stability, then we should observe the swapping of gradients in this sensitivity analysis. Using baseline scans from NKI, the primary structural gradient in childhood (N = 99) as shown in Figure 1f, this was highly correlated (r_s = 0.995) with those derived from adolescents (N = 123). Likewise, the secondary structural gradient in childhood was highly consistent in adolescence (r_s = 0.988). In terms of functional connectivity, the principal gradient in childhood (N = 88) was highly consistent in adolescence (r_s = 0.990, N = 125). The secondary gradient in childhood was again highly similar in adolescence (r_s = 0.984). The same result occurred in the CALM dataset: In the baseline referred subset of CALM, the primary and secondary communicability gradients derived from children (N = 258) and adolescents (N = 53) were near-identical (r_s = 0.991 and r_s = 0.967, respectively). Alignment for the primary and secondary functional gradients derived from children (N = 130) and adolescents (N = 43) were also near-identical (r_s = 0.972 and r_s = 0.983, respectively). These consistencies across development suggest that gradients of communicability and functional connectivity established in childhood are the same as those in adolescence, irrespective of group-level or individual-level analysis. Put simply, our failure to replicate the swapping of gradient order in Dong et al. (2021) is not the result of calculating gradients at the level of individual participants.”

Procrustes Transformation: It is unclear why the authors opted to include a Procrustes transformation in this analysis, especially given that previous related studies (e.g., Dong et al.) did not apply this step. I believe it is crucial to evaluate whether this methodological choice influences the results, particularly in the context of developmental changes in organizational gradients. Specifically, the Procrustes transformation may maximize alignment to the group-level gradients, potentially masking individual-level differences. This could result in a reordering of the gradients (e.g., swapping the first and second gradients), which might obscure true developmental alterations. It would be informative to include an analysis showing the impact of performing vs. omitting the Procrustes transformation, as this could help clarify whether the observed effects are robust or an artifact of the alignment procedure. (Please also refer to my comment on adding a subplot to Figure 1). Additionally, clarifying how exactly the transformation was applied to align gradients across hemispheres, individuals, and/or datasets would help resolve ambiguity. 

The current study investigated individual differences in connectome organisation, rather than group-level trends (Dong et al., 2021). This necessitates aligning individual gradients to the corresponding group-level template using a Procrustes rotation. Without a rotation, there is no way of knowing if you are comparing  ‘like with like’: the manifold eccentricity of a given node may appear to change across individuals simply due to subtle differences in the arbitrary orientation of the underlying manifolds. We also note that prior work examining individual differences in principal alignment have used Procrustes (Xia et al., 2022), who demonstrated emergence of the principal gradient across development, albeit with much smaller effects than Dong and colleagues (2021). Nonetheless, we agree, the Procrustes rotation could be another source of the differences we observed with the previous paper (Dong et al. 2021). We explored the impact of the Procrustes rotation on individual gradients as our next sensitivity analysis. We recalculated everyone’s gradients without Procrustes rotation. We then tested the alignment of each participant with the group-level gradients using Spearman’s correlations, followed by a series of generalised linear models to predict principal gradient alignment using head motion, age, and sex. The expected swapping of the first and second functional gradient (Dong et al., 2021) would be represented by a decrease in the spatial similarity of each child’s principal functional gradient to the principal childhood group-level gradient, at the onset of adolescence (~age 12). However, there is no age effect on this unrotated alignment, suggesting that the lack of gradient swapping in our data does not appear to be the result of the Procrustes rotation. When you use unrotated individual gradients the alignment is remarkably consistent across childhood and adolescence. Alignment is, however, related to head motion, which is often related to age. To emphasise the importance of motion, particularly in relation to development, we conducted a mediation analysis between the relationship between age and principal alignment (without correcting for motion), with motion as a mediator, within the NKI dataset. Before accounting for motion, the relationship between age and principal alignment is significant, but this can be entirely accounted for by motion. In our revised manuscript we have included this additional analysis in the Results sub-section titled “Stability of individual-level gradients across developmental time”, following on from the above point about the effect of group-level versus individual-level analysis (Page 8, Line 400):

“A second possible discrepancy between our results and that of prior work examining developmental change in group-level functional gradients (Dong et al., 2021) was the use of Procrustes alignment. Such alignment of individual-level gradients to group-level templates is a necessary step to ensure valid comparisons between corresponding gradients across individuals, and has been implemented in sliding-window developmental work tracking functional gradient development (Xia et al., 2022). Nonetheless, we tested whether our observation of stable principal functional and communicability gradients may be an artefact of the Procrustes rotation. We did this by modelling how individual-level alignment without Procrustes rotation to the group-level templates varies with age, head motion, and sex, as a series of generalised linear models. We included head motion as the magnitude of the Procrustes rotation has been shown to be positively correlated with mean framewise displacement (Sasse et al., 2024), and prior group-level work (Dong et al., 2021) included an absolute motion threshold rather than continuous motion estimates. Using the baseline referred CALM sample, there was no significant relationship between alignment and age (β = -0.044, 95% CI = [-0.154, 0.066], p = 0.432) after accounting for head motion and sex. Interestingly, however head motion was significantly associated with alignment ( β = -0.318, 95% CI = [-0.428, -.207], p = 1.731 x 10^-8), such that greater head motion was linked to weaker alignment. Note that older children tended to have exhibit less motion for their structural scans (r_s = 0.335, p < 0.001). We observed similar trends in functional alignment, whereby tighter alignment was significantly predicted by lower head motion (β = -0.370, 95% CI = [-0.509, -0.231], p = 1.857 x 10^-7), but not by age (β= 0.049, 95% CI = [-0.090, 0.187], p = 0.490). Note that age and head motion for functional scans were not significantly related (r_s = -0.112, p = 0.137). When repeated for the baseline scans of NKI, alignment with the principal structural gradient was not significantly predicted by either scan age (β = 0.019, 95% CI = [-0.124, 0.163], p = 0.792) or head motion (β = -0.133, 95% CI = [-0.175, 0.009], p = 0.067) together in a single model, where age and motion were negatively correlated (r_s = -0.355, p < 0.001). Alignment with the principal functional gradient was significantly predicted by head motion (β = -0.183, 95% CI = [-0.329, -0.036], p = 0.014) but not by age (β= 0.066, 95% CI = [-0.081, 0.213], p = 0.377), where age and motion were also negatively correlated (r_s = -0.412, p < 0.001). Across modalities and datasets, alignment with the principal functional gradient in NKI was the only example in which there was a significant correlation between alignment and age (r_s = 0.164, p = 0.017) before accounting for head motion and sex. This suggests that apparent developmental effects on alignment are minimal, and where they do exist they are removed after accounting for head motion. Put together this suggests that the lack of order swapping for the first two gradients is not the result of the Procrustes rotation – even without the rotation there is no evidence for swapping”.

“To emphasise the importance of head motion in the appearance of developmental change in alignment, we examined whether accounting for head motion removes any apparent developmental change within NKI. Specifically, we tested whether head motion mediates the relationship between age and alignment (Figure 1X), controlling for sex, given that higher motion is associated with younger children (β= -0.429, 95% CI = [0.552, -0.305], p = 7.957 x 10^-11), and stronger alignment is associated with reduced motion (β = -0.211, 95% CI = [-0.344, -0.078], p = 2.017 x 10^-3). Motion mediated the relationship between age and alignment (β = 0.078, 95% CI = [0.006, 0.146], p = 1.200 x 10^-2), accounting for 38.5% variance in the age-alignment relationship, such that the link between age and alignment became non-significant after accounting for motion (β = 0.066, 95% CI = [-0.081, 0.214], p = 0.378). This firstly confirms our GLM analyses, where we control for motion and find no age associations. Moreover, this suggests that caution is required when associations between age and gradients are observed. In our analyses, because we calculate individual gradients, we can correct for individual differences in head motion in all our analyses. However, other than using an absolute motion threshold and motion-matched child and adolescent groups, individual differences in motion were not accounted for by prior work which demonstrated a flipping of the principal functional gradients with age (Dong et al., 2021)”. 

We further clarify the use of Procrustes rotation as a separate sub-section within the Methods (Page 25, Line 1273):

“Procrustes Rotation

For group-level analysis, for each hemisphere we constructed an affinity matrix using a normalized angle kernel and applied diffusion-map embedding. The left hemisphere was then aligned to the right using a Procrustes rotation. For individual-level analysis, eigenvectors for the left hemisphere were aligned with the corresponding group-level rotated eigenvectors. No alignment was applied across datasets. The only exception to this was for structural gradients derived from the referred CALM cohort. Specifically, we aligned the principal gradient of the left hemisphere to the secondary gradient of the right hemisphere: this was due to the first and second gradients explaining a very similar amount of variance, and hence their order was switched”. 

SC-FC Coupling Metric: The approach used to quantify nodal SC-FC coupling in this study appears to deviate from previously established methods in the field. The manuscript describes coupling as the "Spearman-rank correlation between Euclidean distances between each node and all others within structural and functional manifolds," but this description is unclear and lacks sufficient detail. Furthermore, this differs from what is typically referred to as SC-FC coupling in the literature. For instance, the cited study by Park et al. (2022) utilizes a multiple linear regression framework, where communicability, Euclidean distance, and shortest path length are independent variables predicting functional connectivity (FC), with the adjusted R-squared score serving as the coupling index for each node. On the other hand, the Baum et al. (2020) study, also cited, uses Spearman correlation, but between raw structural connectivity (SC) and FC values. If the authors opt to introduce a novel coupling metric, it is essential to demonstrate its similarity to these previous indices. I recommend providing an analysis (supplementary) showing the correlation between their chosen metric and those used in previous studies (e.g., the adjusted R-squared scores from Park et al. or the SC-FC correlation from Baum et al.). Furthermore, if the metrics are not similar and results are sensitive to this alternative metric, it raises concerns about the robustness of the findings. A sensitivity analysis would therefore be helpful (in case the novel coupling metric is not like previous ones) to determine whether the reported effects hold true across different coupling indices.

This is a great point, and we are happy to take the reviewer’s recommendation. There are multiple different ways of calculating structure-function coupling. For our set of questions, it was important that our metric incorporated information about the structural and functional manifolds, rather than being a separate approach that is unrelated to these low-dimensional embeddings. Put simply, we wanted our coupling measure to be about the manifolds and gradients outlined in the early sections of the results. We note that the multiple linear regression framework was developed by Vázquez-Rodríguez and colleagues (2019), whilst the structure-function coupling computed in manifold space by Park and colleagues (2022) was operationalised as a linear correlation between z-transformed functional connectomes and structural differentiation eigenvectors. To clarify how this coupling was calculated, and to justify why we developed a new coupling method based on manifolds rather than borrow an existing approach from the literature, we have revised the manuscript to make this far clearer for readers (Page 13, line 604):

“To examine the relationship between each node’s relative position in structural and functional manifold space, we turned our attention to structure-function coupling. Whilst prior work typically computed coupling using raw streamline counts and functional connectivity matrices, either as a correlation (Baum et al., 2020) or through a multiple linear regression framework (Vázquez-Rodríguez et al., 2019), we opted to directly incorporate low-dimensional embeddings within our coupling framework. Specifically, as opposed to correlating row-wise raw functional connectivity with structural connectivity eigenvectors (Park et al., 2022), our metric directly incorporates the relative position of each node in low-dimensional structural and functional manifold spaces. Each node was situated in a low-dimensional 3D space, the axes of which were each participant’s gradients, specific to each modality. For each participant and each node, we computed the Euclidean distance with all other nodes within structural and functional manifolds separately, producing a vector of size 200 x 1 per modality. The nodal coupling coefficient was the Spearman correlation between each node’s Euclidean distance to all other nodes in structural manifold space, and that in functional manifold space. Put simply, a strong nodal coupling coefficient suggests that that node occupies a similar location in structural space, relative to all other nodes, as it does in functional space”. 

We also agree with the reviewer’s recommendation to compare this to some of the more standard ways of calculating coupling. We compare our metric with 3 others (Baum et al., 2020; Park et al., 2022; VázquezRodríguez et al., 2019), and find that all metrics capture the core developmental sensorimotor-to-association axis (Sydnor et al., 2021). Interestingly, manifold-based coupling measures captured this axis more strongly than non-manifold measures. We have updated the Results accordingly (Page 14, Line 638):

“To evaluate our novel coupling metric, we compared its cortical spatial distribution to three others (Baum et al., 2020; Park et al., 2022; Vázquez-Rodríguez et al., 2019), using the group-level thresholded structural and functional connectomes from the referred CALM cohort. As shown in Figure 4c, our novel metric was moderately positively correlated to that of a multi-linear regression framework (r_s = 0.494, p_spin = 0.004; Vázquez-Rodríguez et al., 2019) and nodal correlations of streamline counts and functional connectivity (r_s = 0.470, p_spin = 0.005; Baum et al., 2020). As expected, our novel metric was strongly positively correlated to the manifold-derived coupling measure (r_s = 0.661, p_spin < 0.001; Park et al., 2022), more so than the first (Z(198) = 3.669, p < 0.001) and second measure (Z(198) = 4.012, p < 0.001). Structure-function coupling is thought to be patterned along a sensorimotor-association axis (Sydnor et al., 2021): all four metrics displayed weak-tomoderate alignment (Figure 4c). Interestingly, the manifold-based measures appeared most strongly aligned with the sensorimotor-association axis: the novel metric was more strongly aligned than the multi-linear regression framework (Z(198) = -11.564, p < 0.001) and the raw connectomic nodal correlation approach (Z(198) = -10.724, p < 0.001), but the previously-implemented structural manifold approach was more strongly aligned than the novel metric  (Z(198) = -12.242, p < 0.001). This suggests that our novel metric exhibits the expected spatial distribution of structure-function coupling, and the manifold approach more accurately recapitulates the sensorimotor-association axis than approaches based on raw connectomic measures”.

We also added the following to the legend of Figure 4 on page 15:

“d. The inset Spearman correlation plot of the 4 coupling measures shows moderate-to-strong correlations (p_spin < 0.005 for all spatial correlations). The accompanying lollypop plot shows the alignment between the sensorimotor-to-association axis and each of the 4 coupling measures, with the novel measure coloured in light purple (p_spin < 0.007 for all spatial correlations)”. 

Prediction vs. Association Analysis: The term “prediction” is used throughout the manuscript to describe what appear to be in-sample association tests. This terminology may be misleading, as prediction generally implies an out-of-sample evaluation where models trained on a subset of data are tested on a separate, unseen dataset. If the goal of the analyses is to assess associations rather than make true predictions, I recommend refraining from the term “prediction” and instead clarifying the nature of the analysis. Alternatively, if prediction is indeed the intended aim (which would be more compelling), I suggest conducting the evaluations using a k-fold cross-validation framework. This would involve training the Generalized Additive Mixed Models (GAMMs) on a portion of the data and training their predictive accuracy on a held-out sample (i.e. different individuals). Additionally, the current design appears to focus on predicting SC-FC coupling using cognitive or pathological dimensions. This is contrary to the more conventional approach of predicting behavioural or pathological outcomes from brain markers like coupling. Could the authors clarify why this reverse direction of analysis was chosen? Understanding this choice is crucial, as it impacts the interpretation and potential implications of the findings. 

We have replaced “prediction” with “association” across the manuscript. However, for analyses corresponding to Figure 5, which we believe to be the most compelling, we conducted a stratified 5-fold cross-validation procedure, outlined below, repeated 100 times to account for random variation in the train-test splits. To assess whether prediction accuracy in the test splits was significantly greater than chance, we compared our results to those derived from a null dataset in which cognitive factor 2 scores had been permuted across participants. To account for the time-series element and block design of our data, in that some participants had 2 or more observations, we permuted entire participant blocks of cognitive factor 2 scores, keeping all other variables, including covariates, the same. Included in our manuscript are methodological details and results pertaining to this procedure. Specifically, the following has been added to the Results (Page 16, Line 758):

“To examine the predictive value of the second cognitive factor for global and network-level structure-function coupling, operationalised as a Spearman rank correlation coefficient, we implemented a stratified 5-fold crossvalidation framework, and predictive accuracy compared with that of a null data frame with cognitive factor 2 scores permuted across participant blocks (see ‘GAMM cross-validation’ in the Methods). This procedure was repeated 100 times to account for randomness in the train-test splits, using the same model specification as above. Therefore, for each of the 5 network partitions in which an interaction between the second cognitive factor and age was a significant predictor of structure-function coupling (global, visual, somato-motor, dorsal attention, and default-mode), we conducted a Welch’s independent-sample t-test to compare 500 empirical prediction accuracies with 500 null prediction accuracies. Across all 5 network partitions, predictive accuracy of coupling was significantly higher than that of models trained on permuted cognitive factor 2 scores (all p < 0.001). We observed the largest difference between empirical (M = 0.029, SD = 0.076) and null (M = -0.052, SD = 0.087) prediction accuracy in the somato-motor network [t (980.791) = 15.748, p < 0.001, Cohen’s d = 0.996], and the smallest difference between empirical (M = 0.080, SD = 0.082) and null (M = 0.047, SD = 0.081) prediction accuracy in the dorsal attention network [t (997.720) = 6.378, p < 0.001, Cohen’s d = 0.403]. To compare relative prediction accuracies, we ordered networks by descending mean accuracy and conducted a series of Welch’s independent sample t-tests, followed by FDR correction (Figure 5X). Prediction accuracy was highest in the default-mode network (M = 0.265, SD = 0.085), two-fold that of global coupling (t(992.824) = 25.777, p_FDR = 5.457 x 10^-112, Cohen’s d = 1.630, M = 0.131, SD = 0.079). Global prediction accuracy was significantly higher than the visual network (t (992.644) = 9.273, p_FDR = 1.462 x 10^-19, Cohen’s d = 0.586, M = 0.083, SD = 0.085), but visual prediction accuracy was not significantly higher than within the dorsal attention network (t (997.064) = 0.554, p_FDR = 0.580, Cohen’s d = 0.035, M = 0.080, SD = 0.082). Finally, prediction accuracy within the dorsal attention network was significantly stronger than that of the somato-motor network [t (991.566) = 10.158, p_FDR = 7.879 x 10^-23, Cohen’s d = 0.642 M = 0.029, SD = 0.076]. Together, this suggests that out-of-sample developmental predictive accuracy for structure-function coupling, using the second cognitive factor, is strongest in the higher-order default-mode network, and lowest in the lower-order somatosensory network”. 

We have added a separate section for GAMM cross-validation in the Methods (Page 27, Line 1361):

GAMM cross-validation

“We implemented a 5-fold cross validation procedure, stratified by dataset (2 levels: CALM or NKI). All observations from any given participant were assigned to either the testing or training fold, to prevent data leakage, and the cross-validation procedure was repeated 100 times, to account for randomness in data splits. The outcome was predicted global or network-level structure-function coupling across all test splits, operationalised as the Spearman rank correlation coefficient. To assess whether prediction accuracy exceeded chance, we compared empirical prediction accuracy with that of GAMMs trained and tested on null data in which cognitive factor 2 scores were permuted across subjects. The number of observations formed 3 exchangeability blocks (N = 320 with one observation, N = 105 with two observations, and N = 33 with three observations), whereby scores from a participant with two observations were replaced by scores from another participant with two observations, with participant-level scores kept together, and so on for all numbers of observations. We compared empirical and null prediction accuracies using independent sample t-tests as, although the same participants were examined, the shuffling meant that the relative ordering of participants within both distributions was not preserved. For parallelisation and better stability when estimating models fit on permuted data, we used the bam function from the mgcv R package (Wood, 2017)”. 

We also added a justification for why we predicted coupling using behaviour or psychopathology, rather than vice versa (Page 27, Line 1349):

“When using our GAMMs to test for the relationship between cognition and psychopathology and our coupling metrics, we opted to predict structure-function coupling using cognitive or psychopathological dimensions, rather than vice versa, to minimise multiple comparisons. In the current framework, we corrected for 8 multiple comparisons within each domain. This would have increased to 16 multiple comparison corrections for predicting two cognitive dimensions using network-level coupling, and 24 multiple comparison corrections for predicting three psychopathology dimensions. Incorporating multiple networks as predictors within the same regression framework introduces collinearity, whilst the behavioural dimensions were orthogonal: for example, coupling is strongly correlated between the somato-motor and ventral attention networks (r_s = 0.721), between the default-mode and frontoparietal networks (r_s = 0.670), and between the dorsal attention and fronto-parietal networks (r_s = 0.650)”. 

Finally, we noticed a rounding error in the ages of the data frame containing the structure-function coupling values and the cognitive/psychopathology dimensions. We rectified this and replaced the GAMM results, which largely remained the same. 

In typical applications of diffusion map embedding, sparsification (e.g., retaining only the top 10  of the strongest connections) is often employed at the vertex-level resolution to ensure computational feasibility. However, since the present study performs the embedding at the level of 200 brain regions (a considerably coarser resolution), this step may not be necessary or justifiable. Specifically, for FC, it might be more appropriate to retain all positive connections rather than applying sparsification, which could inadvertently eliminate valuable information about lower-strength connections. Whereas for SC, as the values are strictly non-negative, retaining all connections should be feasible and would provide a more complete representation of the structural connectivity patterns. Given this, it would be helpful if the authors could clarify why they chose to include sparsification despite the coarser regional resolution, and whether they considered this alternative approach (using all available positive connections for FC and all non-zero values for SC). It would be interesting if the authors could provide their thoughts on whether the decision to run evaluations at the resolution of brain regions could itself impact the functional and structural manifolds, their alteration with age, and or their stability (in contrast to Dong et al. which tested alterations in highresolution gradients).

This is another great point. We could retain all connections, but we usually implement some form of sparsification to reduce noise, particularly in the case of functional connectivity. But we nonetheless agree with the reviewer’s point. We should check what impact this is having on the analysis. In brief, we found minimal effects of thresholding, suggesting that the strongest connections are driving the gradient (Page 7, Line 304):

“To assess the effect of sparsity on the derived gradients, we examined group-level structural (N = 222) and functional (N = 213) connectomes from the baseline session of NKI. The first three functional connectivity gradients derived using the full connectivity matrix (density = 92%) were highly consistent with those obtained from retaining the strongest 10% of connections in each row (r₁ = 0.999, r₂ = 0.998, r₃ < 0.999, all p < 0.001). Likewise, the first three communicability gradients derived from retaining all streamline counts (density = 83%) were almost identical to those obtained from 10% row-wise thresholding (r₁ = 0.994, r₂ = 0.963, r₃ = 0.955, all p < 0.001). This suggests that the reported gradients are driven by the strongest or most consistent connections within the connectomes, with minimal additional information provided by weaker connections. In terms of functional connectivity, such consistency reinforces past work demonstrating that the sensorimotor-toassociation axis, the major axis within the principal functional connectivity gradient, emerges across both the top- and bottom-ranked functional connections (Nenning et al., 2023)”.

Furthermore, we appreciate the nudge to share our thoughts on whether the difference between vertex versus nodal metrics could be important here, particularly regarding thresholds. To combine this point with R2’s recommendation to expand the Discussion, we have added the following paragraph (Page 19, Line 861): 

“We consider the role of thresholding, cortical resolution, and head motion as avenues to reconcile the present results with select reports in the literature (Dong et al., 2021; Xia et al., 2022). We would suggest that thresholding has a greater effect on vertex-level data, rather than parcel-level. For example, a recent study revealed that the emergence of principal vertex-level functional connectivity gradients in childhood and adolescence are indeed threshold-dependent (Dong et al., 2024). Specifically, the characteristic unimodal organisation for children and transmodal organisation for adolescents only emerged at the 90% threshold: a 95% threshold produced a unimodal organisation in both groups, whilst an 85% threshold produced a transmodal organisation in both groups. Put simply, the ‘swapping’ of gradient orders only occurs at certain thresholds. Furthermore, our results are not necessarily contradictory to this prior report (Dong et al., 2021): developmental changes in high-resolution gradients may be supported by a stable low-dimensional coarse manifold. Indeed, our decision to use parcellated connectomes was partly driven by recent work which demonstrated that vertex-level functional gradients may be derived using biologically-plausible but random data with sufficient spatial smoothing, whilst this effect is minimal at coarser resolutions (Watson & Andrews, 2023). We observed a gradual increase in the variance of individual connectomes accounted for by the principal functional connectivity gradient in the referred subset of CALM, in line with prior vertex-level work demonstrating a gradual emergence of the sensorimotor-association axis as the principal axis of connectivity (Xia et al., 2022), as opposed to a sudden shift. It is also possible that vertex-level data is more prone to motion artefacts in the context of developmental work. Transitioning from vertex-level to parcel-level data involves smoothing over short-range connectivity, thus greater variability in short-range connectivity can be observed in vertex-level data. However, motion artefacts are known to increase short-range connectivity and decrease long-range connectivity, mimicking developmental changes (Satterthwaite et al., 2013). Thus, whilst vertexlevel data offers greater spatial resolution in representation of short-range connectivity relative to parcel-level data, it is possible that this may come at the cost of making our estimates of the gradients more prone to motion”.

Evaluating the consistency of gradients across development: the results shown in Figure 1e are used as evidence suggesting that gradients are consistent across ages. However, I believe additional analyses are required to identify potential sources of the observed inconsistency compared to previous works. The claim that the principal gradient explains a similar degree of variance across ages does not necessarily imply that the spatial structure remains the same. The observed variance explanation is hence not enough to ascertain inconsistency with findings from Dong et al., as the spatial configuration of gradients may still change over time. I suggest the following additional analyses to strengthen this claim. Alignment to group-level gradients: Assess how much of the variance in individual FC matrices is explained by each of the group-level gradients (G1, G2, and G3, for both FC and SC). This analysis could be visualized similarly to Figure 1e, with age on the x-axis and variance explained on the y-axis. If the explained variance varies as a function of age, it may indicate that the gradients are not as consistent as currently suggested. 

This is another great suggestion. In the additional analyses above (new group-level analyses and unrotated gradient analyses) we rule-out a couple of the potential causes of the different developmental trends we observe in our data – namely the stability of the gradients over time. The suggested additional analysis is a great idea, and we have implemented it as follows (Page 8, Line 363):

“To evaluate the consistency of gradients across development, across baseline participants with functional connectomes from the referred CALM cohort (N = 177), we calculated the proportion of variance in individuallevel connectomes accounted for by group-level functional gradients. Specifically, we calculated the proportion of variance in an adjacency matrix A accounted for by the vector v_i as the fraction of the square of the scalar projection of v_i onto A, over the Frobenius norm of A. Using a generalised linear model, we then tested whether the proportion of variance explained varies systematically with age, controlling for sex and headmotion. The variance in individual-level functional connectomes accounted for by the group-level principal functional gradient gradually increased with development (β= 0.111, 95% CI = [0.022, 0.199], p = 1.452 x 10^-2, Cohen’s d = 0.367), as shown in Figure 1g, and decreased with higher head motion ( β = -10.041, 95% CI = [12.379, -7.702], p = 3.900 x 10^-17), with no effect of sex (β= 0.071, 95% CI = [-0.380, 0.523], p = 0.757). We observed no developmental effects on the variance explained by the second (r_s = 0.112, p = 0.139) or third (r_s = 0.053, p = 0.482) group-level functional gradient. When repeated with the baseline functional connectivity for NKI (N = 213), we observed no developmental effects (β = 0.097, 95% CI = [-0.035, 0.228], p = 0.150) on the variance explained by the principal functional gradient after accounting for motion (β= -3.376, 95% CI = [8.281, 1.528], p = 0.177) and sex (β = -0.368, 95% CI = [-1.078, 0.342], p = 0.309). However, we observed significant developmental correlations between age and variance (r_s = 0.137, p = 0.046) explained before accounting for head motion and sex. We observed no developmental effects on the variance explained by the second functional gradient (r_s = -0.066, p = 0.338), but a weak negative developmental effect on the variance explained by the third functional gradient (r_s = -0.189, p = 0.006). Note, however, the magnitude of the variance accounted for by the third functional gradient was very small (all < 1%). When applied to communicability matrices in CALM, the proportion of variance accounted for by the group-level communicability gradient was negligible (all < 1%), precluding analysis of developmental change”. 

“To further probe the consistency of gradients across development, we examined developmental changes in the standard deviation of gradient values, corresponding to heterogeneity, following prior work examining morphological (He et al., 2025) and functional connectivity gradients (Xia et al., 2022). Using a series of generalised linear models within the baseline referred subset of CALM, correcting for head motion and sex, we found that gradient variation for the principal functional gradient increased across development (= 0.219, 95% CI = [0.091, 0.347], p = 0.001, Cohen’s d = 0.504), indicating greater heterogeneity (Figure 1h), whilst gradient variation for the principal communicability gradient decreased across development (β = -0.154, 95% CI = [-0.267, -0.040], p = 0.008, Cohen’s d = -0.301), indicating greater homogeneity (Figure 1h). Note, a paired t-test on the 173 common participants demonstrated a significant effect of modality on gradient variability (t(172) = -56.639, p = 3.663 x 10^-113), such that the mean variability of communicability gradients (M = 0.033, SD = 0.001) was less than half that of functional connectivity (M = 0.076, SD = 0.010). Together, this suggests that principal functional connectivity and communicability gradients are established early in childhood and display age-related refinement, but not replacement”. 

The Issue of Abstraction and Benefits of the Gradient-Based View: The manuscript interprets the eccentricity findings as reflecting changes along the segregation-integration spectrum. Given this, it is unclear why a more straightforward analysis using established graph-theory metrics of segregationintegration was not pursued instead. Mapping gradients and computing eccentricity adds layers of abstraction and complexity. If similar interpretations can be derived directly from simpler graph metrics, what additional insights does the gradient-based framework offer? While the manuscript argues that this approach provides “a more unifying account of cortical reorganization”, it is not evident why this abstraction is necessary or advantageous over traditional graph metrics. Clarifying these benefits would strengthen the rationale for using this method. 

This is a great point, and something we spent quite a bit of time considering when designing the analysis. The central goal of our project was to identify gradients of brain organisation across different datasets and modalities and then test how the organisational principles of those modalities align. In other words, how do structural and functional ‘spaces’ intersect, and does this vary across the cortex? That for us was the primary motivation for operationalising organisation as nodal location within a low-dimensional manifold space (Bethlehem et al., 2020; Gale et al., 2022; Park et al., 2021), using a simple composite measure to achieve compression, rather than as a series of graph metrics. The reason we subsequently calculated those graph metrics and tested for their association was simply to help us interpret what eccentricity within that lowdimensional space means. Manifold eccentricity was moderately positively correlated to graph-theory metrics of integration, leaving a substantial portion of variance unaccounted for, but that association we think is nonetheless helpful for readers trying to interpret eccentricity. However, since ME tells us about the relative position of a node in that low-dimensional space, it is also likely capturing elements of multiple graph theory measures. Following the Reviewer’s question, this is something we decided to test. Specifically, using 4 measures of segregation, including two new metrics requested by the Reviewer in a minor point (weighted clustering coefficient and normalized degree centrality), we conducted a dominance analysis (Budescu, 1993) with normalized manifold eccentricity of the group-level referred CALM structural connectome. We also detail the use of gradient measures in developmental contexts, and how they can be complementary to traditional graph theory metrics. 

We have added the following to the Results section (Page 10, Lines 472 onwards): 

“To further contextualise manifold eccentricity in terms of integration and segregation beyond simple correlations, we conducted a multivariate dominance analysis (Budescu, 1993) of four graph theory metrics of segregation as predictors of nodal normalized manifold eccentricity within the group-level referred CALM structural and functional connectomes (Figure 2c). A dominance analysis assesses the relative importance of each predictor in a multilinear regression framework by fitting 2ⁿ – 1 models (where n is the number of predictors) and calculating the relative increase in adjusted R2 caused by adding each predictor to the model across both main effects and interactions. A multilinear regression model including weighted clustering coefficient, within-module degree Z-score, participation coefficient and normalized degree centrality accounted for 59% of variance in nodal manifold eccentricity in the group-level CALM structural connectome. Withinmodule degree Z score was the most important predictor (40.31% dominance), almost twice that of the participation coefficient (24.03% dominance) and normalized degree centrality (24.05% dominance) which made roughly equal contributions. The least important predictor was the weighted clustering coefficient (11.62% dominance). When the same approach was applied for the group-level referred CALM functional connectome, the 4 predictors accounted for 52% variability. However, in contrast to the structural connectome, functional manifold eccentricity seemed to incorporate the same graph theory metrics in different proportions. Normalized degree centrality was the most important predictor (47.41% dominance), followed by withinmodule degree Z-score (24.27%), and then the participation coefficient (15.57%) and weighted clustering coefficient (12.76%) which made approximately equal contributions. Thus, whilst structural manifold eccentricity was dominated most by within-module degree Z-score and least by the weighted clustering coefficient, functional manifold eccentricity was dominated most by normalized degree centrality and least by the weighted clustering coefficient. This suggests that manifold mapping techniques incorporate different aspects of integration dependent on modality. Together, manifold eccentricity acts as a composite measure of segregation, being differentially sensitive to different aspects of segregation, without necessitating a priori specification of graph theory metrics. Further discussion of the value of gradient-based metrics in developmental contexts and as a supplement to traditional graph theory analyses is provided in the ‘Manifold Eccentricity’ methodology sub-section”. 

We added further justification to the manifold eccentricity Methods subsection (Page 26, line 1283):

“Gradient-based measures hold value in developmental contexts, above and beyond traditional graph theory metrics: within a sample of over 600 cognitively-healthy adults aged between 18 and 88 years old, sensitivity of gradient-based within-network functional dispersion to age were stronger and more consistent across networks compared to segregation (Bethlehem et al., 2020). In the context of microstructural profile covariance, modules resolved by Louvain community detection occupied distinct positions across the principal two gradients, suggesting that gradients offer a way to meaningfully order discrete graph theory analyses (Paquola et al., 2019)”. 

We added the following to the Introduction section outlining the application of gradients as cortex-wide coordinate systems (Page 3, Line 121):

“Using the gradient-based approach as a compression tool, thus forgoing the need to specify singular graph theory metrics a priori, we operationalised individual variability in low-dimensional manifolds as eccentricity (Gale et al., 2022; Park et al., 2021). Crucially, such gradients appear to be useful predictors of phenotypic variation, exceeding edge-level connectomics. For example, in the case of functional connectivity gradients, their predictive ability for externalizing symptoms and general cognition in neurotypical adults surpassed that of edge-level connectome-based predictive modelling (Hong et al., 2020), suggesting that capturing lowdimensional manifolds may be particularly powerful biomarkers of psychopathology and cognition”. 

We also added the following to the Discussion section (Page 18, Line 839):

“By capitalising on manifold eccentricity as a composite measure of segregation across development, we build upon an emerging literature pioneering gradients as a method to establish underlying principles of structural (Paquola et al., 2020; Park et al., 2021) and functional (Dong et al., 2021; Margulies et al., 2016; Xia et al., 2022) brain development without a priori specification of specific graph theory metrics of interest”. 

It is unclear whether the statistical tests finding significant dataset effects are capturing effects of neurotypical vs. Neurodivergent, or simply different scanners/sites. Could the neurotypical portion of CALM also be added to distinguish between these two sources of variability affecting dataset effects (i.e. ideally separating this to the effect of site vs. neurotypicality would better distinguish the effect of neurodivergence).

At a group-level, differences in the gradients between the two cohorts are very minor. Indeed, in the manuscript we describe these gradients as being seemingly ‘universal’. But we agree that we should test whether we can directly attribute any simple main effects of ‘dataset’ are resulting from the different site or the phenotype of the participants. The neurotypical portion of CALM (collected at the same site on the same scanner) helped us show that any minor differences in the gradient alignments is likely due to the site/scanner differences rather than the phenotype of the participants. We took the same approach for testing the simple main effects of dataset on manifold eccentricity. To better parse neurotypicality and site effects at an individual-level, we conducted a series of sensitivity analyses. First, in response to the reviewer’s earlier comment, we conducted a series of nodal generalized linear models for communicability and FC gradients derived from neurotypical and neurodivergent portions of CALM, alongside NKI, and tested for an effect of neurotypicality above and beyond scanner. As at the group level, having those additional scans on a ‘comparison’ sample for CALM is very helpful in teasing apart these effects. We find that neurotypicality affects communicability gradient expression to a greater degree than functional connectivity. We visualised these results and added them to Figure 1. Second, we used the same approach but for manifold eccentricity. Again, we demonstrate greater sensitivity of neurotypicality to communicability at a global-level, but we cannot pin these effects down to specific networks because the effects do not survive the necessary multiple comparison correction. We have added these analyses to the manuscript (Page 13, Line 583): 

“Much as with the gradients themselves, we suspected that much of the simple main effect of dataset could reflect the scanner / site, rather than the difference in phenotype. Again, we drew upon the CALM comparison children to help us disentangle these two explanations. As a sensitivity analysis to parse effects of neurotypicality and dataset on manifold eccentricity, we conducted a series of generalized linear models predicting mean global and network-level manifold eccentricity, for each modality. We did this across all the baseline data (i.e. including the neurotypical comparison sample for CALM) using neurotypicality (2 levels: neurodivergent or neurotypical), site (2 levels: CALM or NKI), sex, head motion, and age at scan (Figure 3X). We restricted our analysis to baseline scans to create more equally-balanced groups. In terms of structural manifold eccentricity (N = 313 neurotypical, N = 311 neurodivergent), we observed higher manifold eccentricity in the neurodivergent participants at a global level (β = 0.090, p = 0.019, Cohen’s d = 0.188) but the individual network level effects did not survive the multiple comparison correction necessary for looking across all seven networks, with the default-mode network being the strongest (β = 0.135, p = 0.027, p_FDR = 0.109, Cohen’s d = 0.177). There was no significant effect of neurodiversity on functional manifold eccentricity (N = 292 neurotypical and N = 177 neurodivergent). This suggests that neurodiversity is significantly associated with structural manifold eccentricity, over and above differences in site, but we cannot distinguish these effects reliably in the functional manifold data”. 

Third, we removed the Scheirer-Ray-Hare test from the results for two reasons. First, its initial implementation did not account for repeated measures, and therefore non-independence between observations, as the same participants may have contributed both structural and functional data. Second, if we wanted to repeat this analysis in CALM using the referred and control portions, a significant difference in group size existed, which may affect the measures of variability. Specifically, for baseline CALM, 311 referred and 91 control participants contributed SC data, whilst 177 referred and 79 control participants contributed FC data. We believe that the ‘cleanest’ parsing of dataset and site for effects of eccentricity is achieved using the GLMs in Figure 3. 

We observed no significant effect of neurodivergence on the magnitude of structure-function coupling across development, and have added the following text (Page 14, Line 632):

“To parse effects of neurotypicality and dataset on structure-function coupling, we conducted a series of generalized linear models predicting mean global and network-level coupling using neurotypicality, site, sex, head motion, and age at scan, at baseline (N = 77 CALM neurotypical, N = 173 CALM neurodivergent, and N = 170 NKI). However, we found no significant effects of neurotypicality on structure-function coupling across development”. 

Since we demonstrated no significant effects of neurotypicality on structure-function coupling magnitude across development, but found differential dataset-specific effects of age on coupling development, we added the following sentence at the end of the coupling trajectory results sub-section (Page 14, line 664):

“Together, these effects demonstrate that whilst the magnitude of structure-function coupling appears not to be sensitive to neurodevelopmental phenotype, its development with age is, particularly in higher-order association networks, with developmental change being reduced in the neurodivergent sample”.  

Figure 1.c: A non-parametric permutation test (e.g. Mann-Whitney U test) could quantitatively identify regions with significant group differences in nodal gradient values, providing additional support for the qualitative findings.

This is a great idea. To examine the effect of referral status on nodal gradient values, whilst controlling for covariates (head motion and sex), we conducted a series of generalised linear models. We opted for this instead of a Mann-Whitney U test, as the former tests for differences in distributions, whilst the direction of the t-statistic for referral status from the GLM would allow us to specify the magnitude and direction of differences in nodal gradient values between the two groups. Again, we conducted this in CALM (referred vs control), at an individual-level, as downstream analyses suggested a main effect of dataset (which is reflected in the highly-similar group-level referred and control CALM gradients). We have updated the Results section with the following text (Page 6, Line 283):

“To examine the effect of referral status on participant-level nodal gradient values in CALM, we conducted a series of generalized linear models controlling for head motion, sex and age at scan (Figure 1d). We restricted our analyses to baseline scans to reduce the difference in sample size for the referred (311 communicability and 177 functional gradients, respectively) and control participants (91 communicability and 79 functional gradients, respectively), and to the principal gradients. For communicability, 42 regions showed a significant effect (p < 0.05) of neurodivergence before FDR correction, with 9 post FDR correction. 8 of these 9 regions had negative t-statistics, suggesting a reduced nodal gradient value and representation in the neurodivergent children, encompassing both lower-order somatosensory cortices alongside higher-order fronto-parietal and default-mode networks. The largest reductions were observed within the prefrontal cortices of the defaultmode network (t = -3.992, p = 6.600 x 10^-5, p_FDR = 0.013, Cohen’s d = -0.476), the left orbitofrontal cortex of the limbic network (t = -3.710, p = 2.070 x 10^-4, p_FDR = 0.020, Cohen’s d = -0.442) and right somato-motor cortex (t = -3.612, p = 3.040 x 10^-4, p_FDR = 0.020, Cohen’s d = -0.431). The right visual cortex was the only exception, with stronger gradient representation within the neurotypical cohort (t = 3.071, p = 0.002, p_FDR = 0.048, Cohen’s d = 0.366). For functional connectivity, comparatively fewer regions exhibited a significant effect (p < 0.05) of neurotypicality, with 34 regions prior to FDR correction and 1 post. Significantly stronger gradient representation was observed in neurotypical children within the right precentral ventral division of the defaultmode network (t = 3.930, p = 8.500 x 10^-5, p_FDR = 0.017, Cohen’s d = 0.532). Together, this suggests that the strongest and most robust effects of neurodivergence are observed within gradients of communicability, rather than functional connectivity, where alterations in both affect higher-order associative regions”. 

In the harmonization methodology, it is mentioned that “if harmonisation was successful, we’d expect any significant effects of scanner type before harmonisation to be non-significant after harmonisation”. However, given that there were no significant effects before harmonization, the results reported do not help in evaluating the quality of harmonization.

We agree with the Reviewer, and have removed the post-harmonisation GLMs, and instead stating that there were no significant effects of scanner type before harmonization. 

Figure 3: It would be helpful to include a plot showing the GAMM predictions versus real observations of eccentricity (x-axis: predictions, y-axis: actual values). 

To plot the GAMM-predicted smooth effects of age, which we used for visualisation purposes only, we used the get_predictions function from the itsadug R package. This creates model predictions using the median value of nuisance covariates. Thus, whilst we specified the entire age range, the function automatically chooses the median of head motion, alongside controlling for sex (default level: male) and, for each dataset-specific trajectory. Since the gamm4 package separates the fitted model into a gam and linear mixed effects model (which accounts for participant ID as a random effect), and the get_predictions function only uses gam, random effects are not modelled in the predicted smooths. Therefore, any discrepancy between the observed and predicted manifold eccentricity values is likely due to sensitivity to default choices of covariates other than age, or random effects. To prevent Figure 3 being too over-crowded, we opted to not include the predictions: these were strongly correlated with real structural manifold data, but less for functional manifold data especially where significant developmental change was absent.

The 30mm threshold for filtering short streamlines in tractography is uncommon. What is the rationale for using such a large threshold, given the potential exclusion of many short-range association fibres?

A minimum length of 30mm was the default for the MRtrix3 reconstruction workflow, and something we have previously used. In a previous project, we systematically varied the minimum fibre length and found that this had minimal impact on network organisation (e.g. Mousley et al. 2025). However, we accept that short-range association fibres may have been excluded and have included this in the Discussion as a methodological limitation, alongside our predictions for how the gradients and structure-function coupling may’ve been altered had we included such fibres (Page 20, Line 955):

“A potential methodological limitation in the construction of structural connectomes was the 30mm tract length threshold which, despite being the QSIprep reconstruction default (Cieslak et al., 2021), may have potentially excluded short-range association fibres. This is pertinent as tracts of different lengths exhibit unique distributions across the cortex and functional roles (Bajada et al., 2019) : short-range connections occur throughout the cortex but peak within primary areas, including the primary visual, somato-motor, auditory, and para-hippocampal cortices, and are thought to dominate lower-order sensorimotor functional resting-state networks, whilst long-range connections are most abundant in tertiary association areas and are recruited alongside tracts of varying lengths within higher-order functional resting-state networks. Therefore, inclusion of short-range association fibres may have resulted in a relative increase in representation of lower-order primary areas and functional networks. On the other hand, we also note the potential misinterpretation of short-range fibres: they may be unreliably distinguished from null models in which tractography is restricted by cortical gyri only (Bajada et al., 2019). Further, prior (neonatal) work has demonstrated that the order of connectivity of regions and topological fingerprints are consistent across varying streamline thresholds (Mousley et al., 2025), suggesting minimal impact”. 

Given the spatial smoothing of fMRI data (6mm FWHM), it would be beneficial to apply connectome spatial smoothing to structural connectivity measures for consistent spatial smoothness.

This is an interesting suggestion but given we are looking at structural communicability within a parcellated network, we are not sure that it would make any difference. The data structural data are already very smooth. Nonetheless we have added the following text to the Discussion (Page 20, Line 968): 

“Given the spatial smoothing applied to the functional connectivity data, and examining its correspondence to streamline-count connectomes through structure-function coupling, applying the equivalent smoothing to structural connectomes may improve the reliability of inference, and subsequent sensitivity to cognition and psychopathology. Connectome spatial smoothing involves applying a smoothing kernel to the two streamline endpoints, whereby variations in smoothing kernels are selected to optimise the trade-off between subjectlevel reliability and identifiability, thus increasing the signal-to-noise ratio and the reliability of statistical inferences of brain-behaviour relationships (Mansour et al., 2022). However, we note that such smoothing is more effective for high-resolution connectomes, rather than parcel-level, and so have only made a modest improvement (Mansour et al., 2022)”.

Why was harmonization performed only within the CALM dataset and not across both CALM and NKI datasets? What was the rationale for this decision?

We thought about this very carefully. Harmonization aims to remove scanner or site effects, whilst retaining the crucial characteristics of interest. Our capacity to retain those characteristics is entirely dependent on them being *fully* captured by covariates, which are then incorporated into the harmonization process. Even with the best set of measures, the idea that we can fully capture ‘neurodivergence’ and thus preserve it in the harmonisation process is dubious. Indeed, across CALM and NKI there are limited number of common measures (i.e. not the best set of common measures), and thus we are limited in our ability to fully capture the neurodivergence with covariates. So, we worried that if we put these two very different datasets into the harmonisation process we would essentially eliminate the interesting differences between the datasets. We have added this text to the harmonization section of the Methods (Page 24, Line 1225):

“Harmonization aims to retain key characteristics of interest whilst removing scanner or site effects. However, the site effects in the current study are confounded with neurodivergence, and it is unlikely that neurodivergence may be captured fully using common covariates across CALM and NKI. Therefore, to preserve variation in neurodivergence, whilst reducing scanner effects, we harmonized within the CALM dataset only”. 

The exclusion of subcortical areas from connectivity analyses is not justified. 

This is a good point. We used the Schaefer atlas because we had previously used this to derive both functional and structural connectomes, but we agree that it would have been good to include subcortical areas (Page 20, Line 977). 

“A potential limitation of our study was the exclusion of subcortical regions. However, prior work has shed light on the role of subcortical connectivity in structural and functional gradients, respectively, of neurotypical populations of children and adolescents (Park et al., 2021; Xia et al., 2022). For example, in the context of the primary-to-transmodal and sensorimotor-to-visual functional connectivity gradients, the mean gradient scores within subcortical networks were demonstrated to be relatively stable across childhood and adolescence (Xia et al., 2022). In the context of structural connectivity gradients derived from streamline counts, which we demonstrated were highly consistent with those derived from communicability, subcortical structural manifolds weighted by their cortical connectivity were anchored by the caudate and thalamus at one pole, and by the hippocampus and nucleus accumbens at the opposite pole, with significant age-related manifold expansion within the caudate and thalamus (Park et al., 2021)”. 

In the KNN imputation method, were uniform weights used, or was an inverse distance weighting applied?

Uniform weights were used, and we have updated the manuscript appropriately.

The manuscript should clarify from the outset that the reported sample size (N) includes multiple longitudinal observations from the same individuals and does not reflect the number of unique participants.

We have rectified the Abstract (Page 2, Line 64) and Introduction (Page 3, Line 138):

“We charted the organisational variability of structural (610 participants, N = 390 with one observation, N = 163 with two observations, and N = 57 with three) and functional (512 participants, N = 340 with one observation, N = 128 with two observations, and N = 44 with three)”.

The term “structural gradients” is ambiguous in the introduction. Clarify that these gradients were computed from structural and functional connectivity matrices, not from other structural features (e.g. cortical thickness).

We have clarified this in the Introduction (Page 3, Line 134):

“Applying diffusion-map embedding as an unsupervised machine-learning technique onto matrices of communicability (from streamline SIFT2-weighted fibre bundle capacity) and functional connectivity, we derived gradients of structural and functional brain organisation in children and adolescents…”

Page 5: The sentence, “we calculated the normalized angle of each structural and functional connectome to derive symmetric affinity matrices” is unclear and needs clarification.

We have clarified this within the second paragraph of the Results section (Page 4, Line 185):

“To capture inter-nodal similarity in connectivity, using a normalised angle kernel, we derived individual symmetric affinity matrices from the left and right hemispheres of each communicability and functional connectivity matrix. Varying kernels capture different but highly-related aspects of inter-nodal similarity, such as correlation coefficients, Gaussian kernels, and cosine similarity. Diffusion-map embedding is then applied on the affinity matrices to derive gradients of cortical organisation”. 

Figure 1.a: “Affine A” likely refers to the affinity matrix. The term “affine” may be confusing; consider using a clearer label. It would also help to add descriptive labels for rows and columns (e.g. region x region).

Thank you for this suggestion! We have replaced each of the labels with “pairwise similarity”. We also labelled the rows and columns as regions.

Figure 1.d: Are the cross-group differences statistically significant? If so, please indicate this in the figure.

We have added the results of a series of linear mixed effects models to the legend of Figure 1 (Page 6, line 252):

“indicates a significant effect of dataset (p < 0.05) on variance explained within a linear mixed effects model controlling for head motion, sex, and age at scan”.

The sentence “whose connectomes were successfully thresholded” in the methods is unclear. What does “successfully thresholded” mean? Additionally, this seems to be the first mention of the Schaefer 100 and Brainnetome atlas; clarify where these parcellations are used. 

We have amended the Methodology section (Page 23, Line 1138):

“For each participant, we retained the strongest 10% of connections per row, thus creating fully connected networks required for building affinity matrices. We excluded any connectomes in which such thresholding was not possible due to insufficient non-zero row values. To further ensure accuracy in connectome reconstruction, we excluded any participants whose connectomes failed thresholding in two alternative parcellations: the 100node Schaefer 7-network (Schaefer et al., 2018) and Brainnetome 246-node (Fan et al., 2016) parcellations, respectively”. 

We have also specified the use of the Schaefer 200-node parcellation in the first sentence on the second Results paragraph.

The use of “streamline counts” is misleading, as the method uses SIFT2-weighted fibre bundle capacity rather than raw streamline counts. It would be better to refer to this measure as “SIFT2-weighted fibre bundle capacity” or “FBC”.

We replaced all instances of “streamline counts” with “SIFT2-weighted fibre bundle capacity” as appropriate.

Figure 2.c: Consider adding plots showing changes in eccentricity against (1) degree centrality, and (2) weighted local clustering coefficient. Additionally, a plot showing the relationship between age and mean eccentricity (averaged across nodes) at the individual level would be informative.

We added the correlation between eccentricity and both degree centrality and the weighted local clustering coefficient and included them in our dominance analysis in Figure 2. In terms of the relationship between age and mean (global) eccentricity, these are plotted in Figure 3. 

Figure 2.b: Considering the results of the following sections, it would be interesting to include additional KDE/violin plots to show group differences in the distribution of eccentricity within 7 different functional networks.

As part of our analysis to parse neurotypicality and dataset effects, we tested for group differences in the distribution of structural and functional manifold eccentricity within each of the 7 functional networks in the referred and control portions of CALM and have included instances of significant differences with a coloured arrow to represent the direction of the difference within Figure 3. 

Figure 3: Several panels lack axis labels for x and y axes. Adding these would improve clarity.

To minimise the amount of text in Figure 3, we opted to include labels only for the global-level structural and functional results. However, to aid interpretation, we added a small schematic at the bottom of Figure 3 to represent all axis labels. 

The statement that “differences between datasets only emerged when taking development into account” seems inaccurate. Differences in eccentricity are evident across datasets even before accounting for development (see Fig 2.b and the significance in the Scheirer-Ray-Hare test).

We agree – differences in eccentricity across development and datasets are evident in structural and functional manifold eccentricity, as well as within structure-function coupling. However, effects of neurotypicality were particularly strong for the maturation of structure-function coupling, rather than magnitude. Therefore, we have rephrased this sentence in the Discussion (page 18, line 832):

“Furthermore, group-level structural and functional gradients were highly consistent across datasets, whilst differences between datasets were emphasised when taking development into account, through differing rates of structural and functional manifold expansion, respectively, alongside maturation of structure-function coupling”.

The handling of longitudinal data by adding a random effect for individuals is not clear in the main text. Mentioning this earlier could be helpful. 

We have included this detail in the second sentence of the “developmental trajectories of structural manifold contraction and functional manifold expansion” results sub-section (page 11, line 503):

“We included a random effect for each participant to account for longitudinal data”. 

Figure 4.b: Why were ranks shown instead of actual coefficient of variation values? Consider including a cortical map visualization of the coefficients in the supplementary material.

We visualised the ranks, instead of the actual coefficient of variation (CV) values, due to considerable variability and skew in the magnitude of the CV, ranging from 28.54 (in the right visual network) to 12865.68 (in the parietal portion of the left default-mode network), with a mean of 306.15. If we had visualised the raw CV values, these larger values would’ve been over-represented. We’ve also noticed and rectified an error in the labelling of the colour bar for Figure 4b: the minimum should be most variable (i.e. a rank of 1). To aid contextualisation of the ranks, we have added the following to the Results (page 14, line 626):

“The distribution of cortical coefficients of variation (CV) varied considerably, with the largest CV (in the parietal division of the left default-mode network) being over 400 times that of the smallest (in the right visual network). The distribution of absolute CVs was positively skewed, with a Fisher skewness coefficient g₁ of 7.172, meaning relatively few regions had particularly high inter-individual variability, and highly peaked, with a kurtosis of 54.883, where a normal distribution has a skewness coefficient of 0 and a kurtosis of 3”. 

Reviewer #2 (Public review):

Some differences in developmental trajectories between CALM and NKI (e.g. Figure 4d) are not explained. Are these differences expected, or do they suggest underlying factors that require further investigation?

This is a great point, and we appreciate the push to give a fuller explanation. It is very hard to know whether these effects are expected or not. We certainly don’t know of any other papers that have taken this approach. In response to the reviewer’s point, we decided to run some more analyses to better understand the differences. Having observed stronger age effects on structure-function coupling within the neurotypical NKI dataset, compared to the absent effects in the neurodivergent portion of CALM, we wanted to follow up and test that it really is that coupling is more sensitive to the neurodivergent versus neurotypical difference between CALM and NKI (rather than say, scanner or site effects). In short, we find stronger developmental effects of coupling within the neurotypical portion of CALM, rather than neurodivergent, and have added this to the Results (page 15, line 701):

“To further examine whether a closer correspondence of structure-function coupling with age is associated with neurotypicality, we conducted a follow-up analysis using the additional age-matched neurotypical portion of CALM (N = 77). Given the widespread developmental effects on coupling within the neurotypical NKI sample, compared to the absent effects in the neurodivergent portion of CALM, we would expect strong relationships between age and structure-function coupling with the neurotypical portion of CALM. This is indeed what we found: structure-function coupling showed a linear negative relationship with age globally (F = 16.76, p_FDR < 0.001, adjusted R² = 26.44%), alongside fronto-parietal (F = 9.24, p_FDR = 0.004, adjusted R² = 19.24%), dorsalattention (F = 13.162, p_FDR = 0.001, adjusted R²= 18.14%), ventral attention (F = 11.47, p_FDR  = 0.002, adjusted R²= 22.78), somato-motor (F = 17.37, p_FDR  < 0.001, adjusted R²= 21.92%) and visual (F = 11.79, p_FDR  = 0.002, adjusted R²= 20.81%) networks. Together, this supports our hypothesis that within neurotypical children and adolescents, structure-function coupling decreases with age, showing a stronger effect compared to their neurodivergent counterparts, in tandem with the emergence of higher-order cognition. Thus, whilst the magnitude of structure-function coupling across development appeared insensitive to neurotypicality, its maturation is sensitive. Tentatively, this suggests that neurotypicality is linked to stronger and more consistent maturational development of structure-function coupling, whereby the tethering of functional connectivity to structure across development is adaptive”. 

In conjunction with the Reviewer’s later request to deepen the Discussion, we have included an additional paragraph attempting to explain the differences in neurodevelopmental trajectories of structure-function coupling (Page 19, Line 924):

“Whilst the spatial patterning of structure-function coupling across the cortex has been extensively documented, as explained above, less is known about developmental trajectories of structure-function coupling, or how such trajectories may be altered in those with neurodevelopmental conditions. To our knowledge, only one prior study has examined differences in developmental trajectories of (non-manifold) structure-function coupling in typically-developing children and those with attention-deficit hyperactivity disorder (Soman et al., 2023), one of the most common conditions in the neurodivergent portion of CALM. Namely, using cross-sectional and longitudinal data from children aged between 9 and 14 years old, they demonstrated increased coupling across development in higher-order regions overlapping with the defaultmode, salience, and dorsal attention networks, in children with ADHD, with no significant developmental change in controls, thus encompassing an ectopic developmental trajectory (Di Martino et al., 2014; Soman et al., 2023). Whilst the current work does not focus on any condition, rather the broad mixed population of young people with neurodevelopmental symptoms (including those with and without diagnoses), there are meaningful individual and developmental differences in structure-coupling. Crucially, it is not the case that simply having stronger coupling is desirable. The current work reveals that there are important developmental trajectories in structure-function coupling, suggesting that it undergoes considerable refinement with age. Note that whilst the magnitude of structure-function coupling across development did not differ significantly as a function of neurodivergence, its relationship to age did. Our working hypothesis is that structural connections allow for the ordered integration of functional areas, and the gradual functional modularisation of the developing brain. For instance, those with higher cognitive ability show a stronger refinement of structurefunction coupling across development. Future work in this space needs to better understand not just how structural or functional organisation change with time, but rather how one supports the other”. 

The use of COMBAT may have excluded extreme participants from both datasets, which could explain the lack of correlations found with psychopathology.

COMBAT does not exclude participants from datasets but simply adjusts connectivity estimates. So, the use of COMBAT will not be impacting the links with psychopathology by removing participants. But this did get us thinking. Excluding participants based on high motion may have systematically removed those with high psychopathology scores, meaning incomplete coverage. In other words, we may be under-representing those at the more extreme end of the range, simply because their head-motion levels are higher and thus are more likely to be excluded. We found that despite certain high-motion participants being removed, we still had good coverage of those with high scores and were therefore sensitive within this range. We have added the following to the revised Methods section (Page 26, Line 1338):

“As we removed participants with high motion, this may have overlapped with those with higher psychopathology scores, and thus incomplete coverage. To examine coverage and sensitivity to broad-range psychopathology following quality control, we calculated the Fisher-Pearson skewness statistic g₁ for each of the 6 Conners t-statistic measures and the proportion of youth with a t-statistic equal to or greater than 65, indicating an elevated or very elevated score. Measures of inattention (g₁ = 0.11, 44.20% elevated), hyperactivity/impulsivity (g₁ = 0.48, 36.41% elevated), learning problems (g₁ = 0.45, 37.36% elevated), executive functioning (g₁ = 0.27, 38.16% elevated), aggression (g₁ = 1.65, 15.58% elevated), and peer relations (g₁ = 0.49, 38% elevated) were positively skewed and comprised of at least 15% of children with elevated or very elevated scores, suggesting sufficient coverage of those with extreme scores”. 

There is no discussion of whether the stable patterns of brain organization could result from preprocessing choices or summarizing data to the mean. This should be addressed to rule out methodological artifacts. 

This is a brilliant point. We are necessarily using a very lengthy pipeline, with many design choices to explore structural and functional gradients and their intersection. In conjunction with the Reviewer’s later suggestion to deepen the Discussion, we have added the following paragraph which details the sensitivity analyses we carried out to confirm the observed stable patterns of brain organization (Page 18, Line 863):

“That is, whilst we observed developmental refinement of gradients, in terms of manifold eccentricity, standard deviation, and variance explained, we did not observe replacement. Note, as opposed to calculating gradients based on group data, such as a sliding window approach, which may artificially smooth developmental trends and summarise them to the mean, we used participant-level data throughout. Given the growing application of gradient-based analyses in modelling structural (He et al., 2025; Li et al., 2024) and functional (Dong et al., 2021; Xia et al., 2022) brain development, we hope to provide a blueprint of factors which may affect developmental conclusions drawn from gradient-based frameworks”.

Although imputing missing data was necessary, it would be useful to compare results without imputed data to assess the impact of imputation on findings. 

It is very hard to know the impact of imputation without simply removing those participants with some imputed data. Using a simulation experiment, we expressed the imputation accuracy as the root mean squared error normalized by the range of observable data in each scale. This produced a percentage error margin. We demonstrate that imputation accuracy across all measures is at worst within approximately 11% of the observed data, and at best within approximately 4% of the observed data, and have included the following in the revised Methods section (Page 27, Line 1348):

“Missing data

To avoid a loss of statistical power, we imputed missing data. 27.50% of the sample had one or more missing psychopathology or cognitive measures (equal to 7% of all values), and the data was not missing at random: using a Welch’s t-test, we observed a significant effect of missingness on age [t (264.479) = 3.029, p = 0.003, Cohen’s d = 0.296], whereby children with missing data (M = 12.055 years, SD = 3.272) were younger than those with complete data (M = 12.902 years, SD = 2.685). Using a subset with complete data (N = 456), we randomly sampled 10% of the values in each column with replacement and assigned those as missing, thereby mimicking the proportion of missingness in the entire dataset. We conducted KNN imputation (uniform weights) on the subset with complete data and calculated the imputation accuracy as the root mean squared error normalized by the observed range of each measure. Thus, each measure was assigned a percentage which described the imputation margin of error. Across cognitive measures, imputation was within a 5.40% mean margin of error, with the lowest imputation error in the Trail motor speed task (4.43%) and highest in the Trails number-letter switching task (7.19%). Across psychopathology measures, imputation exhibited a mean 7.81% error margin, with the lowest imputation error in the Conners executive function scale (5.75%) and the highest in the Conners peer relations scale (11.04%). Together, this suggests that imputation was accurate”.

The results section is extensive, with many reports, while the discussion is relatively short and lacks indepth analysis of the findings. Moving some results into the discussion could help balance the sections and provide a deeper interpretation. 

We agree with the Reviewer and appreciate the nudge to expand the Discussion section. We have added 4 sections to the Discussion. The first explores the importance of the default-mode network as a region whose coupling is most consistently predicted by working memory across development and phenotypes, in terms of its underlying anatomy (Paquola et al., 2025) (Page 20, Line 977):

“An emerging theme from our work is the importance of the default-mode network as a region in which structure-function coupling is reliably predicted by working memory across neurodevelopmental phenotypes and datasets during childhood and adolescence. Recent neurotypical adult investigations combining highresolution post-mortem histology, in vivo neuroimaging, and graph-theory analyses have revealed how the underlying neuroanatomy of the default-mode network may support diverse functions (Paquola et al., 2025), and thus exhibit lower structure-function coupling compared to unimodal regions. The default-mode network has distinct neuroanatomy compared to the remaining 6 intrinsic resting-state functional networks (Yeo et al., 2011), containing a distinctive combination of 5 of the 6 von Economo and Koskinas cell types (von Economo & Koskinas, 1925), with an over-representation of heteromodal cortex, and uniquely balancing output across all cortical types. A primary cytoarchitectural axis emerges, beyond which are mosaic-like spatial topographies. The duality of the default-mode network, in terms of its ability to both integrate and be insulated from sensory information, is facilitated by two microarchitecturally distinct subunits anchored at either end of the cytoarchitectural axis (Paquola et al., 2025). Whilst beyond the scope of the current work, structure-function coupling and their predictive value for cognition may also differ across divisions within the default-mode network, particularly given variability in the smoothness and compressibility of cytoarchitectural landscapes across subregions (Paquola et al., 2025)”. 

The second provides a deeper interpretation and contextualisation of greater sensitivity of communicability, rather than functional connectivity, to neurodivergence (Page 19, Lines 907):

“We consider two possible factors to explain the greater sensitivity of neurodivergence to gradients of communicability, rather than functional connectivity. First, functional connectivity is likely more sensitive to head motion than structural-based communicability and suffers from reduced statistical power due to stricter head motion thresholds, alongside greater inter-individual variability. Second, whilst prior work contrasting functional connectivity gradients from neurotypical adults with those with confirmed ASD diagnoses demonstrated vertex-level reductions in the default-mode network in ASD and marginal increases in sensorymotor communities (Hong et al., 2019), indicating a sensitivity of functional connectivity to neurodivergence, important differences remain. Specifically, whilst the vertex-level group-level differences were modest, in line with our work, greater differences emerged when considering step-wise functional connectivity (SFC); in other words, when considering the dynamic transitions of or information flow through the functional hierarchy underlying the static functional connectomes, such that ASD was characterised by initial faster SFC within the unimodal cortices followed by a lack of convergence within the default-mode network (Hong et al., 2019). This emphasis on information flow and dynamic underlying states may point towards greater sensitivity of neurodivergence to structural communicability – a measure directly capturing information flow – than static functional connectivity”. 

The third paragraph situates our work within a broader landscape of reliable brain-behaviour relationships, focusing on the strengths of combining clinical and normative samples to refine our interpretation of the relationship between gradients and cognition, as well as the importance of equifinality in developmental predictive work (Page 20, line 994):

“In an effort to establish more reliable brain-behaviour relationships despite not having the statistical power afforded by large-scale, typically normative, consortia (Rosenberg & Finn, 2022), we demonstrated the development-dependent link between default-mode structure-function coupling and working memory generalised across clinical (CALM) and normative (NKI) samples, across varying MRI acquisition parameters, and harnessing within- and across-participant variation. Such multivariate associations are likely more reliable than their univariate counterparts (Marek et al., 2022), but can be further optimised using task-related fMRI (Rosenberg & Finn, 2022). The consistency, or lack of, of developmental effects across datasets emphasises the importance of validating brain-behaviour relationships in highly diverse samples. Particularly evident in the case of structure-function coupling development, through our use of contrasting samples, is equifinality (Cicchetti & Rogosch, 1996), a key concept in developmental neuroscience: namely, similar ‘endpoints’ of structure-function coupling may be achieved through different initialisations dependent on working memory. 

The fourth paragraph details methodological limitations in response to Reviewer 1’s suggestions to justify the exclusion of subcortical regions and consider the role of spatial smoothing in structural connectome construction as well as the threshold for filtering short streamlines”. 

While the methods are thorough, it is not always clear whether the optimal approaches were chosen for each step, considering the available data. 

In response to Reviewer 1’s concerns, we conducted several sensitivity analyses to evaluate the robustness of our results in terms of procedure. Specifically, we evaluated the impact of thresholding (full or sparse), level of analysis (individual or group gradients), construction of the structural connectome (communicability or fibre bundle capacity), Procrustes rotation (alignment to group-level gradients before Procrustes), tracking the variance explained in individual connectomes by group-level gradients, impact of head motion, and distinguishing between site and neurotypicality effects. All these analyses converged on the same conclusion: whilst we observe some developmental refinement in gradients, we do not observe replacement. We refer the reviewer to their third point, about whether stable patterns of brain organization were artefactual. 

The introduction is overly long and includes numerous examples that can distract readers unfamiliar with the topic from the main research questions. 

We have removed the following from the Introduction, reducing it to just under 900 words:

“At a molecular level, early developmental patterning of the cortex arises through interacting gradients of morphogens and transcription factors (see Cadwell et al., 2019). The resultant areal and progenitor specialisation produces a diverse pool of neurones, glia, and astrocytes (Hawrylycz et al., 2015). Across childhood, an initial burst in neuronal proliferation is met with later protracted synaptic pruning (Bethlehem et al., 2022), the dynamics of which are governed by an interplay between experience-dependent synaptic plasticity and genomic control (Gottlieb, 2007)”.

“The trends described above reflect group-level developmental trends, but how do we capture these broad anatomical and functional organisational principles at the level of an individual?”

We’ve also trimmed the second Introduction paragraph so that it includes fewer examples, such as removal of the wiring-cost optimisation that underlies structural brain development, as well as removing specific instances of network segregation and integration that occur throughout childhood.

Reviewer #3 (Public review):

The article's main question is how humans handle spurious transitions between object features when learning a predictive model for decision-making. The authors conjecture that humans use semantic knowledge about plausible causal relations as an inductive bias to distinguish true from spurious links.

The authors simulate a successor feature (SF) model, demonstrating its susceptibility to suboptimal learning in the presence of spurious transitions caused by co-occurring but independent causal factors. This effect worsens with an increasing number of planning steps and higher co-occurrence rates. In a preregistered study (N=100), they show that humans are also affected by spurious transitions, but perform somewhat better when true transitions occur between features within the same semantic category. However, no evidence for the benefits of semantic congruency was found in test trials involving novel configurations, and attempts to model these biases within an SF framework remained inconclusive.

Strengths:

(1) The authors tackle an important question.

(2) Their simulations employ a simple yet powerful SF modeling framework, offering computational insights into the problem.

(3) The empirical study is preregistered, and the authors transparently report both positive and null findings.

(4) The behavioral benefit during learning in the congruent vs incongruent condition is interesting

Weaknesses:

(1) A major issue is that approximately one quarter of participants failed to learn, while another quarter appeared to use conjunctive or configural learning strategies. This raises questions about the appropriateness of the proposed feature-based learning framework for this task. Extensive prior research suggests that learning about multi-attribute objects is unlikely to involve independent feature learners (see, e.g., the classic discussion of configural vs. elemental learning in conditioning: Bush & Mosteller, 1951; Estes, 1950).

(2) A second concern is the lack of explicit acknowledgment and specification of the essential role of the co-occurrence of causal factors. With sufficient training, SF models can develop much stronger representations of reliable vs. spurious transitions, and simple mechanisms like forgetting or decay of weaker transitions would amplify this effect. This should be clarified from the outset, and the occurrence rates used in all tasks and simulations need to be clearly stated.

(3) Another problem is that the modeling approach did not adequately capture participant behavior. While the authors demonstrate that the b parameter influences model behavior in anticipated ways, it remains unclear how a model could account for the observed congruency advantage during learning but not at test.

(4) Finally, the conceptualization of semantic biases is somewhat unclear. As I understand it, participants could rely on knowledge such as "the shape of a building robot's head determines the kind of head it will build," while the type of robot arm would not affect the head shape. However, this assumption seems counterintuitive - isn't it plausible that a versatile arm is needed to build certain types of robot heads?

if researchers combine typological and experimental approaches, they will find cases where the two types of evidence do not match. - The linguistic facts observed in natural languages are well-established, but the experiments fail to support the proposed cognitive mechanism (ex. typological data: ‘grammaticalisation’ goes from concrete concepts to abstract concepts, but not viceversa. Hypothesis: it is cognitively easier to do the association in this direction. Experimental data: false, people can do the association in both directions in their brain and it’s quite symmetrical —-> so why does it only happen in one direction in language?) - The cognitive mechanisms proposed by experimental data seem plausible, but the typological evidence is highly disputed among linguists (ex. experimental data: non native speakers simplify language and over time the standard variety also loose complexity. Typological data: imprecise data, variable based on the definition of ‘complexity’ and many more factors can influence the evolution of a language)

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public review):

Summary:

Syed et al. investigate the circuit underpinnings for leg grooming in the fruit fly. They identify two populations of local interneurons in the right front leg neuromere of ventral nerve cord, i.e. 62 13A neurons and 64 13B neurons. Hierarchical clustering analysis identifies 10 morphological classes for both populations. Connectome analysis reveals their circuit interactions: these GABAergic interneurons provide synaptic inhibition either between the two subpopulations, i.e., 13B onto 13A, or among each other, i.e., 13As onto other 13As, and/or onto leg motoneurons, i.e., 13As and 13Bs onto leg motoneurons. Interestingly, 13A interneurons fall into two categories, with one providing inhibition onto a broad group of motoneurons, being called "generalists", while others project to a few motoneurons only, being called "specialists". Optogenetic activation and silencing of both subsets strongly affect leg grooming. As well aas ctivating or silencing subpopulations, i.e., 3 to 6 elements of the 13A and 13B groups, has marked effects on leg grooming, including frequency and joint positions, and even interrupting leg grooming. The authors present a computational model with the four circuit motifs found, i.e., feed-forward inhibition, disinhibition, reciprocal inhibition, and redundant inhibition. This model can reproduce relevant aspects of the grooming behavior.

Strengths:

The authors succeeded in providing evidence for neural circuits interacting by means of synaptic inhibition to play an important role in the generation of a fast rhythmic insect motor behavior, i.e., grooming. Two populations of local interneurons in the fruit fly VNC comprise four inhibitory circuit motifs of neural action and interaction: feed-forward inhibition, disinhibition, reciprocal inhibition, and redundant inhibition. Connectome analysis identifies the similarities and differences between individual members of the two interneuron populations. Modulating the activity of small subsets of these interneuron populations markedly affects the generation of the motor behavior, thereby exemplifying their important role in generating grooming.

We thank the reviewer for their thoughtful and constructive evaluation of our work. 

Weaknesses:

Effects of modulating activity in the interneuron populations by means of optogenetics were conducted in the so-called closed-loop condition. This does not allow for differentiation between direct and secondary effects of the experimental modification in neural activity, as feedforward and feedback effects cannot be disentangled. To do so, open loop experiments, e.g., in deafferented conditions, would be important. Given that many members of the two populations of interneurons do not show one, but two or more circuit motifs, it remains to be disentangled which role the individual circuit motif plays in the generation of the motor behavior in intact animals.

Our optogenetic experiments show a role for 13A/B neurons in grooming leg movements – in an intact sensorimotor system - but we cannot yet differentiate between central and reafferent contributions. Activation of 13As or 13Bs disinhibits motor neurons and that is sufficient to induce walking/grooming. Therefore, we can show a role for the disinhibition motif.

Proprioceptive feedback from leg movements could certainly affect the function of these reciprocal inhibition circuits. Given the synapses we observe between leg proprioceptors and 13A neurons, we think this is likely.

Our previous work (Ravbar et al 2021) showed that grooming rhythms in dusted flies persist when sensory feedback is reduced, indicating that central control is possible. In those experiments, we used dust to stimulate grooming and optogenetic manipulation to broadly silence sensory feedback. We cannot do the same here because we do not yet have reagents to separately activate sparse subsets of inhibitory neurons while silencing specific proprioceptive neurons. More importantly, globally silencing proprioceptors would produce pleiotropic effects and severely impair baseline coordination, making it difficult to distinguish whether observed changes reflect disrupted rhythm generation or secondary consequences of impaired sensory input. Therefore, the reviewer is correct – we do not know whether the effects we observe are feedforward (central), feedback sensory, or both. We have included this in the revised results and discussion section to describe these possibilities and the limits of our current findings.

Additionally, we have used a computational model to test the role of each motif separately and we show that in the results.

Reviewer #2 (Public review):

Summary:

This manuscript by Syed et al. presents a detailed investigation of inhibitory interneurons, specifically from the 13A and 13B hemilineages, which contribute to the generation of rhythmic leg movements underlying grooming behavior in Drosophila. After performing a detailed connectomic analysis, which offers novel insights into the organization of premotor inhibitory circuits, the authors build on this anatomical framework by performing optogenetic perturbation experiments to functionally test predictions derived from the connectome. Finally, they integrate these findings into a computational model that links anatomical connectivity with behavior, offering a systems-level view of how inhibitory circuits may contribute to grooming pattern generation.

Strengths:

(1) Performing an extensive and detailed connectomic analysis, which offers novel insights into the organization of premotor inhibitory circuits.

(2) Making sense of the largely uncharacterized 13A/13B nerve cord circuitry by combining connectomics and optogenetics is very impressive and will lay the foundation for future experiments in this field.

(3) Testing the predictions from experiments using a simplified and elegant model.

We thank the reviewer for their thoughtful and encouraging evaluation of our work. 

Weaknesses:

(1) In Figure 4, while the authors report statistically significant shifts in both proximal inter-leg distance and movement frequency across conditions, the distributions largely overlap, and only in Panel K (13B silencing) is there a noticeable deviation from the expected 7-8 Hz grooming frequency. Could the authors clarify whether these changes truly reflect disruption of the grooming rhythm? 

We reanalyzed the dataset with Linear Mixed Models. We find significant differences in mean frequencies upon silencing these neurons but not upon activation. The experimental groups are also significantly more variable. We revised these panels with updated analysis. We think these data do support our interpretation that the grooming rhythms are disrupted. 

More importantly, all this data would make the most sense if it were performed in undusted flies (with controls) as is done in the next figure.

In our assay conditions, undusted flies groom infrequently. We used undusted flies for some optogenetic activation experiments, where the neuron activation triggers behavior initiation, but we chose to analyze the effect of silencing inhibitory neurons in dusted flies because dust reliably activates mechanosensory neurons and elicits robust grooming behavior enabling us to assess how manipulation of 13A/B neurons alters grooming rhythmicity and leg coordination.

(2) In Figure 4-Figure Supplement 1, the inclusion of walking assays in dusted flies is problematic, as these flies are already strongly biased toward grooming behavior and rarely walk. To assess how 13A neuron activation influences walking, such experiments should be conducted in undusted flies under baseline locomotor conditions.

We agree that there are better ways to assay potential contributions of 13A/13B neurons to walking. We intended to focus on how normal activity in these inhibitory neurons affects coordination during grooming, and we included walking because we observed it in our optogenetic experiments and because it also involves rhythmic leg movements. The walking data is reported in a supplementary figure because we think this merits further study with assays designed to quantify walking specifically. We will make these goals clearer in the revised manuscript and we are happy to share our reagents with other research groups more equipped to analyze walking differences.

(3) For broader lines targeting six or more 13A neurons, the authors provide specific predictions about expected behavioral effects-e.g., that activation should bias the limb toward flexion and silencing should bias toward extension based on connectivity to motor neurons. Yet, when using the more restricted line labeling only two 13A neurons (Figure 4 - Figure Supplement 2), no such prediction is made. The authors report disrupted grooming but do not specify whether the disruption is expected to bias the movement toward flexion or extension, nor do they discuss the muscle target. This is a missed opportunity to apply the same level of mechanistic reasoning that was used for broader manipulations.

Because we cannot unambiguously identify one of the neurons from our sparsest 13A splitGAL4 lines in FANC, we cannot say with certainty which motor neurons they target. That limits the accuracy of any functional predictions.  

(4) Regarding Figure 5: The 70ms on/off stimulation with a slow opsin seems problematic. CsChrimson off kinetics are slow and unlikely to cause actual activity changes in the desired neurons with the temporal precision the authors are suggesting they get. Regardless, it is amazing that the authors get the behavior! It would still be important for the authors to mention the optogenetics caveat, and potentially supplement the data with stimulation at different frequencies, or using faster opsins like ChrimsonR.

We were also intrigued by the behavioral consequences of activating these inhibitory neurons with CsChrimson. We appreciate the reviewer’s point that CsChrimson’s slow off-kinetics limit precise temporal control. To address this, we repeated our frequency analysis using a range of pulse durations (10/10, 50/50, 70/70, 110/110, and 120/120 ms on/off) and compared the mean frequency of proximal joint extension/flexion cycles across conditions. We found no significant difference in frequency (LLMS, p > 0.05), suggesting that the observed grooming rhythm is not dictated by pulse period but instead reflects an intrinsic property of the premotor circuit once activated. We now include these results in ‘Figure 5—figure supplement 1’ and clarify in the text that we interpret pulsed activation as triggering, rather than precisely pacing, the endogenous grooming rhythm. We continue to note in the manuscript that CsChrimson’s slow off-kinetics may limit temporal precision. We will try ChrimsonR in future experiments.

Overall, I think the strengths outweigh the weaknesses, and I consider this a timely and comprehensive addition to the field.

Reviewer #3 (Public review):

Summary:

The authors set out to determine how GABAergic inhibitory premotor circuits contribute to the rhythmic alternation of leg flexion and extension during Drosophila grooming. To do this, they first mapped the ~120 13A and 13B hemilineage inhibitory neurons in the prothoracic segment of the VNC and clustered them by morphology and synaptic partners. They then tested the contribution of these cells to flexion and extension using optogenetic activation and inhibition and kinematic analyses of limb joints. Finally, they produced a computational model representing an abstract version of the circuit to determine how the connectivity identified in EM might relate to functional output. The study, in its current form, makes an important but overclaimed contribution to the literature due to a mismatch between the claims in the paper and the data presented.

Strengths:

The authors have identified an interesting question and use a strong set of complementary tools to address it:

(1) They analysed serial‐section TEM data to obtain reconstructions of every 13A and 13B neuron in the prothoracic segment. They manually proofread over 60 13A neurons and 64 13B neurons, then used automated synapse detection to build detailed connectivity maps and cluster neurons into functional motifs.

(2) They used optogenetic tools with a range of genetic driver lines in freely behaving flies to test the contribution of subsets of 13A and 13B neurons.

(3) They used a connectome-constrained computational model to determine how the mapped connectivity relates to the rhythmic output of the behavior.

Weaknesses:

The manuscript aims to reveal an instructive, rhythm-generating role for premotor inhibition in coordinating the multi-joint leg synergies underlying grooming. It makes a valuable contribution, but currently, the main claims in the paper are not well-supported by the presented evidence.

Major points

(1) Starting with the title of this manuscript, "Inhibitory circuits generate rhythms for leg movements during Drosophila grooming", the authors raise the expectation that they will show that the 13A and 13B hemilineages produce rhythmic output that underlies grooming. This manuscript does not show that. For instance, to test how they drive the rhythmic leg movements that underlie grooming requires the authors to test whether these neurons produce the rhythmic output underlying behavior in the absence of rhythmic input. Because the optogenetic pulses used for stimulation were rhythmic, the authors cannot make this point, and the modelling uses a "black box" excitatory network, the output of which might be rhythmic (this is not shown). Therefore, the evidence (behavioral entrainment; perturbation effects; computational model) is all indirect, meaning that the paper's claim that "inhibitory circuits generate rhythms" rests on inferred sufficiency. A direct recording (e.g., calcium imaging or patch-clamp) from 13A/13B during grooming - outside the scope of the study - would be needed to show intrinsic rhythmogenesis. The conclusions drawn from the data should therefore be tempered. Moreover, the "black box" needs to be opened. What output does it produce? How exactly is it connected to the 13A-13B circuit? 

We modified the title to better reflect our strongest conclusions: “Inhibitory circuits control leg movements during Drosophila grooming”

Our optogenetic activation was delivered in a patterned (70 ms on/off) fashion that entrains rhythmic movements, but this does not rule out the possibility that the rhythm is imposed externally. In the manuscript, we state that we used pulsed light to mimic a flexion-extension cycle and note that this approach tests whether inhibition is sufficient to drive rhythmic leg movements when temporally patterned. While this does not prove that 13A/13B neurons are intrinsic rhythm generators, it does demonstrate that activating subsets of inhibitory neurons is sufficient to elicit alternating leg movements resembling natural grooming and walking.

Our goal with the model was to demonstrate that it is possible to produce rhythmic outputs with this 13A/B circuit, based on the connectome. The “black box” is a small recurrent neural network (RNN) consisting of 40 neurons in its hidden layer. The inputs are the “dust” levels from the environment (the green pixels in Figure 6I), the “proprioceptive” inputs (“efference copy” from motor neurons), and the amount of dust accumulated on both legs. The outputs (all positive) connect to the 13A neurons, the 13B neurons, and to the motor neurons. We refer to it as the “black box” because we make no claims about the actual excitatory inputs to these circuits. Its function is to provide input, needed to run the network, that reflects the distribution of “dust” in the environment as well as the information about the position of the legs.  

The output of the “black box” component of the model might be rhythmic. In fact, in most instances of the model implementation this is indeed the case. However, as mentioned in the current version of the manuscript: “But the 13A circuitry can still produce rhythmic behavior even without those external inputs (or when set to a constant value), although the legs become less coordinated.” Indeed, when we refine the model (with the evolutionary training) without the “black box” (using a constant input of 0.1) the behavior is still rhythmic and sustained. Therefore, the rhythmic activity and behavior can emerge from the premotor circuitry itself without a rhythmic input.

The context in which the 13A and 13B hemilineages sit also needs to be explained. What do we know about the other inputs to the motorneurons studied? What excitatory circuits are there? 

We agree that there are many more excitatory and inhibitory, direct and indirect, connections to motor neurons that will also affect leg movements for grooming and walking. 13A neurons provide a substantial fraction of premotor input. For example, 13As account for ~17.1% of upstream synapses for one tibia extensor (femur seti) motor neuron and ~14.6% for another tibia extensor (femur feti) motor neuron. Our goal was to demonstrate what is possible from a constrained circuit of inhibitory neurons that we mapped in detail, and we hope to add additional components to better replicate the biological circuit as behavioral and biomechanical data is obtained by us and others.  

Furthermore, the introduction ignores many decades of work in other species on the role of inhibitory cell types in motor systems. There is some mention of this in the discussion, but even previous work in Drosophila larvae is not mentioned, nor crustacean STG, nor any other cell types previously studied. This manuscript makes a valuable contribution, but it is not the first to study inhibition in motor systems, and this should be made clear to the reader.

We thank the reviewer for this important reminder.  Previous work on the contribution of inhibitory neurons to invertebrate motor control certainly influenced our research. We have expanded coverage of the relevant history and context in our revised discussion.

(2) The experimental evidence is not always presented convincingly, at times lacking data, quantification, explanation, appropriate rationales, or sufficient interpretation.

We are committed to improving the clarity, rationale, and completeness of our experimental descriptions.  We have revisited the statistical tests applied throughout the manuscript and expanded the Methods.

(3) The statistics used are unlike any I remember having seen, essentially one big t-test followed by correction for multiple comparisons. I wonder whether this approach is optimal for these nested, high‐dimensional behavioral data. For instance, the authors do not report any formal test of normality. This might be an issue given the often skewed distributions of kinematic variables that are reported. Moreover, each fly contributes many video segments, and each segment results in multiple measurements. By treating every segment as an independent observation, the non‐independence of measurements within the same animal is ignored. I think a linear mixed‐effects model (LMM) or generalized linear mixed model (GLMM) might be more appropriate.

We thank the reviewer for raising this important point regarding the statistical treatment of our segmented behavioral data. Our initial analysis used independent t-tests with Bonferroni correction across behavioral classes and features, which allowed us to identify broad effects. However, we acknowledge that this approach does not account for the nested structure of the data. To address this, we re-analyzed key comparisons using linear mixed-effects models (LMMs) as suggested by the reviewer. This approach allowed us to more appropriately model within-fly variability and test the robustness of our conclusions. We have updated the manuscript based on the outcomes of these analyses.

(4) The manuscript mentions that legs are used for walking as well as grooming. While this is welcome, the authors then do not discuss the implications of this in sufficient detail. For instance, how should we interpret that pulsed stimulation of a subset of 13A neurons produces grooming and walking behaviours? How does neural control of grooming interact with that of walking?

We do not know how the inhibitory neurons we investigated will affect walking or how circuits for control of grooming and walking might compete. We speculate that overlapping pre-motor circuits may participate because both have similar extension flexion cycles at similar frequencies, but we do not have hard experimental data to support. This would be an interesting area for future research. Here, we focused on the consequences of activating specific 13A/B neurons during grooming because they were identified through a behavioral screen for grooming disruptions, and we had developed high-resolution assays and familiarity with the normal movements in this behavior.

(5) The manuscript needs to be proofread and edited as there are inconsistencies in labelling in figures, phrasing errors, missing citations of figures in the text, or citations that are not in the correct order, and referencing errors (examples: 81 and 83 are identical; 94 is missing in text).

We have proofread the manuscript to fix figure labeling, citation order, and referencing errors.

Reviewing Editor Comments:

In addition to the recommendations listed below, a common suggestion, given the lack of evidence to support that 13A and 13B are rhythm-generating, is to tone down the title to something like, for example, "Inhibitory circuits control leg movements during grooming in Drosophila" (or similar).

We changed the title to Inhibitory circuits control leg movements during Drosophila  grooming

Reviewer #1 (Recommendations for the authors):

(1) Naming of movements of leg segments:

The authors refer to movements of leg segments across the leg, i.e., of all joints, as "flexion" and "extension". For example, in Figure 4A and at many other places. This naming is functionally misleading for two reasons: (i) the anatomical organization of an insect leg differs in principle from the organization of the mammalian leg, which the manuscript often refers to. While the organization of a mammalian limb is planar the organization of the insect limb shows a different plane as compared to the body length axis (for detailed accounts see Ritzmann et al. 2004; Büschges & Ache, 2024); (ii) the reader cannot differentiate between places in the text, where "flexion" and "extension" refer to movements of the tibia of the femur-tibia joint, e.g. in the graphical abstract, in Figure 3 and its supplements, and other places, e.g. Figure 4 and its supplements, where these two words refer to movements of leg segments of other joints, e.g. thorax-coxa, coxa-trochanter and tarsal joints. The reviewer strongly suggests naming the movements of the leg segments according to the individual joint and its muscles.

We accept this helpful suggestion. We now include a description of the leg segments and joints in the revised Introduction and refer to which leg segments we mean   

“The adult Drosophila leg consists of serially arranged joints—bodywall/thoraco-coxal (Th-C), coxa–trochanter (C-Tr), trochanter–femur (Tr-F), femur–tibia (F-Ti), tibia–tarsus (Ti-Ta)—each powered by opposing flexor and extensor muscles that transmit force through tendons (Soler et al., 2004). The proximal joints, Th-C and C-Tr, mediate leg protraction–retraction and elevation–depression, respectively (Ritzmann et al., 2004; Büschges & Ache, 2025). The medial joint, F-Ti, acts as the principal flexion–extension hinge and is controlled by large tibia extensor motor neurons and flexor motor neurons (Soler et al., 2004; Baek and Mann 2009; Brierley et al., 2012; Azevedo et al., 2024; Lesser et al., 2024). By contrast, distal joints such as Ti-Ta and the tarsomeres contribute to fine adjustments, grasping, and substrate attachment (Azevedo et al., 2024).”

We also clarified femur-tibia joints in the graphical abstract, modified Figure 3 legend and added joints at relevant places.

(2)  Figures 3, 4, and 5 with supplements:

The authors optogenetically silence and activate (sub)populations of 13A and 13B interneurons. Changes in frequency of movements and distance between legs or leg movements are interpreted as the effect of these experimental paradigms. No physiological recordings from leg motoneurons or leg muscles are shown. While I understand the notion of the authors to interpret a movement as the outcome of activity in a muscle, it needs to be remembered that it is well known that fast cyclic leg movements, including those for grooming, cannot be used to conclude on the underlying neural activity. Zakotnik et al. (2006) and others provided evidence that such fast cyclic movements can result from the interaction of the rhythmic activity of one leg muscle only, together with the resting tension of its silent antagonist. Given that no physiological recordings are presented, this needs to be mentioned in the discussion, e.g., in the section "Inhibitory Innervation Imbalance.......".

Added studies from Heitler, 1974; Bennet-Clark, 1975; Zakotnik et al., 2006; Page et al., 2008 in discussion.

(3) Introduction and Discussion:

The authors refer extensively to work on the mammalian spinal cord and compare their own work with circuit elements found in the spinal cord. From the perspective of the reviewer this notion is in conflict with acknowledging prior research work on the role of inhibitory network interactions for other invertebrates and lower vertebrates: such are locust flight system (for feedforward inhibition, disinhibition), crustacean stomatogastric nervous system (reciprocal inhibition), clione swimming system (reciprocal inhibition, feedforward inhibition, disinhibition), leech swimming system (reciprocal inhibition, disinhibition, feedforward inhibition), xenopus swimming system (reciprocal inhibition). The next paragraph illustrates this criticism/suggestion for stick insect neural circuits for leg stepping.

(4) Discussion:

"Feedforward inhibition" and "Disinhibition": it is already been described that rhythmic activity of antagonistic insect leg motoneuron pools arises from alternating synaptic inhibition and disinhibition of the motoneurons from premotor central pattern generating networks, e.g., Büschges (1998); Büschges et al. (2004); Ruthe et al. (2024).

We have added these references to the revised Discussion.

(5) Circuit motifs of the simulation, i.e., mutual inhibition between interneurons and onto motoneurons and sensory feedback influences and pathways share similarities to those formerly used by studies simulating rhythmic insect leg movements, for example, Schilling & Cruse 2020, 2023 or Toth et al. 2012. For the reader, it appears relevant that the progress of the new simulation is explained in the light of similarities and differences to these former approaches with respect to the common circuit motifs used.

We now put our work in the context of other models in the Discussion section: “Similar circuit motifs, namely reciprocal inhibitions between pre-motor neurons and the sensory feedback have been modeled before, in particular neuroWalknet, and such simple motifs do not require a separate CPG component to generate rhythmic behavior in these models (Schilling & Cruse 2020, 2023). However, our model is much simpler than the neuroWalknet - it controls a 2D agent operating on an abstract environment (the dust distribution), without physics. In real animals or complex mechanical models such as NeuroMechFly (Lobato-Rios et al), a more explicit central rhythm generation may be advantageous for the coordination across many more degrees of freedom.”

Reviewer #2 (Recommendations for the authors):

I might have missed this, but I couldn't find any mention of how the grooming command pathways, described by previous work from the authors' lab, recruit these predicted grooming pattern-generating neurons. This should be mentioned in the connectome analysis and also discussed later in the discussion.

13A neurons are direct downstream targets of previously described grooming command neurons. Specifically, the antennal grooming command neuron aDN (Hampel et al., 2015) synapses onto two primary 13As (γ and α; 13As-i) that connect to proximal extensor and medial flexor motor neurons, as well as four other 13As (9a, 9c, 9i, 6e) projecting to body wall extensor motor neurons. The 13As-i also form reciprocal connections with 13As-ii, providing a potential substrate for oscillatory leg movements. aDN connects to homologous 13As on both sides, consistent with the bilateral coordination needed for antennal sweeping. 

The head grooming/leg rubbing command neuron DNg12 (Guo et al., 2022)  synapses directly onto ~50 13As, predominantly those connected to proximal motor neurons. 

While sometimes the structural connectivity suggests pathways for generating rhythmic movements, the extensive interconnections among command neurons and premotor circuits indicate that multiple motifs could contribute to the observed behaviors. Further work will be needed to determine how these inputs are dynamically engaged during normal grooming sequences. We have now added it to the discussion.

I encourage the authors to be explicit about caveats wherever possible: e.g., ectopic expression in genetic tools, potential for other unexplored neurons as rhythm generators (rather than 13A/B), given that the authors never get complete silencing phenotypes, CsChrimson kinetics, neurotransmitter predictions, etc.

We now explain these caveats as follows: Ectopic expression is noted in Figure 1—figure supplement 1, and we added the following to the Discussion: “While our experiments with multiple genetic lines labeling 13A/B neurons consistently implicate these cells in leg coordination, ectopic expression in some lines raises the possibility that other neurons may also contribute to this phenotype. In addition, other excitatory and inhibitory neural circuits, not yet identified, may also contribute to the generation of rhythmic leg movements. Future studies should identify such neurons that regulate rhythmic timing and their interactions with inhibitory circuits.”

We also added a caveat regarding CsChrimson kinetics in the Results. Finally, our identification of these neurons as inhibitory is based on genetic access to the GABAergic population (we use GAD-spGAL4 as part of the intersection which targets them), rather than on predictions of neurotransmitter identity.

Reviewer #3 (Recommendations for the authors):

Detailed list of figure alterations:

(1) Figure 1:

(a) Figure 1B and Figure 1 - Figure Supplement 1 lack information on individual cells - how can we tell that the cells targeted are indeed 13A and 13B, and which ones they are? Since off-target expression in neighboring hemilineages isn't ruled out, the interpretation of results is not straightforward.

The neurons labeled by R35G04-DBD and GAD1-AD are identified as 13A and 13B based on their stereotyped cell body positions and characteristic neurite projections into the neuropil, which match those of 13A and 13B neurons reconstructed in the FANC and MANC connectome. While we have not generated flip-out clones in this genotype, we do isolate 13A neurons more specifically later in the manuscript using R35G04-DBD intersected with Dbx-AD, and show single-cell morphology consistent with identified 13A neurons. The purpose of including this early figure was to motivate the study by showing that silencing this population, which includes 13A/13B neurons, strongly reduces grooming in dusted flies. 

Regarding Figure 1—Figure Supplement 1:

This figure showed the expression patterns of all lines used throughout the manuscript. Panels C and D illustrated lines with minimal to no ectopic expression. Panels A and B show neurons with posterior cell bodies that may correspond to 13A neurons not reconstructed in our dataset but described in Soffers et al., 2025 and Marin et al., 2025 and we have provided detailed information about all VNC expressions in the figure legend.

(b) Figure 1D lacks explanation of boxplots, asterisks, genotypes/experimental design.

Added.

(c) Figures 1E-F and video 1 lack quantification, scale bars.

Added quantification.

(2) Figure 2:

(a) Figure 2A, Figure 2 - Supplement 3: What are the details of the hierarchical clustering? What metric was used to decide on the number of clusters? 

We have used FANC packages to perform NBLAST clustering (Azevedo et al., 2024, Nature). We now include the full protocol in Methods.  The details are as follows:

We performed hierarchical clustering on pairwise NBLAST similarity scores computed using navis.nblast_allbyall(). The resulting similarity matrix was symmetrized by averaging it with its transpose, and converted into a distance matrix using the transformation:

distance=(1−similarity)\text{distance} = (1 - \text{similarity})distance=(1−similarity)

This ensures that a perfect NBLAST match (similarity = 1) corresponds to a distance of 0.

Clustering was performed using Ward’s linkage method (method='ward' in scipy.cluster.hierarchy.linkage), which minimizes the total within-cluster variance and is well-suited for identifying compact, morphologically coherent clusters.

We did not predefine the number of clusters. Instead, clusters were visualized using a dendrogram, where branch coloring is based on the default behavior of scipy.cluster.hierarchy.dendrogram(). By default, this function applies a visual color threshold at 70% of the maximum linkage distance to highlight groups of similar elements. In our dataset, this corresponded to a linkage distance of approximately 1–1.5, which visually separated morphologically distinct neuron types (Figures 2A and Figure 2—figure supplement 3A). This threshold was used only as a visual aid and not as a hard cutoff for quantitative grouping.

The Methods section says that the classification "included left-right comparisons". What does that mean? What are the implications of the authors only having proofread a subset of neurons in T1L (see below)? 

All adult leg motor neurons and 13A neurons (except one, 13A-ε) have neurite arbors restricted to the local, ipsilateral neuropil associated with the nearest leg.  Although 13B neurons have contralateral cell bodies, their projections are also entirely ipsilateral. The Tuthill Lab, with contributions from our group, focused proofreading efforts on the left front neuropil (T1L) in FANC. This is also where the motor neuron to muscle mapping has been most extensively done. We reconstructed/proofread the 13A and 13B neurons from the right side as well (T1R). We see similar clustering based on morphology and connectivity here as well.  

Reconstructions lack scale bars and information on orientation (also in other figures), and the figures for the 13B analysis are not consistent with the main figure (e.g., labelling of clusters in panel B along x,y axes).

Added.  

(b) Figure 2B: Since the cosine similarity matrix's values should go from -1 to 1, why was a color map used ranging from 0 to 1? 

While cosine similarity values can theoretically range from -1 to 1, in our case, all vector entries (i.e., synaptic weights) are non-negative, as they reflect the number of synapses from each 13A neuron to its downstream targets. This means all pairwise cosine similarities fall within the 0 to 1 range. 

Why are some neurons not included in this figure, like 1g, 2b, 3c-f (also in Supplement 3)?

The few 13A neurons that don’t connect to motor neurons are not shown in the figure.

(c) Figures 2C and D: the overlaid neurites are difficult to distinguish from one another. If the point here is to show that each 13A neuron class innervates specific motor neurons, then this is not the clearest way of doing that. For instance, the legend indicates that extensors are labelled in red, and that MNs with the highest number of synapses are highlighted in red - does that work? I could not figure out what was going on. On a more general point: if two cells are connected, does that not automatically mean that they should overlap in their projection patterns?

We intended these panels to illustrate that 13A neurons synapse onto overlapping regions of motor neurons, thereby creating a spatial representation of muscle targets. However, we agree that overlapping multiple neurons in a single flat projection makes the figure difficult to interpret. We have therefore removed Figures 2C and 2D.

While neurons must overlap at least somewhere if they form a synaptic connection, the amount of their neurites that overlap can vary, and more extensive overlap suggests more possible connections. Because the synapses are computationally predicted, examining the overlap helps to confirm that these predictions are consistent.

While connected neurons must overlap locally at their synaptic sites, they do not necessarily show extensive or spatially structured overlap of their projections. For example, descending neurons or 13B interneurons may form synapses onto motor neurons without exhibiting a topographically organized projection pattern. In contrast, 13A→MN connectivity is organized in a structured manner: specialist 13A neurons align with the myotopic map of MN dendrites, whereas generalist 13As project more broadly and target MN groups across multiple leg segments, reflecting premotor synergies. This spatial organization—combining both joint-specific and multi-joint representations—was a key finding we wished to highlight, and we have revised the Results text to make this clearer.

(d) Figure 2 - Figure Supplement 1: Why are these results presented in a way that goes against the morphological clustering results, but without explanation? Clusters 1-3 seem to overlap in their connectivity, and are presented in a mixed order. Why is this ignored? Are there similar data for 13B?

The morphological clusters 1–3 do exhibit overlapping connectivity, but this is consistent with both their anatomical similarity and premotor connectivity. Specifically, Cluster 1 neurons connect to SE and TrE motor neurons, Cluster 2 connects only to TrE motor neurons, and Cluster 3 targets multiple motor pools, including SE and TrE (Figure 2—Figure Supplement 1B). This overlap is also reflected in the high pairwise cosine similarity among Clusters 1–3 shown in Figure 2B. Thus, their similar connectivity profiles align with their proximity in the NBLAST dendrogram.

Regarding 13B neurons: there is no clear correlation between morphological clusters and downstream motor targets, as shown in the cosine similarity matrix (Figure 2—figure supplement 3). Moreover, even premotor 13B neurons that fall within the same morphological cluster do not connect to the same set of motor neurons (Figure 3—figure supplement 1F). For example, 13B-2a connects to LTrM and tergo-trochanteral MNs, 13B-2b connects to TiF MNs, and 13B-2g connects to Tr-F, TiE, and tergo-T MNs. Together, these results demonstrate that 13A neurons are spatially organized in a manner that correlates with their motor neuron targets, whereas 13B neurons lack such spatially structured organization, suggesting distinct principles of connectivity for these two inhibitory premotor populations.

(e) Figure 2 - Figure Supplement 2: A comparison is made here between T1R (proofread) and T1L (largely not proofread). A general point is made here that there are "similar numbers of neurons and cluster divisions". First, no quantitative comparison is provided, making it difficult to judge whether this point is accurate. Second, glancing at the connectivity diagram, I can identify a large number of discrepancies. How should we interpret those? Can T1L be proofread? If this is too much of a burden, results should be presented with that as a clear caveat.

The 13A and 13B neurons in the T1L hemisegment are fully proofread (Lesser et al, 2024, current publication); the T1R has been extensively analyzed as well.  To compare the clustering and match identities of 13A and 13B neurons on the left and the right, We mirrored the 13A neurons from the left side and used NBLAST to match them with their counterparts on the right.

While individual synaptic counts differ between sides in the FANC dataset (T1L generally showing higher counts), the number of 13A neurons, their clustering, and the overall patterns of connectivity are largely conserved between T1L and T1R.

Importantly, each 13A cluster targets the same subset of motor neurons on both sides, preserving the overall pattern of connectivity. The largest divergence is seen in cluster 9, which shows more variable connectivity.  

(f) Figure 2 - Figure Supplements 4 & 5: Why did the authors choose to present the particular cell type in Supplement 4?  Why are the cell types in Supplement 5 presented differently? Labels in Supplement 5 are illegible, but I imagine this is due to the format of the file presented to reviewers. Why are there no data for 13B?

We chose to present the particular cell type in Supplement 4 because it corresponds to cell types targeted in the genetic lines used in our behavioral experiments. The 13A neuron shown is also one of the primary neurons in this lineage. This example illustrates its broader connectivity beyond the inhibitory and motor connections emphasized in the main figures.

In Supplement 5, we initially aimed to highlight that the major downstream targets of 13A neurons are motor neurons. We have now removed this figure and instead state in the text that the major downstream targets are MNs.

We did not present 13B neurons in the same format because their major downstream targets are not motor neurons. Instead, we emphasize their role in disinhibition and their connections to 13A neurons, as shown in a specific example in Figure 3—figure supplement 2. This 13B neuron also corresponds to a cell type targeted in the genetic line used in our behavioral experiments.

(3) Figure 3:

(a) Figure 3A: the collection of diagrams is not clear. I'd suggest one diagram with all connections included repeated for each subpanel, with each subpanel highlighting relevant connections and greying out irrelevant ones to the type of connection discussed. The nomenclature should be consistent between the figure and the legend (e.g., feedforward inhibition vs direct MN inhibition in A1.

The intent of Figure 3A is to highlight individual circuit motifs by isolating them in separate panels. Including all connections in every sub panel would likely reduce clarity and make it harder to follow each motif. For completeness, we show the full set of connections together in Panel D. We updated the nomenclature as suggested. 

(b) Figure 3B: Why was the medial joint discussed in detail? Do the thicknesses of the lines represent the number of synapses? There should be a legend, in that case. Why are the green edges all the same thickness? Are they indeed all connected with a similarly low number of synapses?

We focused on the medial joint (femur-tibia joint) because it produces alternating flexion and extension of the tibia during both head sweeps and leg rubbing, which are the main grooming actions we analyzed. During head grooming, the tarsus is typically suspended in the air, so the cleaning action is primarily driven by tibial movements generated at the medial joint. 

The thickness of the edges represents the number of synapses, and we have now clarified this in the legend. The green edges represent connections from 13B neurons, which were manually added to the graph, as described in the Methods section. 13B neurons are smaller than 13A neurons and form significantly fewer total downstream synapses. For example, the 13B neuron shown in Figure 3—figure supplement 2 makes a total of 155 synapses to all downstream neurons, with only 22 synapses to its most strongly connected partner, a 13A neuron. The relatively sparse connectivity of 13B neurons is shown in thinner or uniform edge weights in this graph.

(C) Figure 3C: This is a potentially important panel, but the connections are difficult to interpret. Moreover, the text says, "This organizational motif applies to multiple joints within a leg as reciprocal connections between generalist 13A neurons suggest a role in coordinating multi-joint movements in synergy". To what extent is this a representative result? The figure also has an error in the legend (it is not labelled as 3C).

This statement is true and based on the connectivity of these neurons. We now added

“Data for 13A-MN connections shown in Figure 2—figure supplement 1 I9, I6, I7, H9, H4, and H5; 13A-13A connections shown in Figure 3—figure supplement 1C.” to the figure legend.

Thanks, we fixed the labelling error.

(d) Figure 3 - Figure Supplement 1: Panel A is very difficult to interpret. Could a hierarchical diagram be used, or some other representation that is easier to digest?

Panel A provides a consolidated view of all upstream and downstream interconnections among individual 13A and 13B neurons, allowing readers to quickly assess which neurons connect to which others without having to examine all subpanels. For a hierarchical representation, we have provided individual neuron-level diagrams in Panels C–F. 

(e) Figure 3 - Figure Supplement 2: Why was this cell type selected?

We selected this 13B because it is involved in the disinhibition of 13A neurons and is also present in the genetic line used for our behavioral experiments. 

(f) Figure 3 - Figure Supplement 3: The diagram is confusing, with text aligned randomly, and colors lacking some explanations. Legend has odd formatting.

The diagram layout and text alignment are designed to reflect the logical grouping of proprioceptors, 13A neurons, and motor neurons. To improve clarity, we have added node colors, included a written explanation for edge colors, and corrected the formatting of the figure legend.

(4) Figure 4:

(a) Figure 4A: This has no quantification, poor labelling, and odd units (centiseconds?). The colours between the left and right panels also don't align.

We have fixed these issues.

(b) Figure 4D-K: The ranges on the different axes are not the same (e.g., y axis on box plots, x axis on histograms). This obscures the fact that the differences between experimental and control, which in many cases are not big, are not consistent between the various controls. Moreover, the data that are plotted are, as far as I can tell (which is also to say: this should be explained), one value per frame. With imaging at 100Hz, this means that an enormous number of values are used in each analysis. Very small differences can therefore be significant in a statistical sense. However, how different something is between conditions is important (effect size), and this is not taken int account in this manuscript. For instance, in 4D-J, the differences in the mean seem to be minimal. Should that not be taken into consideration? A point in case is panel D in Figure 4 - Figure Supplement 1: even with near identical distributions, a statistically significant difference is detected. The same applies to Figure 4 - Figure Supplements 1-3. Also, what do the boxes and whiskers in the box plots show, exactly?

We have re-plotted all summary panels using linear mixed-effects models (LMMs) as suggested. In the updated plots, each dot represents the mean value for a single animal, and bar height represents the group mean. Whiskers indicate the 95% confidence interval around the group mean. This approach avoids inflating sample size by using per-frame values and provides a more accurate view of both variability and effect size. 

(e) Figure 4 - Figure Supplement 1: There are 6 cells labelled in the split line; only 4 are shown in A3. Is cluster 6 a convincing match between EM and MCFO?

We indeed report four neurons targeted by the split-GAL4 line in flip out clones. Generating these clones was technically challenging. In our sample (n=23), we may not have labeled all of the neurons.  Alternatively, two neurons may share very similar morphology and connectivity, making it difficult to tell them apart. We have added this clarification to the revised figure legend.

It is interesting to see data on walking in panel K, but why were these analyses not done on any of the other manipulations? What defect produced the reduction in velocity, exactly? How should this be interpreted?

Our primary focus was on grooming, but we did observe changes in walking, so we report illustrative examples. We initially included a panel showing increased walking velocity upon 13A activation, but this effect did not survive FDR correction and was removed in the revised version. We instead included data for 13A silencing which did not affect the frequency of joint movements during walking. However, spatial aspects of walking were affected: the distance between front leg tips during stance was reduced, indicating that although flies continued to walk rhythmically, the positioning of the legs was altered. This suggests that these specific 13A neurons may influence coordination and limb placement during walking without disrupting basic rhythmicity. As reviewer #2 also noted, dust may itself affect walking, so we have chosen not to further pursue this aspect in the current study.

(f) Figure 4 - Figure Supplement 2: panel A is identical to Figure 1 - Figure Supplement 1C. This figure needs particular attention, both in content and style. Why present data on silencing these neurons in C-D, but not in E-F?

We removed the panel Figure 1 - Figure Supplement 1C and kept it in Figure 4 - Figure Supplement 2 A. E-F also shows data on silencing, as C’.

(g) Figure 4 - Figure Supplement 3: In panel B, the authors should more clearly demonstrate the identity of 4b and 4a. Why present such a limited number of parameters in F and G?

The cells shown in panel B represent the best matches we could identify between the light-level expression pattern and EM reconstructions. In panels F and G, we focused on bout duration, as leg position/inter-leg distance and frequency were already presented (in Figure 4). Together, these parameters demonstrate the role of 13B neurons in coordinating leg movements. Maximum angular velocity of proximal joints was not significantly affected and is therefore not included.

(5) Figure 5:

(a) Figure 5B: Lacks a quantification of the periodic nature of the behavior, which is required to compare to experimental conditions, e.g., in panel C.

Added

(b) Figure 5C: Requires a quantification; stimulus dynamics need to be incorporated.

Added

(c) Figure 5D: More information is needed. Does "Front leg" mean "leg rub", and "Head" "head sweep"? How do the dynamics in these behaviors compare to normal grooming behavior?

Yes, head grooming is head sweeps and Front leg grooming is leg rub. Comparison added, shown in 5E-F

(d) Figure 5E: How should we interpret these plots? Do these look like normal grooming/walking?

We have now included the comparison.

(e) Figure 5F: Needs stats to compare it to 5B'.

Done

(6) Figure 6:

(a) Figure 6A: I think the circuit used for the model is lacking the claw/hook extension - 13Bs connection. Any other changes? What is the rationale?

13Bs upstream of these particular 13As do not receive significant connections from claw/hook neurons (there’s only one ~5 synapses connection from one hook extension to one 13B neurons, which we neglected for the modeling purpose). 

(b) Figure 6B and C: Needs labels, legend; where is 13B?

In the figure legend we now added: “The 13B neurons in this model do not connect to each other, receive excitatory input from the black box, and only project to the 13As (inhibitory). Their weight matrix, with only two values, is not shown.” We added the colorbar and corrected the color scheme.

(c) Figure 6D-H: plots are very difficult to interpret. Units are also missing (is "Time" correct?).

The units are indeed Time in frames (of simulation). We added this to the figure and the legend. We clarified the units of all variables in these panels. Corrected the color scheme and added their meaning to the legend text.

(d) Figure 6I: I think the authors should consider presenting this in a different format.

(e)  Figure 6 J and K (also Figure Supplement): lacks labels.

We added labels for the three joints, increased the size of fonts for clarity, and added panel titles on the top.

More specific suggestions:

(1) It would be helpful if the titles of all figures reflected the take-away message, like in Figure 2.

(2) "Their dendrites occupy a limited region of VNC, suggesting common pre-synaptic inputs" - all dendrites do, so I'd suggest rephrasing to be more precise.

(3) "We propose that the broadly projecting primary neurons are generalists, likely born earlier, while specialists are mostly later-born secondary neurons" - this needs to be explained.

We added the explanation.

We propose that the broadly projecting primary neurons are generalists, likely born earlier, while specialists are mostly later-born secondary neurons. This is consistent with the known developmental sequence of hemilineages, where early-born primary neurons typically acquire larger arbors and integrate across broader premotor and motor targets, whereas later-born secondary neurons often have more spatially restricted projections and specialized roles[18,19,81,82,85]. Our morphological clustering supports this idea: generalist 13As have extensive axonal arbors spanning multiple leg segments, whereas specialist neurons are more narrowly tuned, connecting to a few MN targets within a segment. Thus, both their morphology and connectivity patterns align with the expectation from birth-order–dependent diversification within hemilineages.

(4) "We did not find any correlation between the morphology of premotor 13B and motor connections" - this needs to be explained, as morphology constrains connectivity.

We agree that morphology often constrains connectivity. However, in contrast to 13A neurons—where morphological clusters strongly predict MN connectivity—we did not observe such a correlation for 13B neurons. As we noted in our response to comment 2d, 13B neurons can form synapses onto MNs without exhibiting extensive or spatially structured overlap of their axonal projections with MN dendrites. This suggests that 13B→MN connectivity may be governed by more local, synapse-specific rules rather than by large-scale morphological positioning, in contrast to the spatially organized premotor map we observe for 13As.

(5) "Based on their connectivity, we hypothesized that continuously activating them might reduce extension and increase flexion. Conversely, silencing them might increase extension and reduce flexion." - these clear predictions are then not directly addressed in the results that follow.

We have now expanded this section.

(6) "Thus, 13A neurons regulate both spatial and temporal aspects of leg coordination" "Together, 13A and 13B neurons contribute to both spatial and temporal coordination during grooming" - are these not intrinsically linked? This needs to be explained/justified.

The spatial (leg positioning, joint angles) and temporal (frequency, rhythm) aspects are often linked, but they can be at least partially dissociated. This has been shown in other systems: for example, Argentine ants reduce walking speed on uneven terrain primarily by decreasing stride frequency while maintaining stride length (Clifton et al., 2020), and Drosophila larvae adjust crawling speed mainly by modulating cycle period rather than the amplitude of segmental contractions (Heckscher et al., 2012). Consistent with these findings, we observe that 13A neuron manipulation in dusted flies significantly alters leg positioning without changing the frequency of walking cycles. Thus, leg positioning can be perturbed while the number of extension–flexion cycles per second remains constant, supporting the view that spatial and temporal features are at least partially dissociable.

(7) "Connectome data revealed that 13B neurons disinhibit motor pools (...) One of these 13B neurons is premotor, inhibiting both proximal and tibia extensor MN" - these are not possible at the same time.

We show that the 13B population contains neurons with distinct connectivity motifs:

some inhibit premotor 13A neurons (leading to disinhibition of motor pools), while others directly inhibit motor neurons. The split-GAL4 line we use labels three 13B neurons—two that inhibit the primary 13A neuron 13A-9d-γ (which targets proximal extensor and medial flexor MNs) and one that is premotor, directly inhibiting both proximal and tibia extensor MNs. Although these functions may appear mutually exclusive, their combined action could converge to a similar outcome: disinhibition of proximal extensor and medial flexor MNs while simultaneously inhibiting medial extensor MNs. This suggests that the labeled 13B neurons act in concert to bias the network toward a specific motor state rather than producing contradictory effects.

(8) "we often observed that one leg became locked in flexion while the other leg remained extended, (indicating contribution from additional unmapped left right coordination circuits)." - Are these results not informative? I'd suggest the authors explain the implications of this more, rather than mentioning it within brackets like this.

We agree with the reviewer that these results are highly informative. The observation that one leg can remain locked in flexion while the other stays extended suggests that additional left–right coordination circuits are engaged during grooming. This cross-talk is likely mediated by commissural interneurons downstream of inhibitory premotor neurons, which have not yet been systematically studied. Dissecting these circuits will require a dedicated project combining bilateral connectomic reconstruction, studying downstream targets of these commissural neurons, and functional interrogation, which is beyond the scope of the current study.

(9) "Indeed, we observe that optogenetic activation of specific 13A and 13B neurons triggers grooming movements. We also discover that" - this phrasing suggests that this has already been shown.external

We replaced ‘indeed’ with “Consistent with this connectivity,”

(10) "But the 13A circuitry can still produce rhythmic behavior even without those  sensory inputs (or when set to a constant value), although the legs become less coordinated." - what does this mean?

We can train (fine-tune) the model without the descending inputs from the “black box” and the behavior will still be rhythmic, meaning that our modeled 13A circuit alone can produce rhythmic behavior, i.e. the rhythm is not generated externally (by the “black box”). We added Figure 7 to the MS and re-wrote this paragraph. In the revised manuscript we now state: “But the 13A circuitry can still produce rhythmic behavior even without those excitatory inputs from the “black box” (or when set to a constant value), although the legs become less coordinated (because they are “unaware” of each other’s position at any time). Indeed, when we refine the model (with the evolutionary training) without the “black box” (using instead a constant input of 0.1) the behavior is still rhythmic although somewhat less sustained (Figure 7). This confirms that the rhythmic activity and behavior can emerge from the modeled pre-motor circuitry itself, without a rhythmic input.”

(11) "However, to explore the possibility of de novo emergent periodic behavior (without the direct periodic descending input) we instead varied the model's parameters around their empirically obtained values." - why do the authors not show how the model performs without tuning it first? What are the changes exactly that are happening as a result of the tuning? Are there specific connections that are lost? Do I interpret Figure 6B and C correctly when I think that some connections are lost (e.g., an SN-MN connection)? How does that compare to the text, which states that "their magnitudes must be at least 80% of the empirical weights"?

Without the fine-tuning we do not get any behavior (the activation levels saturate). So, we tolerate 20% divergence from the empirically established weights and we keep the signs the same. However, in the previous version we allowed the weights to decrease below 20% of the empirical weight (as long as the sign didn’t change) but not above (the signs were maintained and synapses were not added or removed). We thank the reviewer for observing this important discrepancy. In the current version we ensured that the model’s weights are bounded in both directions (the tolerance = 0.2), but we also partially relaxed the constraint on adjacency matrix re-scaling (see Methods, the “The fine-tuning of the synaptic weights” section, where we now clarify more precisely how the evolving model is fitted to the connectome constraints). We then re-ran the fine-tuning process. The Figure 6B and C is now corrected with the properly constrained model, as well as other panels in the figure.  We also applied a better color scheme (now, blue is inhibitory and red is excitatory) for Fig. 6B and C.

(12) "Interestingly, removing 13As-ii-MN connections to the three MNs (second row of the 13A → MN matrices in Figures 6B and C) does not have much effect on the leg movement (data not shown). It seems sufficient for this model to contract only one of the two antagonistic muscles per joint, while keeping the other at a steady state." - this is not clear.

We repeated this test with the newly fine-tuned model and re-wrote the result as follows:  “...when we remove just the 13A-i-MN connections (which control the flexors of the right leg) we likewise get a complete paralysis of the leg. However, removing the 13A-ii-MN (which control the extensors of the right leg) has only a modest effect on the leg movement. So, we need the 13A-i neurons to inhibit the flexors (via motor neurons), but not extensors, in order to obtain rhythmic movements.”

(13) The Discussion needs to reference the specific Results in all relevant sections.

We have revised the discussion to explicitly reference the specific results.

(14) "Flexors and extensors should alternate" - there are circumstances in which flexors and extensors should co-contract. For instance, co-contraction modulates joint stiffness for postural stability and helps generate forces required for fast movements.

Thanks for pointing this out. We added “However, flexor–extensor co-contraction can also be functionally relevant, such as for modulating joint stiffness during postural stabilization or for generating large forces required for fast movements (Zakotnik et al., 2006; Günzel et al., 2022; Ogawa and Yamawaki 2025). Some generalist 13A neurons could facilitate co-contraction across different leg segments, but none target antagonistic motor neurons controlling the same joint. Therefore, co-contraction within a single joint would require the simultaneous activation of multiple 13A neurons.”

(15) "While legs alternate between extension and flexion, they remain elevated during grooming. To maintain this posture, some MNs must be continuously activated while their antagonists are inactivated." - this is not necessarily correct. Small limbs, like those of Drosophila, can assume gravity-independent rest angles (10.1523/JNEUROSCI.5510-08.2009).

We added it to discussion

(16) The discussion "Spatial Mapping of premotor neurons in the nerve cord" seems to me to be making obvious points, and does not need to be included.

We have now revised this section to highlight the significance of 13A spatial organization, emphasizing premotor topographic mapping, multi-joint movement modules, and parallels to myotopic, proprioceptive, and vertebrate spinal maps.

(17) Key point, albeit a small one: "Normal activity of these inhibitory neurons is critical for grooming" - the use of the word critical is problematic, and perhaps typical of the tone of the manuscript. These animals still groom when many of these neurons are manipulated, so what does "critical" really mean?

In this instance, we now changed “critical” to “important”. We observed that silencing or activating a large number (>8) 13A neurons or few 13A and B neurons together completely abolishes grooming in dusted flies as flies get paralyzed or the limbs get locked in extreme poses. Therefore we think we have a justification for the statement that these neurons are critical for grooming.  These neurons may contribute to additional behaviors, and there may be partially redundant circuits that can also support grooming. We have revised the manuscript  with the intention of clarifying both what we have observed and the limits.

This was before the acrophonic principle was introduced which associated certain pictures with certain sounds

Author Response:

Evaluation Summary:

Since DBS of the habenula is a new treatment, these are the first data of its kind and potentially of high interest to the field. Although the study mostly confirms findings from animal studies rather than bringing up completely new aspects of emotion processing, it certainly closes a knowledge gap. This paper is of interest to neuroscientists studying emotions and clinicians treating psychiatric disorders. Specifically the paper shows that the habenula is involved in processing of negative emotions and that it is synchronized to the prefrontal cortex in the theta band. These are important insights into the electrophysiology of emotion processing in the human brain.

The authors are very grateful for the reviewers’ positive comments on our study. We also thank all the reviewers for the comments which has helped to improve the manuscript.

Reviewer #1 (Public Review):

The study by Huang et al. report on direct recordings (using DBS electrodes) from the human habenula in conjunction with MEG recordings in 9 patients. Participants were shown emotional pictures. The key finding was a transient increase in theta/alpha activity with negative compared to positive stimuli. Furthermore, there was a later increase in oscillatory coupling in the same band. These are important data, as there are few reports of direct recordings from the habenula together with the MEG in humans performing cognitive tasks. The findings do provide novel insight into the network dynamics associated with the processing of emotional stimuli and particular the role of the habenula.

Recommendations:

How can we be sure that the recordings from the habenula are not contaminated by volume conduction; i.e. signals from neighbouring regions? I do understand that bipolar signals were considered for the DBS electrode leads. However, high-frequency power (gamma band and up) is often associated with spiking/MUA and considered less prone to volume conduction. I propose to also investigate that high-frequency gamma band activity recorded from the bipolar DBS electrodes and relate to the emotional faces. This will provide more certainty that the measured activity indeed stems from the habenula.

We thank the reviewer for the comment. As the reviewer pointed out, bipolar macroelectrode can detect locally generated potentials, as demonstrated in the case of recordings from subthalamic nucleus and especially when the macroelectrodes are inside the subthalamic nucleus (Marmor et al., 2017). However, considering the size of the habenula and the size of the DBS electrode contacts, we have to acknowledge that we cannot completely exclude the possibility that the recordings are contaminated by volume conduction of activities from neighbouring areas, as shown in Bertone-Cueto et al. 2019. We have now added extra information about the size of the habenula and acknowledged the potential contamination of activities from neighbouring areas through volume conduction in the ‘Limitation’:

"Another caveat we would like to acknowledge that the human habenula is a small region. Existing data from structural MRI scans reported combined habenula (the sum of the left and right hemispheres) volumes of ~ 30–36 mm3 (Savitz et al., 2011a; Savitz et al., 2011b) which means each habenula has the size of 2~3 mm in each dimension, which may be even smaller than the standard functional MRI voxel size (Lawson et al., 2013). The size of the habenula is also small relative to the standard DBS electrodes (as shown in Fig. 2A). The electrodes used in this study (Medtronic 3389) have electrode diameter of 1.27 mm with each contact length of 1.5 mm, and contact spacing of 0.5 mm. We have tried different ways to confirm the location of the electrode and to select the contacts that is within or closest to the habenula: 1.) the MRI was co-registered with a CT image (General Electric, Waukesha, WI, USA) with the Leksell stereotactic frame to obtain the coordinate values of the tip of the electrode; 2.) Post-operative CT was co-registered to pre-operative T1 MRI using a two-stage linear registration using Lead-DBS software. We used bipolar signals constructed from neighbouring macroelectrode recordings, which have been shown to detect locally generated potentials from subthalamic nucleus and especially when the macroelectrodes are inside the subthalamic nucleus (Marmor et al., 2017). Considering that not all contacts for bipolar LFP construction are in the habenula in this study, as shown in Fig. 2, we cannot exclude the possibility that the activities we measured are contaminated by activities from neighbouring areas through volume conduction. In particular, the human habenula is surrounded by thalamus and adjacent to the posterior end of the medial dorsal thalamus, so we may have captured activities from the medial dorsal thalamus. However, we also showed that those bipolar LFPs from contacts in the habenula tend to have a peak in the theta/alpha band in the power spectra density (PSD); whereas recordings from contacts outside the habenula tend to have extra peak in beta frequency band in the PSD. This supports the habenula origin of the emotional valence related changes in the theta/alpha activities reported here."

We have also looked at gamma band oscillations or high frequency activities in the recordings. However, we didn’t observe any peak in high frequency band in the average power spectral density, or any consistent difference in the high frequency activities induced by the emotional stimuli (Fig. S1). We suspect that high frequency activities related to MUA/spiking are very local and have very small amplitude, so they are not picked up by the bipolar LFPs measured from contacts with both the contact area for each contact and the between-contact space quite large comparative to the size of the habenula.

A

B

Figure S1. (A) Power spectral density of habenula LFPs across all time period when emotional stimuli were presented. The bold blue line and shadowed region indicates the mean ± SEM across all recorded hemispheres and the thin grey lines show measurements from individual hemispheres. (B) Time-frequency representations of the power response relative to pre-stimulus baseline for different conditions showing habenula gamma and high frequency activity are not modulated by emotional

References:

Savitz JB, Bonne O, Nugent AC, Vythilingam M, Bogers W, Charney DS, et al. Habenula volume in post-traumatic stress disorder measured with high-resolution MRI. Biology of Mood & Anxiety Disorders 2011a; 1(1): 7.

Savitz JB, Nugent AC, Bogers W, Roiser JP, Bain EE, Neumeister A, et al. Habenula volume in bipolar disorder and major depressive disorder: a high-resolution magnetic resonance imaging study. Biological Psychiatry 2011b; 69(4): 336-43.

Lawson RP, Drevets WC, Roiser JP. Defining the habenula in human neuroimaging studies. NeuroImage 2013; 64: 722-7.

Marmor O, Valsky D, Joshua M, Bick AS, Arkadir D, Tamir I, et al. Local vs. volume conductance activity of field potentials in the human subthalamic nucleus. Journal of Neurophysiology 2017; 117(6): 2140-51.

Bertone-Cueto NI, Makarova J, Mosqueira A, García-Violini D, Sánchez-Peña R, Herreras O, et al. Volume-Conducted Origin of the Field Potential at the Lateral Habenula. Frontiers in Systems Neuroscience 2019; 13:78.

Figure 3: the alpha/theta band activity is very transient and not band-limited. Why refer to this as oscillatory? Can you exclude that the TFRs of power reflect the spectral power of ERPs rather than modulations of oscillations? I propose to also calculate the ERPs and perform the TFR of power on those. This might result in a re-interpretation of the early effects in theta/alpha band.

We agree with the reviewer that the activity increase in the first time window with short latency after the stimuli onset is very transient and not band-limited. This raise the question that whether this is oscillatory or a transient evoked activity. We have now looked at this initial transient activity in different ways: 1.) We quantified the ERP in LFPs locked to the stimuli onset for each emotional valence condition and for each habenula. We investigated whether there was difference in the amplitude or latency of the ERP for different stimuli emotional valence conditions. As showing in the following figure, there is ERP with stimuli onset with a positive peak at 402 ± 27 ms (neutral stimuli), 407 ± 35 ms (positive stimuli), 399 ± 30 ms (negative stimuli). The flowing figure (Fig. 3–figure supplement 1) will be submitted as figure supplement related to Fig. 3. However, there was no significant difference in ERP latency or amplitude caused by different emotional valence stimuli. 2.) We have quantified the pure non-phase-locked (induced only) power spectra by calculating the time-frequency power spectrogram after subtracting the ERP (the time-domain trial average) from time-domain neural signal on each trial (Kalcher and Pfurtscheller, 1995; Cohen and Donner, 2013). This shows very similar results as we reported in the main manuscript, as shown in Fig. 3–figure supplement 2. These further analyses show that even though there were event related potential changes time locked around the stimuli onset, and this ERP did NOT contribute to the initial broad-band activity increase at the early time window shown in plot A-C in Figure 3. The figures of the new analyses and following have now been added in the main text:

"In addition, we tested whether stimuli-related habenula LFP modulations primarily reflect a modulation of oscillations, which is not phase-locked to stimulus onset, or, alternatively, if they are attributed to evoked event-related potential (ERP). We quantified the ERP for each emotional valence condition for each habenula. There was no significant difference in ERP latency or amplitude caused by different emotional valence stimuli (Fig. 3–figure supplement 1). In addition, when only considering the non phase-locked activity by removing the ERP from the time series before frequency-time decomposition, the emotional valence effect (presented in Fig. 3–figure supplement 2) is very similar to those shown in Fig.3. These additional analyses demonstrated that the emotional valence effect in the LFP signal is more likely to be driven by non-phase-locked (induced only) activity."

A

B

Fig. 3–figure supplement 1. Event-related potential (ERP) in habenula LFP signals in different emotional valence (neutral, positive and negative) conditions. (A) Averaged ERP waveforms across patients for different conditions. (B) Peak latency and amplitude (Mean ± SEM) of the ERP components for different conditions.

Fig. 3–figure supplement 2. Non-phase-locked activity in different emotional valence (neutral, positive and negative) conditions (N = 18). (A) Time-frequency representation of the power changes relative to pre-stimulus baseline for three conditions. Significant clusters (p < 0.05, non-parametric permutation test) are encircled with a solid black line. (B) Time-frequency representation of the power response difference between negative and positive valence stimuli, showing significant increased activity the theta/alpha band (5-10 Hz) at short latency (100-500 ms) and another increased theta activity (4-7 Hz) at long latencies (2700-3300 ms) with negative stimuli (p < 0.05, non-parametric permutation test). (C) Normalized power of the activities at theta/alpha (5-10 Hz) and theta (4-7 Hz) band over time. Significant difference between the negative and positive valence stimuli is marked by a shadowed bar (p < 0.05, corrected for multiple comparison).

References:

Kalcher J, Pfurtscheller G. Discrimination between phase-locked and non-phase-locked event-related EEG activity. Electroencephalography and Clinical Neurophysiology 1995; 94(5): 381-4.

Cohen MX, Donner TH. Midfrontal conflict-related theta-band power reflects neural oscillations that predict behavior. Journal of Neurophysiology 2013; 110(12): 2752-63.

Figure 4D: can you exclude that the frontal activity is not due to saccade artifacts? Only eye blink artifacts were reduced by the ICA approach. Trials with saccades should be identified in the MEG traces and rejected prior to further analysis.

We understand and appreciate the reviewer’s concern on the source of the activity modulations shown in Fig. 4D. We tried to minimise the eye movement or saccade in the recording by presenting all figures at the centre of the screen, scaling all presented figures to similar size, and presenting a white cross at the centre of the screen preparing the participants for the onset of the stimuli. Despite this, participants my still make eye movements and saccade in the recording. We used ICA to exclude the low frequency large amplitude artefacts which can be related to either eye blink or other large eye movements. However, this may not be able to exclude artefacts related to miniature saccades. As shown in Fig. 4D, on the sensor level, the sensors with significant difference between the negative vs. positive emotional valence condition clustered around frontal cortex, close to the eye area. However, we think this is not dominated by saccades because of the following two reasons:

1.) The power spectrum of the saccadic spike artifact in MEG is characterized by a broadband peak in the gamma band from roughly 30 to 120 Hz (Yuval-Greenberg et al., 2008; Keren et al., 2010). In this study the activity modulation we observed in the frontal sensors are limited to the theta/alpha frequency band, so it is different from the power spectra of the saccadic spike artefact.

2.) The source of the saccadic spike artefacts in MEG measurement tend to be localized to the region of the extraocular muscles of both eyes (Carl et al., 2012).We used beamforming source localisation to identify the source of the activity modulation reported in Fig. 4D. This beamforming analysis identified the source to be in the Broadmann area 9 and 10 (shown in Fig. 5). This excludes the possibility that the activity modulation in the sensor level reported in Fig. 4D is due to saccades. In addition, Broadman area 9 and 10, have previously been associated with emotional stimulus processing (Bermpohl et al., 2006), Broadman area 9 in the left hemisphere has also been used as the target for repetitive transcranial magnetic stimulation (rTMS) as a treatment for drug-resistant depression (Cash et al., 2020). The source localisation results, together with previous literature on the function of the identified source area suggest that the activity modulation we observed in the frontal cortex is very likely to be related to emotional stimuli processing.

References:

Yuval-Greenberg S, Tomer O, Keren AS, Nelken I, Deouell LY. Transient induced gamma-band response in EEG as a manifestation of miniature saccades. Neuron 2008; 58(3): 429-41.

Keren AS, Yuval-Greenberg S, Deouell LY. Saccadic spike potentials in gamma-band EEG: characterization, detection and suppression. NeuroImage 2010; 49(3): 2248-63.

Carl C, Acik A, Konig P, Engel AK, Hipp JF. The saccadic spike artifact in MEG. NeuroImage 2012; 59(2): 1657-67.

Bermpohl F, Pascual-Leone A, Amedi A, Merabet LB, Fregni F, Gaab N, et al. Attentional modulation of emotional stimulus processing: an fMRI study using emotional expectancy. Human Brain Mapping 2006; 27(8): 662-77.

Cash RFH, Weigand A, Zalesky A, Siddiqi SH, Downar J, Fitzgerald PB, et al. Using Brain Imaging to Improve Spatial Targeting of Transcranial Magnetic Stimulation for Depression. Biological Psychiatry 2020.

The coherence modulations in Fig 5 occur quite late in time compared to the power modulations in Fig 3 and 4. When discussing the results (in e.g. the abstract) it reads as if these findings are reflecting the same process. How can the two effect reflect the same process if the timing is so different?

As the reviewer pointed out correctly, the time window where we observed the coherence modulations happened quite late in time compared to the initial power modulations in the frontal cortex and the habenula (Fig. 4). And there was another increase in the theta band activities in the habenula area even later, at around 3 second after stimuli onset when the emotional figure has already disappeared. Emotional response is composed of a number of factors, two of which are the initial reactivity to an emotional stimulus and the subsequent recovery once the stimulus terminates or ceases to be relevant (Schuyler et al., 2014). We think these neural effects we observed in the three different time windows may reflect different underlying processes. We have discussed this in the ‘Discussion’:

"These activity changes at different time windows may reflect the different neuropsychological processes underlying emotion perception including identification and appraisal of emotional material, production of affective states, and autonomic response regulation and recovery (Phillips et al., 2003a). The later effects of increased theta activities in the habenula when the stimuli disappeared were also supported by other literature showing that, there can be prolonged effects of negative stimuli in the neural structure involved in emotional processing (Haas et al., 2008; Puccetti et al., 2021). In particular, greater sustained patterns of brain activity in the medial prefrontal cortex when responding to blocks of negative facial expressions was associated with higher scores of neuroticism across participants (Haas et al., 2008). Slower amygdala recovery from negative images also predicts greater trait neuroticism, lower levels of likability of a set of social stimuli (neutral faces), and declined day-to-day psychological wellbeing (Schuyler et al., 2014; Puccetti et al., 2021)."

References:

Schuyler BS, Kral TR, Jacquart J, Burghy CA, Weng HY, Perlman DM, et al. Temporal dynamics of emotional responding: amygdala recovery predicts emotional traits. Social Cognitive and Affective Neuroscience 2014; 9(2): 176-81.

Phillips ML, Drevets WC, Rauch SL, Lane R. Neurobiology of emotion perception I: The neural basis of normal emotion perception. Biological Psychiatry 2003a; 54(5): 504-14.

Haas BW, Constable RT, Canli T. Stop the sadness: Neuroticism is associated with sustained medial prefrontal cortex response to emotional facial expressions. NeuroImage 2008; 42(1): 385-92.

Puccetti NA, Schaefer SM, van Reekum CM, Ong AD, Almeida DM, Ryff CD, et al. Linking Amygdala Persistence to Real-World Emotional Experience and Psychological Well-Being. Journal of Neuroscience 2021: JN-RM-1637-20.

Be explicit on the degrees of freedom in the statistical tests given that one subject was excluded from some of the tests.

We thank the reviewers for the comment. The number of samples used for each statistics analysis are stated in the title of the figures. We have now also added the degree of freedom in the main text when parametric statistical tests such as t-test or ANOVAs have been used. When permutation tests (which do not have any degrees of freedom associated with it) are used, we have now added the number of samples for the permutation test.

Reviewer #2 (Public Review):

In this study, Huang and colleagues recorded local field potentials from the lateral habenula in patients with psychiatric disorders who recently underwent surgery for deep brain stimulation (DBS). The authors combined these invasive measurements with non-invasive whole-head MEG recordings to study functional connectivity between the habenula and cortical areas. Since the lateral habenula is believed to be involved in the processing of emotions, and negative emotions in particular, the authors investigated whether brain activity in this region is related to emotional valence. They presented pictures inducing negative and positive emotions to the patients and found that theta and alpha activity in the habenula and frontal cortex increases when patients experience negative emotions. Functional connectivity between the habenula and the cortex was likewise increased in this band. The authors conclude that theta/alpha oscillations in the habenula-cortex network are involved in the processing of negative emotions in humans.

Because DBS of the habenula is a new treatment tested in this cohort in the framework of a clinical trial, these are the first data of its kind. Accordingly, they are of high interest to the field. Although the study mostly confirms findings from animal studies rather than bringing up completely new aspects of emotion processing, it certainly closes a knowledge gap.

In terms of community impact, I see the strengths of this paper in basic science rather than the clinical field. The authors demonstrate the involvement of theta oscillations in the habenula-prefrontal cortex network in emotion processing in the human brain. The potential of theta oscillations to serve as a marker in closed-loop DBS, as put forward by the authors, appears less relevant to me at this stage, given that the clinical effects and side-effects of habenula DBS are not known yet.

We thank the reviewers for the favourable comments about the implication of our study in basic science and about the value of our study in closing a knowledge gap. We agree that further studies would be required to make conclusions about the clinical effects and side-effects of habenula DBS.

Detailed comments:

The group-average MEG power spectrum (Fig. 4B) suggests that negative emotions lead to a sustained theta power increase and a similar effect, though possibly masked by a visual ERP, can be seen in the habenula (Fig. 3C). Yet the statistics identify brief elevations of habenula theta power at around 3s (which is very late), a brief elevation of prefrontal power a time 0 or even before (Fig. 4C) and a brief elevation of Habenula-MEG theta coherence around 1 s. It seems possible that this lack of consistency arises from a low signal-to-noise ratio. The data contain only 27 trails per condition on average and are contaminated by artifacts caused by the extension wires.

With regard to the nature of the activity modulation with short latency after stimuli onset: whether this is an ERP or oscillation? We have now investigated this. In summary, by analysing the ERP and removing the influence of the ERP from the total power spectra, we didn’t observe stimulus emotional valence related modulation in the ERP, and the modulation related to emotional valence in the pure induced (non-phase-locked) power spectra was similar to what we have observed in the total power shown in Fig. 3. Therefore, we argue that the theta/alpha increase with negative emotional stimuli we observed in both habenula and prefrontal cortex 0-500 ms after stimuli onset are not dominated by visual or other ERP.

With regard to the signal-to-noise ratio from only 27 trials per condition on average per participant: We have tried to clean the data by removing the trials with obvious artefacts characterised by increased measurements in the time domain over 5 times the standard deviation and increased activities across all frequency bands in the frequency domain. After removing the trials with artefacts, we have 27 trials per condition per subject on average. We agree that 27 trials per condition on average is not a high number, and increasing the number of trials would further increase the signal-to-noise ratio. However, our studies with EEG recordings and LFP recordings from externalised patients have shown that 30 trials was enough to identify reduction in the amplitude of post-movement beta oscillations at the beginning of visuomotor adaption in the motor cortex and STN (Tan et al., 2014a; Tan et al., 2014b). These results of motor error related modulation in the post-movement beta have been repeated by other studies from other groups. In Tan et al. 2014b, with simultaneous EEG and STN LFP measurements and a similar number of trials (around 30), we also quantified the time-course of STN-motor cortex coherence during voluntary movements. This pattern has also been repeated in a separate study from another group with around 50 trials per participant (Talakoub et al., 2016). In addition, similar behavioural paradigm (passive figure viewing paradigm) has been used in two previous studies with LFP recordings from STN from different patient groups (Brucke et al., 2007; Huebl et al., 2014). In both studies, a similar number of trials per condition around 27 was used. The authors have identified meaningful activity modulation in the STN by emotional stimuli. Therefore, we think the number of trials per condition was sufficient to identify emotional valence induced difference in the LFPs in the paradigm.

We agree that the measurement of coherence can be more susceptible to noise and suffer from the reduced signal-to-noise ratio in MEG recording. In Hirschmann et al. 2013, 5 minutes of resting recording and 5 minutes of movement recording from 10 PD patients were used to quantify movement related changes in STN-cortical coherence and how this was modulated by levodopa (Hirschmann et al., 2013). Litvak et al. (2012) have identified movement-related changes in the coherence between STN LFP and motor cortex with recording with simultaneous STN LFP and MEG recordings from 17 PD patients and 20 trials in average per participant per condition (Litvak et al., 2012). With similar methods, van Wijk et al. (2017) used recordings from 9 patients and around on average in 29 trials per hand per condition, and they identified reduced cortico-pallidal coherence in the low-beta decreases during movement (van Wijk et al., 2017). So the trial number per condition participant we used in this study are comparable to previous studies.

The DBS extension wires do reduce signal-to-noise ratio in the MEG recording. therefore the spatiotemporal Signal Space Separation (tSSS) method (Taulu and Simola, 2006) implemented in the MaxFilter software (Elekta Oy, Helsinki, Finland) has been applied in this study to suppress strong magnetic artifacts caused by extension wires. This method has been proved to work well in de-noising the magnetic artifacts and movement artifacts in MEG data in our previous studies (Cao et al., 2019; Cao et al., 2020). In addition, the beamforming method proposed by several studies (Litvak et al., 2010; Hirschmann et al., 2011; Litvak et al., 2011) has been used in this study. In Litvak et al., 2010, the artifacts caused by DBS extension wires was detailed described and the beamforming was demonstrated to effectively suppress artifacts and thereby enable both localization of cortical sources coherent with the deep brain nucleus. We have now added more details and these references about the data cleaning and the beamforming method in the main text. With the beamforming method, we did observe the standard movement-related modulation in the beta frequency band in the motor cortex with 9 trials of figure pressing movements, shown in the following figure for one patient as an example (Figure 5–figure supplement 1). This suggests that the beamforming method did work well to suppress the artefacts and help to localise the source with a low number of trials. The figure on movement-related modulation in the motor cortex in the MEG signals have now been added as a supplementary figure to demonstrate the effect of the beamforming.

Figure 5–figure supplement 1. (A) Time-frequency maps of MEG activity for right hand button press at sensor level from one participant (Case 8). (B) DICS beamforming source reconstruction of the areas with movement-related oscillation changes in the range of 12-30 Hz. The peak power was located in the left M1 area, MNI coordinate [-37, -12, 43].

References:

Tan H, Jenkinson N, Brown P. Dynamic neural correlates of motor error monitoring and adaptation during trial-to-trial learning. Journal of Neuroscience 2014a; 34(16): 5678-88.

Tan H, Zavala B, Pogosyan A, Ashkan K, Zrinzo L, Foltynie T, et al. Human subthalamic nucleus in movement error detection and its evaluation during visuomotor adaptation. Journal of Neuroscience 2014b; 34(50): 16744-54.

Talakoub O, Neagu B, Udupa K, Tsang E, Chen R, Popovic MR, et al. Time-course of coherence in the human basal ganglia during voluntary movements. Scientific Reports 2016; 6: 34930.

Brucke C, Kupsch A, Schneider GH, Hariz MI, Nuttin B, Kopp U, et al. The subthalamic region is activated during valence-related emotional processing in patients with Parkinson's disease. European Journal of Neuroscience 2007; 26(3): 767-74.

Huebl J, Spitzer B, Brucke C, Schonecker T, Kupsch A, Alesch F, et al. Oscillatory subthalamic nucleus activity is modulated by dopamine during emotional processing in Parkinson's disease. Cortex 2014; 60: 69-81.

Hirschmann J, Ozkurt TE, Butz M, Homburger M, Elben S, Hartmann CJ, et al. Differential modulation of STN-cortical and cortico-muscular coherence by movement and levodopa in Parkinson's disease. NeuroImage 2013; 68: 203-13.

Litvak V, Eusebio A, Jha A, Oostenveld R, Barnes G, Foltynie T, et al. Movement-related changes in local and long-range synchronization in Parkinson's disease revealed by simultaneous magnetoencephalography and intracranial recordings. Journal of Neuroscience 2012; 32(31): 10541-53.

van Wijk BCM, Neumann WJ, Schneider GH, Sander TH, Litvak V, Kuhn AA. Low-beta cortico-pallidal coherence decreases during movement and correlates with overall reaction time. NeuroImage 2017; 159: 1-8.

Taulu S, Simola J. Spatiotemporal signal space separation method for rejecting nearby interference in MEG measurements. Physics in Medicine and Biology 2006; 51(7): 1759-68.

Cao C, Huang P, Wang T, Zhan S, Liu W, Pan Y, et al. Cortico-subthalamic Coherence in a Patient With Dystonia Induced by Chorea-Acanthocytosis: A Case Report. Frontiers in Human Neuroscience 2019; 13: 163.

Cao C, Li D, Zhan S, Zhang C, Sun B, Litvak V. L-dopa treatment increases oscillatory power in the motor cortex of Parkinson's disease patients. NeuroImage Clinical 2020; 26: 102255.

Litvak V, Eusebio A, Jha A, Oostenveld R, Barnes GR, Penny WD, et al. Optimized beamforming for simultaneous MEG and intracranial local field potential recordings in deep brain stimulation patients. NeuroImage 2010; 50(4): 1578-88.

Litvak V, Jha A, Eusebio A, Oostenveld R, Foltynie T, Limousin P, et al. Resting oscillatory cortico-subthalamic connectivity in patients with Parkinson's disease. Brain 2011; 134(Pt 2): 359-74.

Hirschmann J, Ozkurt TE, Butz M, Homburger M, Elben S, Hartmann CJ, et al. Distinct oscillatory STN-cortical loops revealed by simultaneous MEG and local field potential recordings in patients with Parkinson's disease. NeuroImage 2011; 55(3): 1159-68.

I doubt that the correlation between habenula power and habenula-MEG coherence (Fig. 6C) is informative of emotion processing. First, power and coherence in close-by time windows are likely to to be correlated irrespective of the task/stimuli. Second, if meaningful, one would expect the strongest correlation for the negative condition, as this is the only condition with an increase of theta coherence and a subsequent increase of theta power in the habenula. This, however, does not appear to be the case.

The authors included the factors valence and arousal in their linear model and found that only valence correlated with electrophysiological effects. I suspect that arousal and valence scores are highly correlated. When fed with informative yet highly correlated variables, the significance of individual input variables becomes difficult to assess in many statistical models. Hence, I am not convinced that valence matters but arousal not.

For the correlation shown in Fig. 6C, we used a linear mixed-effect modelling (‘fitlme’ in Matlab) with different recorded subjects as random effects to investigate the correlations between the habenula power and habenula-MEG coherence at an earlier window, while considering all trials together. Therefore the reported value in the main text and in the figure (k = 0.2434 ± 0.1031, p = 0.0226, R2 = 0.104) show the within subjects correlation that are consistent across all measured subjects. The correlation is likely to be mediated by emotional valence condition, as negative emotional stimuli tend to be associated with both high habenula-MEG coherence and high theta power in the later time window tend to happen in the trials with.

The arousal scores are significantly different for the three valence conditions as shown in Fig. 1B. However, the arousal scores and the valence scores are not monotonically correlated, as shown in the following figure (Fig. S2). The emotional neutral figures have the lowest arousal value, but have the valence value sitting between the negative figures and the positive figures. We have now added the following sentence in the main text:

"This nonlinear and non-monotonic relationship between arousal scores and the emotional valence scores allowed us to differentiate the effect of the valence from arousal."

Table 2 in the main text show the results of the linear mixed-effect modelling with the neural signal as the dependent variable and the valence and arousal scores as independent variables. Because of the non-linear and non-monotonic relationship between the valence and arousal scores, we think the significance of individual input variables is valid in this statistical model. We have now added a new figure (shown below, Fig. 7) with scatter plots showing the relationship between the electrophysiological signal and the arousal and emotional valence scores separately using Spearman’s partial correlation analysis. In each scatter plot, each dot indicates the average measurement from one participant in one emotional valence condition. As shown in the following figure, the electrophysiological measurements linearly correlated with the valence score, but not with the arousal scores. However, the statistics reported in this figure considered all the dots together. The linear mixed effect modelling taking into account the interdependency of the measurements from the same participant. So the results reported in the main text using linear mixed effect modelling are statistically more valid, but supplementary figure here below illustrate the relationship.

Figure S2. Averaged valence and arousal ratings (mean ± SD) for figures of the three emotional condition. (B) Scatter plots showing the relationship between arousal and valence scores for each emotional condition for each participant.

Figure 7. Scatter plots showing how early theta/alpha band power increase in the frontal cortex (A), theta/alpha band frontal cortex-habenula coherence (B) and theta band power increase in habenula stimuli (C) changed with emotional valence (left column) and arousal (right column). Each dot shows the average of one participant in each categorical valence condition, which are also the source data of the multilevel modelling results presented in Table 2. The R and p value in the figure are the results of partial correlation considering all data points together.

Page 8: "The time-varying coherence was calculated for each trial". This is confusing because coherence quantifies the stability of a phase difference over time, i.e. it is a temporal average, not defined for individual trials. It has also been used to describe the phase difference stability over trials rather than time, and I assume this is the method applied here. Typically, the greatest coherence values coincide with event-related power increases, which is why I am surprised to see maximum coherence at 1s rather than immediately post-stimulus.

We thank the reviewer for pointing out this incorrect description. As the reviewer pointed out correctly, the method we used describe the phase difference stability over trials rather than time. We have now clarified how coherence was calculated and added more details in the methods:

"The time-varying cross trial coherence between each MEG sensor and the habenula LFP was first calculated for each emotional valence condition. For this, time-frequency auto- and cross-spectral densities in the theta/alpha frequency band (5-10 Hz) between the habenula LFP and each MEG channel at sensor level were calculated using the wavelet transform-based approach from -2000 to 4000 ms for each trial with 1 Hz steps using the Morlet wavelet and cycle number of 6. Cross-trial coherence spectra for each LFP-MEG channel combination was calculated for each emotional valence condition for each habenula using the function ‘ft_connectivityanalysis’ in Fieldtrip (version 20170628). Stimulus-related changes in coherence were assessed by expressing the time-resolved coherence spectra as a percentage change compared to the average value in the -2000 to -200 ms (pre-stimulus) time window for each frequency."

In the Morlet wavelet analysis we used here, the cycle number (C) determines the temporal resolution and frequency resolution for each frequency (F). The spectral bandwidth at a given frequency F is equal to 2F/C while the wavelet duration is equal to C/F/pi. We used a cycle number of 6. For theta band activities around 5 Hz, we will have the spectral bandwidth of 25/6 = 1.7 Hz and the wavelet duration of 6/5/pi = 0.38s = 380ms.

As the reviewer noticed, we observed increased activities across a wide frequency band in both habenula and the prefrontal cortex within 500 ms after stimuli onset. But the increase of cross-trial coherence starts at around 300 ms. The increase of coherence in a time window without increase of power in either of the two structures indicates a phase difference stability across trials in the oscillatory activities from the two regions, and this phase difference stability across trials was not secondary to power increase.

Reviewer #3 (Public Review):

This paper describes the oscillatory activity of the habenula using local field potentials, both within the region and, through the use of MEG, in connection to the prefrontal cortex. The characteristics of this activity were found to vary with the emotional valence but not with arousal. Sheding light on this is relevant, because the habenula is a promising target for deep brain stimulation.

In general, because I am not much on top of the literature on the habenula, I find difficult to judge about the novelty and the impact of this study. What I can say is that I do find the paper is well-written and very clear; and the methods, although quite basic (which is not bad), are sound and rigourous.

We thank the reviewer for the positive comments about the potential implication of our study and on the methods we used.

On the less positive side, even though I am aware that in this type of studies it is difficult to have high N, the very low N in this case makes me worry about the robustness and replicability of the results. I'm sure I have missed it and it's specified somewhere, but why is N different for the different figures? Is it because only 8 people had MEG? The number of trials seems also a somewhat low. Therefore, I feel the authors perhaps need to make an effort to make up for the short number of subjects in order to add confidence to the results. I would strongly recommend to bootstrap the statistical analysis and extract non-parametric confidence intervals instead of showing parametric standard errors whenever is appropriate. When doing that, it must be taken into account that each two of the habenula belong to the same person; i.e. one bootstraps the subjects not the habenula.

We do understand and appreciate the concern of the reviewer on the low sample numbers due to the strict recruitment criteria for this very early stage clinical trial: 9 patients for bilateral habenula LFPs, and 8 patients with good quality MEGs. Some information to justify the number of trials per condition for each participant has been provided in the reply to the Detailed Comments 1 from Reviewer 2. The sample number used in each analysis was included in the figures and in the main text.

We have used non-parametric cluster-based permutation approach (Maris and Oostenveld, 2007) for all the main results as shown in Fig. 3-5. Once the clusters (time window and frequency band) with significant differences for different emotional valence conditions have been identified, parametric statistical test was applied to the average values of the clusters to show the direction of the difference. These parametric statistics are secondary to the main non-parametric permutation test.

In addition, the DICS beamforming method was applied to localize cortical sources exhibiting stimuli-related power changes and cortical sources coherent with deep brain LFPs for each subject for positive and negative emotional valence conditions respectively. After source analysis, source statistics over subjects was performed. Non-parametric permutation testing with or without cluster-based correction for multiple comparisons was applied to statistically quantify the differences in cortical power source or coherence source between negative and positive emotional stimuli.

References:

Maris E, Oostenveld R. Nonparametric statistical testing of EEG- and MEG-data. Journal of Neuroscience Methods 2007; 164(1): 177-90.

Related to this point, the results in Figure 6 seem quite noisy, because interactions (i.e. coherence) are harder to estimate and N is low. For example, I have to make an effort of optimism to believe that Fig 6A is not just noise, and the result in Fig 6C is also a bit weak and perhaps driven by the blue point at the bottom. My read is that the authors didn't do permutation testing here, and just a parametric linear-mixed effect testing. I believe the authors should embed this into permutation testing to make sure that the extremes are not driving the current p-value.

We have now quantified the coherence between frontal cortex-habenula and occipital cortex-habenula separately (please see more details in the reply to Reviewer 2 (Recommendations for the authors 6). The new analysis showed that the increase in the theta/alpha band coherence around 1 s after the negative stimuli was only observed between prefrontal cortex-habenula and not between occipital cortex-habenula. This supports the argument that Fig. 6A is not just noise.

① Fırça başı, kök ucuna doğru yönelen eğik bir açıyla konumlandırılır; kılların bir kısmı diş etine bir kısmı diş yüzeyine temas eder ve hafif titreşimli bir hareket uygulandıktan sonra fırça başı giderek oklüzal veya insizal yöne doğru döndürülür.

Yumuşak bir fırça başını oklüzal düzleme paralel yerleştirin, fırça başı üç ila dört dişi kapsayacak şekilde konumlandırın ve çenede en distal dişten başlayın.

very 1984

Author Response

Reviewer #1 (Public Review):

1) It would be helpful to include some sort of comparison in Fig. 4, e.g. the regressions shown in Fig 3, to indicate to what extent the ICCl data corresponds to the "control range" of frequency tuning.

Figure 4 was modified to show the frequency range typically found in the ICCls. This range is based on results from Wagner et al., 2007, which extensively surveyed ICCls responses. This modification shows that our ICCls recordings in the ruff-removed owls cover the normal frequency hearing range of the owl.

2) A central hypothesis of the study is that the frequency preference of the high-frequency neurons is lower in ruff-removed owls because of the lowered reliability caused by a lack of the ruff. Yet, while lower, the frequency range of many neurons in juvenile and ruff-removed owls seems sufficiently high to be still responsive at 7-8 kHz. I think it would be important to know to what extent neurons are still ITD sensitive at the "unreliable high frequencies" even if the CFs are lower since the "optimization" according to reliability depends not on the best frequency of each neuron per se, but whether neurons are less ITD sensitive at the higher, less reliable frequencies.

The concern regarding the frequency range that elicits responsivity was largely addressed above. Specifically, Figure L1 showing frequency tuning of frontally tuned ICx neurons in ruff-removed owls indicates that while there is some variability of tuning across neurons, there is little responsivity above 6 kHz. In contrast, equivalent analysis in juvenile owls (Figure L3), shows there is much more responsiveness and variability across neurons to high and low frequencies. This evidence supports our hypothesis that the juvenile owl brain is still highly plastic, which facilitates learning during development. Although the underlying data was already reported in Figure 7 of our previously submitted manuscript, we can include Figures L1 and L2, potentially as supplemental figures, if considered useful by editors and reviewers. Nevertheless, this argumentation was further expanded in the revised text (Line 229).

Figure L1. Frequency tuning of frontally-tuned ICx neurons in ruff-removed owls. Tuning curves are normalized by the max response. Thick black line indicates the average tuning curve. Dashed black line indicates basal response.

Figure L2. ITD sensitivity across frequencies in ruff-removed owl. Two example neurons shown in a and b. ITD tuning for tones (colored) and broadband (black) plotted by firing rate (non-normalized). Solid colored lines indicate responses to frequencies that are within the neuron’s preferred frequency range (i.e. above the half-height, see Methods), dashed lines indicate frequencies outside of the neuron’s frequency range.

Figure L3. Frequency tuning of frontally-tuned ICx neurons in juvenile owls. Tuning curves are normalized by the max response. Thick black line indicates the average tuning curve. Dashed black line indicates basal response.

3) It would be interesting to have an estimate of the time scale of experience dependency that induces tuning changes. Do the authors have any data on this question? I appreciate the authors' notion that the quantifications in Fig 7 might indicate that juvenile owls are already "beginning to be shaped by ITD reliability" (line 323 in Discussion). How many days after hearing onset would this correspond to? Does this mean that a few days will already induce changes?

While tracking changes induced by ruff-removal over development were outside of the scope of this study, many other studies have assessed experience-dependent plasticity in the barn owl. The recordings in this study were performed approximately 20 days after hearing onset, suggesting that the juveniles had ample time to begin learning. These points were expanded upon in the discussion (Lines 254, 280-283).

Reviewer #2 (Public Review):

1) Why is IPD variability plotted instead of ITD variability (or indeed spatial reliability)? The relationship between these measures is likely to vary across frequency, which makes it difficult to compare ITD variability across frequency when IPDs are plotted. Normalizing data across frequencies also makes it difficult to compare different locations and acoustical conditions. For example, in Fig.1a and Fig.1b, the data shown for 3 kHz at ~160 degrees seems quantitatively and visually quite different, but the difference (in Fig.1c) appears to be negligible.

Justification of why IPD variability is used as an estimate of ITD variability was added to introduction (Lines 55-60), results (Line 100) and methods (Lines 371-374) sections of the manuscript, explaining the fact that because ITD detection is based on phase locking by auditory nerve and ITD detector neurons tuned to narrow frequency bands, responses of ITD detector neurons forwarded to downstream midbrain regions are therefore determined by IPD variability. Additionally, ITD is calculated by dividing IPD by frequency, which makes comparisons of ITD reliability across frequency mathematically uninformative.

2) How well do the measures of ITD reliability used reflect real-world listening? For example, the model used to calculate ITD reliability appears to assume the same (flat) spectral profile for targets and distractors, which are presented simultaneously with the same temporal envelope, and a uniform spatial distribution of sounds across space. It is therefore unclear how robust the study's results are to violations of these assumptions.

While we agree that our analysis cannot completely capture real-world listening for the barn owl, a general analysis using similar flat spectral profiles for targets and concurrent sounds provides a broad assessment of reliability of ITD cues. While a full recapitulation of real-world listening is beyond the scope of this study (i.e. recording natural scenes from the ear canals of wild barn owls), we included additional analyses of ITD reliability in Figure 1-figure supplement 1, described above.

3) Does facial ruff removal produce an isolated effect on ITD variability or does it also produce changes in directional gain, and the relationship between spatial cues and sound location? Although the study considers this issue in some places (e.g. Fig.2, Fig.5), a clearer presentation of the acoustical effects of facial ruff removal and their implications (for all locations, not just those to the front), as well as an attempt to understand how these acoustical changes lead to the observed changes in ITD reliability, would greatly strengthen the study. In addition, Fig.1 shows average ITD reliability across owls, but it would be helpful to know how consistent these measures are across owls, given individual variability in Head-Related Transfer Functions (HRTFs). This potentially has implications for the electrophysiological experiments, if the HRTFs of those animals were not measured. One specific question that is potentially very relevant is whether the facial ruff attenuates sounds presented behind the animal and whether it does so in a frequency-dependent way. In addition, if facial ruff removal enables ILDs to be used for azimuth, then ITDs may also become less necessary at higher frequencies, even if their reliability remains unchanged.

Additional analysis was conducted to generate representation of changes in directional gain induced by ruff removal, added to new figure (Fig 5). This analysis shows that changes in gain following ruff-removal are largely frequency-independent: there is a de-attenuation of peripherally and rearwardly located sounds, but the highest gain remains for high frequencies in frontal space. There is an additional increase in gain for high frequencies from rearward space, these changes would not explain the changes in frequency tuning we report. As mentioned in new additions to the manuscript, the changes at the most rearward-located auditory spatial locations are unlikely to have an effect on the auditory midbrain. No studies in the barn owl have found neurons in the ICx or optic tectum tuned to >120° (Knudsen, 1982; Knudsen, 1984; Cazettes et al., 2014). In addition, variability of IPD reliability across owls was analyzed and reported in the amended Figure 1, which notes very little changes across owls. In this analysis, we did realize that the file of one of the HRTFs obtained from von Campenhausen et al. 2006 was mislabeled, which explains slight differences in revised Fig 1b. Nevertheless, added analysis of IPD reliability across owls indicates that the pattern in ITD reliability is stable across owls (Fig. 1d,e), which supports our decision to not record HRTFs from owls used in this study. Finally, we added to the discussion that clarifies that the use of ILD for azimuth would not provide the same resolution as ITD would (Lines 295-303). We also do not believe that the use of ILD for azimuth would make “ITDs… less necessary at higher frequencies”, given that the ICCls is still computing ITD at these high frequencies (Fig 4), and that ILDs also have higher resolution at higher frequencies, with and without the facial ruff (Olsen et al, 1989; Keller et al., 1998; von Campenhausen et al., 2006).

1) It is unclear why some analyses (Fig.5, Fig.7) are focused on frontal locations and frontally-tuned neurons. It is also unclear why neurons with a best ITDs of 0 are described as frontally tuned since locations behind the animal produce an ITD of 0 also. Related to this, in Fig.1, facial ruff removal appears to reduce IPD variability at low frequencies for locations to the rear (~160 degrees), where the ITD is likely to be close to 0. Neurons with a best ITD of 0 might therefore be expected to adjust their frequency tuning in opposite directions depending on whether they are tuned to frontal or rearward locations.

An extensive explanation was added to the methods detailing why we do not believe the neurons recorded in this study are tuned to the rear. Namely, studies mapping the barn owl’s ICx and optic tectum have not reported neurons tuned to locations >120°, with the number of neurons representing a given spatial location decreasing with eccentricity (Knudsen, 1982; Knudsen, 1984; Cazettes et al., 2014). While we agree that there does seem to be a change in ITD reliability at ~160° following ruff-removal, the result is largely similar to the change that occurs in frontal space (Fig 1b), which is consistent with the ruff-removed head functioning as a sphere. Thus, we wouldn’t expect rearwardly-tuned neurons, if they could be readily found, to adjust their frequency tuning to higher frequencies. Finally, we want to clarify that we focused our analyses on frontally-tuned neurons because frontal space is where we observed the largest change in ITD reliability. Text was added to the Discussion section to clarify this point (Lines 313-321).

2) The study suggests that information about high-frequency ITDs is not passed on to the ICX if the ICX does not contain neurons that have a high best frequency. However, neurons might be sensitive to ITDs at frequencies other than the best frequency, particularly if their frequency tuning is broader. It is also unclear whether the best frequency of a neuron always corresponds to the frequency that provides the most reliable ITD information, which the study implicitly assumes.

The concern about ITD sensitivity at non-preferred frequencies was addressed under the essential revision #3, as well as under Reviewer 1’s concerns.

Author Response:

Reviewer #2 (Public Review):

1. The novelty of the current observation of two types of links is overstated, for example, in the abstract: "Our data reveal the existence of two molecular connectors/spacers which likely contribute to the nanometer scale precise stacking of the ROS disks" (Line 25). In fact, both of these links have been shown before (Usukura and Yamada, 1981; Roof and Heuser, 1982; Corless and Schneider, 1987; Corless et al., 1987; Kajimura et al., 2000). These previous studies deserve to be recognized. Of special note is the paper by Usukura and Yamada whose images of the disc rim connectors are by no means less convincing than shown in the current manuscript. On the other hand, the novelty and impact of the data related to peripherin appears to be understated, particularly in the abstract.

We changed the abstract line 27 to: “Our data confirm the existence of two previously observed molecular connectors …”, cite the recommended references in the introduction (lines 54-55), the results (lines 131-132), and the discussion (lines 282/285). To highlight the previous reports, we rephrased the sentence in lines 132-133, “In agreement with these previous findings, we observed structures that connect membranes of two adjacent disks …”; the discussion is rephrased in lines 280-281, “Similar connectors have been observed previously ...” and “… and their statistical analysis confirmed the existence of two distinct connector species.”, and in lines 291-292, “Based on previous studies combined with our quantitative analysis, we put forward a hypothesis for the molecular identity of the disk rim connector which agrees in part with recent models”.

1. Notably, ROM-1 has not been found in peripherin oligomers larger than octamers (e.g. Loewen and Molday, 2000 and subsequent studies by Naash and colleagues). This should be discussed in the context of the current model.

We agree that this is an important aspect. We pick subvolumes along all disk rims, and on average we obtain the ordered scaffold as shown in the manuscript. We expect heterogeneity in the data because of the different degrees of oligomerization and the exclusion of ROM1 from higher oligomers. Our analysis required substantial classification to achieve convergence to a stable average, indeed indicating heterogeneity in the rim structure. However, we could not resolve additional structures to sufficient quality. It might be that this heterogeneity is what ultimately limits our achievable resolution. We added these thoughts in the discussion starting in lines 377-378, “PRPH2-Rom1 oligomers isolated from native sources exhibit varying degrees of polymerization (Loewen and Molday, 2000), and ROM1 is excluded from larger oligomers (Milstein et al., 2020). We could not resolve this heterogeneity as additional structures to sufficient quality by subvolume averaging, but in combination with the inherent flexibility of the disk rim, this heterogeneity might be the reason for the restricted resolution of our averages.”

1. The following statement should be reconsidered given the established role of cysteine-150 in peripherin oligomerization: "We hypothesize that the necessary cysteine residues are located in the head domain of the tetramers (Figure 5B), ..." It has been firmly established that only one cysteine (C150) located in the intradiscal loop is not engaged in intramolecular interactions and is essential for peripherin oligomerization.

Thank you for this advice. We agree and rephrased our discussion in lines 368-371, “The intermolecular disulfide brides are exclusively formed by the PRPH2-C150 and ROM1-C153 cysteine residues, which are located in the luminal domain (Zulliger et al., 2018). We hypothesize that these disulfide bonds (Figure 5B), are responsible for the contacts across rows (Figure 3) ...”

1. Line 340: "A model involving V-shaped tetramers for membrane curvature formation was proposed recently (Milstein et al., 2020), but it comprises two rows of tetramers which are linked in a head-tohead manner. Our analysis instead resolves three rows organized side-by side in situ (Figure 5A)." I am confused by this statement: doesn't your model also show long rows connected head-to-head? The real difference is that Milstein and colleagues proposed four tetramers per rim whereas the current data reveal three.

Thank you for pointing out this imprecise description. The model proposed by Milstein and the model in the old version of our manuscript, both propose linkage between tetramers via their disk luminal domains. In our manuscript, we refer to the luminal domain as the head domain. However, to our understanding, the Milstein model suggests two rows of tetramers, where one tetramer in the first row is rotated 180° with respect to a tetramer in the second row (therefore head-to-head), while our data indicate that the V-shaped repeats which we originally hypothesized to be tetramers are only rotated ~63° with respect to one another and are therefore rather oriented side-by-side:

Fig. 2: Comparison of models for the organization of the ROS disk rim as proposed in in Milstein et al., 2020 (top panel)

and in our work (lower panel). We now rephrased lines 383-385, “Instead, our analysis in situ resolves three rows of repeats which are also linked by the luminal domain but are rather organized side-by-side (Figure 5A).”

1. Line 347: "Our data indicate that the luminal domains of tetramers hold the disk rim scaffold together (Figure 3C), which is supported by the fact that most pathological mutations of PRPH2 affect its luminal domain (Boon et al., 2008; Goldberg et al., 2001). It is possible that these mutations impair the formation of tetramers, rows of tetramers, and their disulfide bond-stabilized oligomerization. These alterations could impede or completely prevent disk morphogenesis which, in turn, would disrupt the structural integrity of ROS, compromise the viability of the retina and ultimately lead to blindness." This is not an original idea, as many studies showed that disruptions in peripherin oligomerization lead to anatomical defects in disc formation and subsequent photoreceptor cell death.

Thank you for pointing this out. Our data are indeed in good agreement with the results made by many groups and further expand on them. We rephrased the manuscript in several places to clarify this relationship: in the abstract lines 32-34, “Our Cryo-ET data provide novel quantitative and structural information on the molecular architecture in ROS and substantiate previous results on proposed mechanisms underlying pathologies of certain PRPH2 mutations leading to blindness.”; in the introduction lines 78-79, “… allowed us to obtain 3D molecular-resolution images of vitrified ROS in a close-to-native state providing further evidence for previously suggested mechanisms leading to ROS dysfunction”; and in the discussion lines 393-397, “In good agreement with previous work, it is possible that these mutations impair the formation of complexes, and their disulfide bond-stabilized oligomerization (Chang et al., 2002; Conley et al., 2019; Zulliger et al., 2018). Hence, these alterations could impede or completely prevent disk morphogenesis …”. Also, additional relevant publications are cited in line 395.

1. In regards to the distance between disc rims and plasma membrane, the authors cite the data obtained with frogs (10 nm) but not a more relevant, previously reported measurement in mice (Gilliam et al, 2012). The value of 18 nm reported in that study is much closer to the currently reported value.

We appreciate the reference to this excellent paper. We added it in lines 335-337, “This value was derived from amphibians (Roof and Heuser, 1982) and deviates considerably from recent results (18 nm, (Gilliam et al., 2012)) and from our current measurements in mice (~25 nm).” Our aim was to point out that a model for ROS organization that is often cited and is otherwise well-founded (BatraSafferling et al., 2006) makes a wrong assumption about distance in the context of the mammalian systems. 7. The authors are (correctly) being very careful in assigning the molecular identity of disc interior connectors to PDE6. However, they are more confident in assigning the disc rim connectors to GARP2, which is reflected in the labeling of these links in figure

1. Their arguments are valid, but these links are not attached to peripherin (a protein considered to be the membrane binding partner for GARPs), which is not immediately consistent with this hypothesis. Perhaps it would be fair to re-label the corresponding links in figure 5 as "disc rim connectors".

That is an excellent and fair suggestion. We changed Figure 5 accordingly.

1. On a similar note, the disc rim connectors seem to be located where ABCA4 is presumed to be localized within the rim, which may not be just a coincidence. The authors already have tomograms obtained from ABCA4 knockout animals. Is it possible to analyze whether these links are preserved in these tomograms?

We agree, this is an important question to address. Unfortunately, neither the biological preparation nor the tomograms of the ABCA4 knockout were as good in quality as for the WT. Still, we frequently see connectors at the disk rim, especially after denoising of the tomograms.

Fig. 3: connectors at disk rims in WT (left) and ABCA4 knockout mice (right).

Sometimes it appears the connectors between adjacent disks are linked via an intradisk densities, which was already observed in Corless et al., 1987. We thought that these densities could be ABCA4 and tried to find them with two approaches in our WT tomograms (data not shown). In the first approach using a segmentation similar to what we did for the connectors between disks, we found an order of magnitude fewer intradisk connectors than (inter)disk rim connectors. In the second approach, we used the positions of segmented (inter)disk rim connectors and classified rotational averages which focused on the disk luminal space next to the contact point of a connector with the disk membrane. Again, less than 10% of the disk rim connector subvolumes were assigned to classes with an additional luminal density. Both experiments indicate that disk rim connectors sometimes occur with an additional luminal density. In total, we found less than 100 of these intradisk densities, an observation which seems to be preserved in WT and ABCA4 KO. Based on this small number of positions/locations, however, we cannot draw any conclusion. Therefore, we did not add this point to the manuscript.

Author Response

Reviewer #1 (Public Review):

This is a very interesting paper describing membrane potential dynamics of hippocampal principal cells during UP/DOWN transitions and sharp-wave ripples. Using whole-cell in combination with linear LFP recordings in head-fixed awake mice, the authors show striking differences of membrane potential responses in principal cells from the dentage gyrus, CA3 and CA1 sectors. The authors propose that switches between a dominant inhibitory excitable state and a disinhibited non-excitable state control the intra-hippocampal dynamics during UP/DOWN transitions.

Obtaining intracellular recordings in vivo is commendable. The authors provide valuable data and analysis. While data show clear trends and some of the conclusions are well supported, the authors may need to clarify the following potential confounds, which can actually impact their conclusions and interpretation:

1- All the analysis is based in z-scored membrane potential responses but the mean resting membrane potential is never reported. For DG granule cells recorded in awake conditions, the membrane potential is usually hyperpolarized so that most of the effect may be due to reversed GABAa mediated currents. Similarly, for those cells exhibiting the non-expected polarization during UP/DOWN states there may be drifts around reversal potentials explaining their behavior. Moreover, regional trends on passive and active membrane parameters and connectivity can actually explain part of the variability. A longitudinal comparison of state Vm and spikes in fig.5 suggests that some of the largest depolarized responses are not correlated with firing. Authors should evaluate this angle, ideally showing the distribution of membrane potential values across cells and regions and confronting this with the different membrane potential responses.

We added Figure 1 - figure supplement 4, which now describes the mean resting membrane potential, input resistance, burst propensity, and spikes per burst for the recorded cells. These data are provided in Figure 1 - source data 1 together with a recording identifier that can be used to link each cell to all other figure panels and data files. We further added Figure 1 - figure supplement 1, which provides examples of morphological information for our recordings, Figure 1 - figure supplement 2 that shows examples of bursts from morphologically identified neurons, and Figure 1 - figure supplement 3 that shows the locations of recorded cells.

In addition, we added Figure 5 - figure supplement 4 that includes the resting Vm and proximodistal location of cells in relation to their UP-DOWN modulation. We did not detect any significant trends with respect to brain state modulation. DG cells are more hyperpolarized compared to CA3 and CA1 cells and are closest to the reversal potential for GABAa (Figure 1 - figure supplement 4). The lack of any clear trends with respect to the resting Vm suggests that drifts around the GABAa reversal potential are unlikely to be a major factor driving variability in the observed UDS modulation.

2- While there are some trends for each hippocampal regions, there is also individual variability across cells during UP/DOWN transitions (fig.5) and near ripples (fig.6). What part of this variability can be explained by proximodistal and/or deep-superficial differences of cell location and identity? Can authors provide some morphological validation, even if in only a subset of cells? For CA3, proximodistal heterogeneity for intrinsic properties and entorhinal input responses are well documented in intracellular recordings both in vitro and in vivo. What is the location of CA3 cell contributing to this study? For CA1 cells, deep-superficial trends of GABAergic perisomatic inhibition and connectivity with input pathways dominate firing responses. Regarding DG cells, are all they from the upper blade?

We now provide morphological validation for a subset of cells (Figure 1 - figure supplement 1). Since we patch multiple cells in each experiment it is not possible to unequivocally determine their depth within the cell layer, although it is possible to confirm that they are granule cells or pyramidal cells in experiments where all labeled cells are principal neurons (Figure 1 - figure supplement 1). In addition, we added Figure 1 - figure supplement 3 that shows the proximodistal locations of recorded cells. With respect to the DG cells 20/22 are from the upper blade, with only two granule cells recorded in the lower blade (Figure 1 - figure supplement 3).

We added Figure 5 - figure supplement 4 that includes the resting Vm and proximodistal location of each cell as a function of UP-DOWN modulation. We did not detect any significant trends with respect to UDS modulation.

In addition, we added Figure 6 - figure supplement 1 that includes the resting Vm and proximodistal location of each cell as a function of ripple modulation. This figure shows that the most depolarized CA3 cells tend to hyperpolarize most during ripples, consistent with the fact that these cells are furthest away from the GABAa reversal potential and experience the highest driving force. No other significant trends were detected, although we would like to note that our recordings do not span the full proximodistal axis and may hence not be ideally suited to test the dependence of our results on proximodistal location.

3- AC-coupled LFP recordings cannot provide unambiguous identification of the sign of phasic CSD signals, because fluctuations accompanying UP/DOWN states alter the baseline reference. This is actually the case, given changes of membrane potential accompanying UP/DOWN transitions. I recommend reading Brankack et al. 1993 doi: 10.1016/0006-8993(93)90043-m. The authors should acknowledge this limitation and discuss how it could influence their results. One potential solution to get rid of this effect is using principal/independent component analysis for blind source separation.

We acknowledge the inherent limitations of AC-coupled recordings in regards to CSD analysis (Brankack et al., 1993). However, we do not believe these limitations affect our analysis or results for the reasons illustrated in Figure R1. Specifically, we do not attempt to measure the low frequency (< 1 Hz) CSD content directly. Instead, we extract the envelope of the rectified fast CSD transients. In the original submission we referred to this envelope signal as “DG CSD magnitude”, which may have been confusing. In the revised manuscript we use “DG CSD activity” instead to remove any suggestion that the low frequency CSD signal was directly measured. Notice that because of the rectification step the envelope signal is insensitive to the actual polarity of the fast transient CSD fluctuations. Using the envelope, we identify UP states as time periods when the rate and amplitude of EC input current transients, rather than the DC level, increases, in accordance with previous publications (Isomura et al., 2006). We further validated that the extracted UP/DOWN states reflect modulation of pupil diameter and ripple rate, quantities that are independently measured.

Figure R1. Deriving slow envelope signal from AC coupled recordings. (A) In this example the true CSD signal contains both a slow component (8 Hz) and a fast component (80 Hz) that is amplitude modulated by the slow component. Such phase-amplitude coupling is well known between theta and gamma oscillations in the hippocampus. The true CSD shows a current sink with time-varying magnitude. (B) The power spectral density (PSD) estimate of the signal in (A) shows both the slow (8 Hz) and fast (three peaks near 80 Hz) components. (C) Assume LFP recordings are obtained with a high-pass filter that has eliminated the slow component. Consequently, the estimated CSD signal contains only fast fluctuations. Furthermore, instead of a time-varying current sink it shows quickly alternating sinks and sources (both negative and positive values). The slow component can be visualized as the amplitude envelope (interrupted red line) of the signal. (D) PSD estimate shows that the slow component is absent from the extracted CSD signal. (E) Rectifying the CSD estimate (black) and then filtering (red) approximately recovers the true slow component (red interrupted). This is how the DG CSD activity signal is obtained. (F) PSD estimate of the rectified and filtered CSD signal recovers the slow component (interrupted red vertical line).

Reviewer #2 (Public Review):

In this manuscript "Inhibition is the hallmark of CA3 intracellular dynamics around awake ripples" the authors obtained Vm recordings from CA1, CA3 and DG neurons while also obtaining local field potentials across the CA1 and DG layers. This enabled them to identify periods of up and down state transitions, and to detect sharp-wave ripples (SWRs). Using these data, they then came to the conclusion that compared to CA1 and DG, the Vm of more CA3 neurons is hyperpolarized at the approximate time of SWRs.

Unfortunately, for the following reasons, the current manuscript does not necessarily support this conclusion:

Recordings are obtained in mice who are recently (same day) recovering from craniotomy surgery/anesthesia and have no training on head fixation. This means that the behavioral state is abnormal, and the animal may have residual anesthesia effects.

The main surgery for implanting the head-fixation apparatus and marking the coordinates for multisite and pipette insertion was carried out at least two days before the experiment. On the day of the experiment animals were briefly lightly anesthetized (<1 hr, at <1% isoflurane at 1 lit/min) for the sole purpose of resecting the dura at the two sites for multisite probe and pipette insertion. This procedure was carried out on the same day as the experiment in order to minimize the time the brain was exposed and optimize the quality of the recordings. Experiments began at least six hours after this short procedure. Furthermore, animals were given time to get familiarized with the behavioral apparatus before recordings began and showed no signs of distress.

Previous studies show that about 95% of isoflurane is eliminated within minutes by exhalation (Holaday et al., 1975). The further elimination of isoflurane proceeds with a fast phase with half-time of about 7-9 min and a slower phase with half-time of about 100-115 min (Chen et al., 1992), with the faster phase reflecting elimination from the brain (Litt et al., 1991). Given these considerations there should be negligible residual isoflurane from the short anesthesia six hours later when recordings are initiated.

In order to further investigate whether the short and light anesthesia during the day of recordings has any effect on the results reported in the paper, we carried out additional experiments in which we performed the surgery, including dura removal, 3 days before the recording session. The animals were habituated under head-fixation on the spherical treadmill for two hour periods each of the two days following the surgery. On the third day after surgery, we carried out recordings without any surgical procedures or anesthesia. The durations of UP and DOWN states without same day anesthesia were similar to those obtained in our previous experiments (Figure 2- figure supplement 4). The additional CA3 whole-cell recordings obtained in these new experiments have the same hyperpolarization features typical of our previous recordings. These additional experiments argue that the brief anesthesia on the day of recordings has no significant effect on the results.

Most of the paper is dedicated to dynamics around up-down state transitions, not focused on ripples.

We changed the title to “Up-Down states and ripples differentially modulate membrane potential dynamics across DG, CA3, and CA1 in awake mice” to reflect the analysis of both UP-DOWN state transitions and ripples. The two analyses are linked as the brain state modulation accounts for the slow Vm modulation around ripples.

Vm should be examined raw first, then split into fast and slow -the cell lives with the raw Vm.

The raw Vm can be obtained by adding the slow and fast Vm components. Hence the behavior of the Vm around ripples can be obtained by adding the panels of columns 1 and 3 in Figure 6. Decomposing into the slow and fast components illustrates how the slow modulation around ripples is due to brain state modulation of the slow component of the Vm (Figure 6).

While some (assumed) CA3 principal cells were hyperpolarized around the time of ripples, saying inhibition is the hallmark of CA3 dynamics around ripples is an exaggeration, especially because it does not seem mechanistically tied to anything else.

While a small fraction of CA3 cells is excited around ripples, the majority is inhibited. We suggest that the inhibition of the majority of CA3 neurons can account for the sparse and selective activation of CA3 around ripples.

The use of ripple onset time is questionable, since the detected onset of the ripple depends on the detector settings, amplifier signal-to-noise ratio, etc. The best and most widely used (including by a subset of these authors) metric is the ripple peak time.

We added Figure 6 - figure supplement 2, which shows that the Vm modulation around peak ripple power is the same as the modulation around ripple start, except for a small time shift due to the fact that the ripple power peaks shortly after ripple start. Our focus on ripple onset facilitates characterizing the timing of pre-ripple activity, such as the Vm depolarization observed before ripple onset for DG and CA1 neurons.

There is not enough raw data (or quality metrics) shown to judge the quality of the data, especially for the whole cell recordings. For instance what was the input resistance of the neurons? Was the access resistance constant?

We added Figure 1 - figure supplement 4, which now describes the mean resting membrane potential, input resistance, burst propensity, and spikes per burst for the recorded cells. These data are provided in Figure 1 - source data 1 together with a recording identifier that can be used to link each cell to all other figure panels and data files. We further added Figure 1- figure supplement 1, which provides examples of morphological information for our recordings, Figure 1 - figure supplement 2 that shows examples of bursts from morphologically identified neurons, and Figure 1 - figure supplement 3 that shows the locations of recorded cells.

There is not enough explanation regarding why the reported results on the spiking of CA1 and CA3 neurons in SWRs is so different than previously published. In general, whole cell recording is not the most reliable way to record spike timing, and the presented whole cell data differ from previously published juxtacellular and extracellular recording methods, which better preserve physiological spiking activity.

The CA1 neurons in this study depolarize and elevate their firing around ripples, consistent with previous intracellular and extracellular recordings. Our study reveals hyperpolarization of the majority of CA3 cells while only a small fraction is depolarized. This is consistent with the sparse activation of CA3 around ripples previously reported with extracellular studies. The overall firing rate change of CA3 neurons around ripples is a balance between the firing rate elevation of the small subset of activated cells and the net decrease in firing across the rest of the population. Since the baseline firing rate of CA3 pyramidal neurons in quiet wakefulness and sleep is low, the ripple-associated inhibition may not be readily observable in the spiking of individual CA3 neurons due to a “floor effect”. The overall rate of CA3 neurons we record increases before ripple onset, consistent with previous studies (Fig. 6D4). The subthreshold hyperpolarization of the majority of neurons provides novel insights into the mechanisms ensuring sparse and selective activation of the CA3 population around ripples.

The number of neurons from each area is not reported.

The number of cells was (indirectly) reported as the number of rows in Figs. 3-7. We now report the number of cells explicitly: 22 DG cells, 32 CA3 cells, and 32 CA1 cells.

There is no verification of cell type so it is inappropriate to assume that all neurons are the principal neurons.

We added Figure 1 - figure supplement 1, which shows morphological identification of recorded cells. We patch multiple cells in each experiment, but we can confirm the morphological identity of principal neurons when all stained cells have morphology of dentate granule cells or CA3/CA1 pyramidal neurons. The properties of morphologically identified cells in Figure 1 - figure supplement 1 are typical of all recorded cells (morphologically identified neurons from Figure 1 - figure supplement 1 are shown as diamonds in Figure 1- figure supplement 4, while the rest are shown as dots). There were no significant differences between the two groups (p > 0.05 t-test; p > 0.05 Wilcoxon rank sum test).

Are the fluctuations in the CA3 Vm generally smaller than for CA1 and DG because of physiology or technical reasons?

The recordings were done in exactly the same way across areas, arguing against technical reasons for any differences observed across the hippocampal subfields.

Reviewer #3 (Public Review):

During slow wave sleep and quiet immobility, communication between the hippocampus and the neocortex is thought to be important for memory formation notably during periods of hippocampal synchronous activity called sharp-wave ripple events. The cellular mechanisms of sharp-wave ripple initiation in the hippocampus are still largely unknown, notably during awake immobility. In this paper, the authors addressed this question using patch-clamp recordings of principal cells in different hippocampal subfields (CA3, CA1 and the dentate gyrus) combined with extracellular recordings in awake head-fixed mice as well as computer modeling. Using the current source density (CSD) profile of local field potential (LFP) recordings in the molecular layer of the dentate gyrus as a proxy of UP/DOWN state activity in the entorhinal cortex they report the preferential occurrence of sharp-wave ripple (recorded in area CA1) during UP states with a higher probability toward the end of the UP state (unlike eye blinks which preferentially occur during DOWN states). Patch-clamp recordings reveal that a majority of dentate granule cells get depolarized during UP state while a majority of CA3 pyramidal cells get hyperpolarized and CA1 pyramidal cells show a more mixed behavior. Closer examination of Vm behavior around state transitions revealed that CA3 pyramidal cells are depolarized and spike at the DOWN/UP transition (with some cells depolarizing even earlier) and then progressively hyperpolarize during the course of the UP state while DGCs and CA1 pyramidal cells tend to depolarize and fire throughout the UP state. Interestingly, CA3 pyramidal cells also tend to be hyperpolarized during ripples (except for a minority of cells that get depolarized and could be instrumental in ripple generation), while DGCs and CA1 pyramidal cells tend to be depolarized and fire. The strong activation of dentate granule cells during ripples is particularly interesting and deserves further investigations. The observation that the probability of ripple occurrence increases toward the end of the UP state, when CA3 pyramidal cells are maximally hyperpolarized, suggests that the inhibitory state of the CA3 hippocampal network could be permissive for ripple generation possibly by de-inactivation of voltage-gated channels thus increasing their excitability (i.e. ability to get excited). Altogether, these results confirm previous work on the impact of slow oscillations on the membrane potential of hippocampal neurons in vivo under anesthesia but also point to specificities possibly linked to the awake state. They also invite to revisit previous models derived from in vitro recordings attributing synchronous activity in CA3 to a global build-up of excitatory activity in the network by suggesting a role for Vm hyperpolarization in preserving the excitability of the CA3 network.

1) In light of recent report of heterogeneity within hippocampal cell types (and notably description of a new CA3 pyramidal cell type instrumental for sharp-wave ripple generation) (Hunt et al., 2018), the small minority of CA3 pyramidal cells depolarized during ripples deserve more attention. These cells are indeed likely key in the generation of sharp wave ripple. Several analyses could be performed in order to decipher whether they have specific intrinsic properties (baseline Vm, firing threshold, burst propensity), whether they are located in specific sub-areas of CA3 (a versus b, deep versus superficial) and whether they are distinctively modulated during UP/DOWN states.

Following the reviewer’s suggestion we now analyze the properties and UDS modulation of the CA3 neurons that are depolarized around ripples (Figure 6 - figure supplement 3). These neurons have comparable resting Vm, spike thresholds, and burst propensity as the rest of the CA3 population (p > 0.05, t-test). These CA3 cells had lower firing probability in the DOWN state. The locations of the depolarized cells are distributed across CA3c,b and are not clustered compared to the rest of the cells (Figure R2).

Figure R2. Proximodistal locations of CA3 cells that depolarize during ripples. Same as Figure 1 - figure supplement 3, but CA3 cells showing depolarization in their ripple-triggered average (RTA) response are marked with black dots. There was no significant difference in the proximodistal locations of these cells compared to the rest of the CA3 population (p > 0.05, t-test).

The population of athorny cells described in Hunt et al. represents a small percentage of CA3 cells (10-20%) that are concentrated in the CA3a region, which we do not sample in our recordings. Hence, the depolarized cells are unlikely to correspond to the athorny cells reported in Hunt et al.

2) The authors use CSD analysis in the DG as a proxy of synaptic inputs coming from the EC to define alternating periods of UP and DOWN states. I have few questions concerning this procedure: 1- It is unclear if only periods when animals was still/immobile were analyzed. 2- How coherent were these periods with slow oscillations recorded in the cortex (which are also recorded with the linear probe?).

The analysis was restricted to periods of immobility, which comprise the majority of the recording time as the animals are not performing any task. Cortical LFPs exhibit high coherence for low frequencies (<1 Hz) with the rectified DG CSD signal (Figure R3), although the contribution of volume conduction to this effect cannot be ruled out.

Figure R3. Coherence between DG CSD power and cortical LFP. (Top) population average magnitude squared coherence between DG CSD power (rectified CSD from the DG molecular layer) and cortical LFP across all recorded datasets. Notice the elevated coherence at low frequencies (< 1 Hz, vertical interrupted line) as well as the peak at theta ( 7-8 Hz). Volume conduction from other brain areas (i.e. the hippocampus) contributes to the cortical LFP and may be responsible for the coherence at theta, as well as at low frequencies. (Bottom) Each row in the pseudocolor image shows the coherence between DG CSD power and cortical LFP for a given dataset.

3- How long did these periods last? Did they occur during classically described hippocampal states (LIA/SIA) or do they correspond to a different state (Wolansky et al., J Neurosci 2006).

The distribution of UP and DOWN state durations is shown in Figure 2 - figure supplement 4.

We also added Figure 2 - supplementary figure 8 that shows the distribution of LIA and SIA transitions as a function of UDS phase. The LIA and SIA states were computed based on LFPs from CA1 stratum radiatum as described in (Hulse et al., 2017). The detected LIA→SIA transitions map very closely to UP→DOWN transitions. The SIA→LIA transitions are also concentrated around DOWN→UP transitions, but the distribution is broader compared to the LIA→SIA transitions. These observations are consistent with UP states broadly overlapping with LIA and DOWN states with SIA.

3) To better characterize hippocampal CSD profiles around ripples and UP/Down states transitions, could you plot ripple and UDS transition-triggered average CSD profiles across hippocampal subfields?

We added Figure 2 - supplementary figure 7 that shows average CSD profiles around UP/DOWN state transitions and ripples.

4) The duration of UP states appears longer than that reported in anesthetized animals. To ascertain this fact could the authors quantify and report mean UP and DOWN states durations? Shorter DOWN states would decrease the probability to detect ripple. Could the authors correct for this bias in their analysis of ripple occurrence during UP and DOWN states?

We report the medians and means of the distributions of UP and DOWN durations in Figure 2 - figure supplement 4. Ripples occur almost exclusively during the UP states, with almost no ripples occurring in DOWN states. Furthermore, the duration of UP and DOWN states is comparable suggesting that the duration of DOWN states does not bias the probability of ripple detection. We also added Figure 2 - figure supplement 2B, showing the rate (in Hz) of ripple occurrence as a function of UDS phase, which explicitly controls for UDS phase occupancy.

The duration of UP and DOWN states in quiet wakefulness depend on the behavior of the animal, attentional state, and external stimuli and need not be the same as in anesthesia or sleep when the animal is not behaving and is less responsive to external stimuli. To provide validation that the extracted UP and DOWN states in quiet wakefulness indeed correspond to genuine brain states, we show that the pupil diameter and ripple rates which are independently extracted are strongly modulated around the extracted UP and DOWN states.

5) The authors report a high coherence between the Vm of an example CA3 pyramidal cells and UP/DOWN state in DG. Was it a general property of a majority of CA3 pyramidal cells? The coherence values should be reported for all CA3 pyramidal cells.

We added Figure 2 - figure supplement 1, which reports the coherence of all cells across the subfields with the rectified DG CSD. The coherence values are similar across cells and subfields. We also report correlations between the slow component of the Vm and DG CSD activity for all cells in Figure 3. Neurons in CA3 exhibit negative correlations in contrast to DG and CA1, with the absolute values of the correlations similar across the subfields.

6) Was the high coherence between DG CSD magnitude and CA3 Vm specific to these slow oscillatory periods or a more general feature of the DG/CA3 functional coupling. For example, was it also observed during theta/movement periods?

Figure 2 - figure supplement 1 reports the coherence of all cells across the subfields with rectified DG CSD over the entire recording duration. Mice do not perform any tasks during the recordings so periods of immobility and quiet wakefulness comprise the majority of the recording session and are the focus of our analysis. During some occasional theta periods there is increased coherence in the theta frequency band (figure R4).

7) Fig. 6 shows depolarization and increase firing in DGCs up to 150 ms prior to ripple onset. However, ripples sometime occur in bursts with one ripple following others. Could such phenomenon explain the firing prior to ripples? (which would in fact correspond to firing during a previous ripple). What is the behavior of firing rate and Vm of different cells types if analysis is restricted to isolated ripples? This analysis is notably important in CA3 where feedback inhibition following a first ripple could lead to hyperpolarization « during » the next ripple.

We added a new figure (Figure 7 - figure supplement 2) that compares Vm aligned to the onset of isolated single ripples vs. ripple doublets. The pre-ripple depolarization in DG and CA1 is similar for isolated ripples and ripple doublets arguing against the hypothesis that pre-ripple responses are a reflection of ripple bursts.

Author Response:

Reviewer #1 (Public Review):

The lateral entorhinal cortex (LEC) receives direct inputs from the olfactory bulb (OB) but their odor response properties have not been well characterized despite a recent increase in interests in the role of LEC in olfactory behaviors. In this study, Bitzenhofer and colleagues provide unprecedented details of odor response properties of layer 2 cells in LEC. The authors first show that LEC neurons respond to odors with a rapid burst of activity time-locked to inhalation onset, similarly to the piriform cortex (PCx), but distinct from the OB. Firing rates of LEC ensembles conveyed information about odor identify whereas timing of spikes odor intensity. The authors then examined the difference between two major cell types in LEC layer 2 - fan cells and pyramidal neurons, and found that, on average, fan cells responded earlier than pyramidal neurons, and pyramidal neurons, but not fan cells, changed their peak timing in response to changes in concentrations, providing a basis for temporal coding of odor concentrations. Additionally, the authors show that inactivation of LEC impairs odor discrimination based on either identify or intensity, and demonstrate different cellular properties of fan cells and pyramidal neurons. Finally, the authors also examined the odor response properties of hippocampal CA1 neurons, and showed that odor identify can be decoded by firing rate responses, while decoding of odor concentration depended on spike timing.

The authors performed a large amount of experiments, and provide an impressive set of data regarding odor response properties of LEC layer 2 neurons in a cell type specific manner. The results reported are very interesting, and will be a point of reference for future studies on odor coding and processing in the LEC. The manuscript is clearly written, and data are well analyzed and presented clearly. I have only relatively minor concerns or suggestions.

1. The authors infer the time at which "mice could discriminate odors" from the time at which d-prime becomes significantly different between baseline and odor stimulation conditions (line 111 and line 121). However, the statistical test applied to these data does not guarantee that an observer can accurately discriminate odors. For example, a small p-value can be obtained even when discrimination accuracy is only slightly above chance if there are many trials. The statement such as "mice could discriminate two odors by as early as 225 ms after inhalation onset" (line 111) can be misleading because this might sound as if mice can accurately discriminate odors at this timepoint, while this is not necessarily the case (as indicated by the d-prime value).

We have added plots of performance accuracy over time under control conditions (LED off) to Figure 2-supplement 1. These plots of fraction of correct responses (binned every 50 ms) show that mice (n = 6) are making choices significantly different from chance within 200 ms of odor inhalation. We changed the wording in the Results to now say: “Moreover, by analyzing lick timing, we determined that the discriminability measure d’ became significantly different under control conditions as early as 225 ms after inhalation onset and performance accuracy increased within 200 ms of inhalation (Fig. 2b, Figure 2-supplement 1).”

1. Optogenetic identification can be a little tricky when identifying excitatory neurons as in this study. Please discuss some rational or difficulty regarding how to distinguish those that are activated directly by light from those activated indirectly (i.e. synaptically). Do the results hold if the authors use only those that the authors are more confident about identification?

We only used the cells that were confidently identified using a combination of two criteria. First, tagged cells had to show a significant increase in firing (p_Rate <0.01) during the 5 ms LED illumination period versus 100 randomly selected time windows before LED stimulation. Cells also had to respond with a fixed latency to reduce the chance of including cells recruited by polysynaptic excitation. Further, we used the stimulus associated spike latency test (SALT) as detailed in Kvitsiani et al., 2013. To be judged as tagged, units had to show significantly less spike jitter during the 5 ms LED illumination than 100 randomly selected time windows before LED stimulation (p_SALT<0.01). Only those cells with BOTH p_Rate<0.01 and p_Salt<0.01 were considered as tagged (both methods typically agreed for most cells). Moreover, slice work testing synaptic connections between LEC layer 2 cells found extremely low levels of connectivity between fan and pyramidal cells Nilssen et al., J. Neuroscience, 2018. This makes it unlikely that LED-induced firing of fan or pyramidal cells would recruit indirectly (synaptically) excited cells.

1. The authors sort odor response profiles by peak timing, and indicate that odor responses peak at different timing that tiles respiration cycles. However, this analysis does not indicate the reliability of peak timing. Sorting random activity by "peak timing" could generate similar figure. One way to show the reliability or significance of peaks is to cross-validate. For instance, one can use a half of the trials to sort, and plot the rest of the trials. If the peak timing is reliable, the original pattern will be replicated by the other half, and those neurons that are not reliable will lose their peaks. Please use such a method so that we can evaluate the reliability of peaks.

We analyzed the data as suggested by this reviewer as shown below (Author response image 1). Plotting only the odd trials sorted by the odd trials in the dataset (top) looked identical to the data from all trails used in Figure 1g. More importantly, plotting only the even trials sorted by the odd trials (bottom), though noisier due to trial-by-trial variation, showed the same general structure of tiling throughout the respiration cycle for OB cells.

Author response image 1

Reviewer #2 (Public Review):

In this study, Bitzenhofer et al recorded odor-evoked activity in the LEC and examined the coding of odor identity and intensity using extracellular recordings in head-fixed mice, and used the standard suite of quantitative tools to interpret these data (decoding analyses, dimensionality reduction, etc). In addition, they performed behavioral experiments to show the necessity of LEC in odor identity and intensity discrimination, and deploy some elegant and straightforward 'circuit-busting' slice physiology experiments to characterize this circuit. Importantly, they performed some of their experiments in Ntng1-cre and Calb-cre mice, which allowed them to differentiate between the two major classes of LEC principal neurons, fan cells and pyramidal cells, respectively. Many of their results are contrasted with what has previously been observed in the piriform cortex (PCx), where odor coding has been studied much more extensively.

Their major conclusions are:

Cells in the LEC respond rapidly to odor stimuli. Within the first 300 ms after inhalation, odor identity is encoded by the ensemble of active neurons, while odor intensity (more specifically, responses to different concentrations) is encoded by the timing of the LEC response; specifically, the synchrony of the response. These coding strategies have been described in the PCx by Bolding & Franks. Bolding also found two populations of responses to different concentrations: one population of responses was rapid and barely changed with concentration and the second population of responses had onset latencies that decreased with increasing concentration. Roland et al also found two populations of responses using calcium imaging in anesthetized mice: one population of responses was concentration-dependent and another population was 'concentration-invariant'. However, neither Bolding nor Roland were able to determine whether these populations of responses emerged from distinct populations of cells. Here, the authors elegantly register these two response types in LEC to different cell types: fan cells respond early and stably, and pyramidal cells response latencies decrease with concentration. This is a novel and important finding. They also showed that, unlike PCx or LEC where concentration primarily affects timing rather than rate/number, odor concentration in CA1 is only reflected in the timing of responses.

Using optogenetic suppression of LEC in a 2AFC task, the authors purport to show that LEC is required for both the discrimination of odor identity and odor intensity. If true, this is an important result, but see below.

In slice experiments, the authors characterize the differential connectivity of fan and pyramidal cells to direct olfactory bulb input, input from PCx, and inhibitory inputs from SOM and PV cells. This work is elegant, novel, and important, although it is a little out of place in this manuscript. As such, their findings are irrelevant/orthogonal to the rest of the results in this study. But fine.

The simultaneous recordings from three different stations along the olfactory pathway are impressive.

Major concern

My major concern with this manuscript regards the behavioral experiments. The authors show that blue light over the LEC in GAD2-Cre/Ai32 mice completely abolishes (i.e. to chance) the mouse's ability to perform a 2AFC task discriminating between either two different odorants or one odorant at different concentrations. Their interpretation is that LEC is required for rapid odor-driven behavior. The sensory component of the task is so easy, and the effect is so striking that I find this result surprising and almost too good to be true. The authors do control for a blue-light distraction effect by repeating the experiments in mice that don't express ChR2, but do not control for the effect of rapidly shutting down a large part of the sensory/limbic system. If they did this experiment in the bulb I would be impressed with how clean the result was but not conceptually surprised by the outcome. I think a different negative control is needed here to convince me that the LEC is necessary for this simple sensory discrimination task. For example, the authors could activate all the interneurons (i.e. use this protocol) in another part of the brain, ideally in the olfactory pathway not immediately upstream of the LEC, and show that the behavior is not affected.

This reviewer suggests a negative control experiment for the effects we observe on behavior when optogenetically silencing LEC. However, we disagree that it would be informative to silence other olfactory pathways in search of those that do not affect behavior. Our strong effects on behavior are also in complete agreement with recent findings that muscimol inactivation of LEC abolishes discrimination of learned odor associations (Extended Data Figure 8, Lee et. al., Nature, 2021).

More specifically, both the presentation and the interpretation of the data are confusing. First, there is a lack of detail about the behavioral task. I was not sure exactly when the light comes on and goes off, when the cue was presented, and when the reward was presented. In the manuscript they say (line 108) "…used to suppress activity during odor delivery on a random subset…". There is nothing more about this in the figure legend or Methods. The only clue to this is the dotted line in the 'LED On' example at the bottom of Fig. 2a. The authors also say that (line 660) "Trials were initiated with a 50 ms tone." When exactly was the tone presented? In the absence of any other information, I assume it was presented at odor onset. When was the reward presented? Lines 106-7 say "Mice were free to report their choice (left or right lick) at any time within 2 s of odor onset." Presumably this means the reward was presented to one of the ports for 2 seconds, starting at odor onset.

The LED is applied during odor delivery, the 50 ms tone immediately precedes odor delivery, and water reward is dispensed after the first lick at the correct lick port during the choice period. The choice period begins with the odor onset and odor delivery is terminated by the first lick at either the correct or incorrect port. If there is no lick at either port, odor delivery lasts 1s and is followed by an extended choice period (terminated by correct or incorrect lick) lasting 1s. To clarify the behavior protocol, we have included a schematic of the trial structure in Figure 2-supplement 1.

These details matter because the authors want to claim that "LEC is essential for rapid odor-driven behavior." The data presented in support of this claim are (1) that mice perform this task at chance levels in LED On trials, presumably based on which port the mouse licked first (this is the 'essential' part), and (2) that in control in LED Off trials, d' becomes statistically different from baseline after ~200 ms (this is the 'rapid' part).

To further support the argument that LEC is required for rapid odor-driven behavior, we now show a plot of % correct responses over time from first odor inhalation.

On first reading, these suggested that shutting off LEC makes odor discrimination worse and/or slower. However, the supplementary data clarifies several things. First, the mice never Miss (Fig.2S.2a & c), meaning then they always lick. Second, in LED Off trials (F2S2 & e), the mice make few mistakes, and these only occur immediately after inhalation, presumably meaning the mice occasionally guess, possibly in response to the auditory cue. Thus, the mean time to lick is much shorter for Error trials than Correct trials. To state the obvious, the mice often wait >300 ms before they lick, and when they do wait, they never make mistakes. Now, in the LED On trials, the mice almost always lick within the first 300 ms and perform at chance levels, with the distribution of lick times for Correct and Error trials almost overlapping. In fact, although the authors claim LEC is required for rapid odor discrimination, the mean time to lick on Correct trials appears to decrease in LED On trials. This makes me think that the mice are making ballistic guesses in response to the tone in LED On cases, which doesn't necessarily implicate a dependence on LEC for odor discrimination.

We do not believe that mice are making ballistic guesses in response to the tone for LED on trials. First, although a 50 ms tone immediately precedes odor delivery, all data in Figure 2-supplement 1 shows lick times aligned to the first inhalation of odor. Thus, time 0 ms is not the tone or subsequent odor onset but rather a variable time point coinciding with the first odor inhalation (the delay from odor onset to first inhalation is ~300 ms, the average respiration interval under our conditions). In fact, we excluded trials if mice made premature licks between the time of odor onset and first odor inhalation. We re-analyzed these trials to test the reviewer’s idea that mice were more likely to make fast ballistic guesses when the LEC was silenced. However, we saw no evidence that mice made more premature licks in trials with LED on (Author response image 2).

Author response image 2

The authors' interpretation of their data would be more solid if, for example, there were a delay between the auditory cue and odor delivery and/or if the reward was only available with some delay after the odor offset. Here, however, it seems just as likely as not that the mice are making ballistic guesses in response to the tone in LED On cases, which doesn't necessarily involve dependence on LEC for odor discrimination. Here, the divergence of d' from baseline in the control (i.e LED Off) condition seems mostly because mice take longer to correctly discriminate under control conditions. While this is not formally contradictory to LEC is essential for rapid odor-driven behavior", it is nevertheless a bit contrived and misleading. An interesting (thought) experiment is what would happen if the authors presented a tone but no odor. I would guess that the mice would continue licking randomly in Light On trials.

While a delay between odor delivery and reward would have been useful for some aspects of interpreting the behavior, we would have lost the ability to examine the role of LEC in response timing. To address this reviewer’s concern, we have added a section to the Discussion mentioning caveats related to the interpretation of experiments using acute optogenetic silencing to understand behavior.

Author Response:

Reviewer #1 (Public Review):

The authors make juxtacellular recordings on awake mice, which should yield clear responses of actions potentials, and employ a number of manipulations to silence pathways. They also record from a "non"-whisker secondary thalamic region, LP, as a null hypothesis to establish if certain effects are related to "behavior" - read arousal or saliency". I have no major qualms.

In light of Petersen's paper (Cell Reports 2014) on cholinergic effects on spike rates in primary whisker somatosensory cortex, I can imagine that the authors considered measuring from cholinergic neurons in nucleus basalis during whisking. I'll assume that this is easier said than done. As such, the current manuscript passes my threshold for publication modulo issues raised below that are related to anatomy.

The cholinergic experiments are an interesting idea. However, inactivation of S1 did not change the relationship between POm and whisking, suggesting that cholinergic modulation of S1 and thereby corticothalamic output are not the key mechanism. It is conceivable that acetylcholine modulates POm directly, but the critical experiments would involve extensive manipulations of POm (a whole additional study). Nevertheless, we have added a reference to Eggermann & Petersen and discussed this issue further in the revision.

I provide a figure-by-figure critique:

(1) Recent work from Deschênes et al (Neuron 2016) points to a description of whisking in terms of Angle = Set-point_angle - Whisking-amplitude [1 + cosine(Phase - Phase_0)], where Phase is a rapidly varying, typically rhythmic function of time. Why not use this notation as opposed to yet another descriptive statistic and report the kinetics as the time averaged parameters , i.e., the most forward position, and ,Whisking-amplitude>, i.e., the half-amplitude of the average whisk?

We are not entirely sure what the reviewer means by “another descriptive statistic” as we do not introduce new approaches for analyzing whisking in this paper. (Perhaps the reviewer refers to “median angle”, which is an average of all the whisker positions on a single frame. We use this measurement because our videos contain the entire whisker field rather than just a single whisker as in our other studies, e.g. Hong et al 2018, Rodgers et al 2021). We based our parameterization of median angle on two publications: Hill et al (2011 Neuron) and Moore et al (2015 PLoS Biology). Moore et al describes whisking as a function of phase, amplitude, and midpoint:

where 𝜃(t) is the median whisker angle at time 𝑡 , 𝜙 is the phase as computed by the Hilbert transform of the filtered whisker angle, 𝜃^Amplitude is the difference between the most protracted and retracted whisker positions over a single cycle, and 𝜃^midpoint is the central angle of a single whisk cycle. As we understand the reviewer, we are using the formulation they describe. We are happy to consider alternate formulations if we are missing something.

A critical issue is to confirm where the recording were made. This the authors should supply at least a typical record of anatomy from their POm as well as VPM and LP recording. The beauty of the juxtacellular technique is that neurons can be labeling after the recording

We used the juxtacellular recording technique for its superior recording quality. We did not label individual cells after recording because we recorded multiple cells per animal over several days. The number of cells would complicate matching of filled cells to recorded physiological data, and biotin filling is not stable over multiple days (beyond 36 hours). Instead, as described in the original manuscript, we tracked the relative locations of all inserted pipettes and labeled the final track with DiI. Cells were roughly localized along the tracks using relative microdrive depths. Due to the morphological homogeneity of thalamic neurons, filling individual cells would not be more informative than labelling the recording site with DiI. New Figure 1 – figure supplement 1 includes representative histology images from our recordings in POm, VPM, LP, and M1.

(2) Did the authors make sure that the mystacial pad is not moving by imaging the pad as opposed to just the shaft of the whiskers? The top view in Figure 1A makes this hard to check.

To address this concern, we provide new data, in which both the cut and uncut sides of the face of mice were imaged. We measured the movement of the mystacial pads as motion energy – the mean absolute difference in pixel values across video frames. The motor nerve surgery almost completely abolished movement of the mystacial pad. A new figure panel (Figure 2B) demonstrates the movement of the normal and paralyzed mystacial pads.

Further, did the authors perform post-hoc anatomy to insure that both the ramus buccolabialis inferior and ramus buccolabialis superior muscles were cut? This is critical; it is also easy to leave the maxillolabialis (external retractor) innervated if the cut is too far rostral.

We did not attempt to cut muscles. We only cut the motor nerve. We did not examine the face post mortem, as it was obvious that both whisker and mystacial pad movement were absent (as in new Figure 2B).

(3/4) As relevant background, the text should note that whisker primary motor cortex maintains a copy of the envelope of the whisking, i.e., an ill-defined summation of set-point and amplitudes, even if the sensory input (Ahrens & Kleinfeld J Neurophysiol 2004) or motor output (Fee et al. J Neurophysiol 1997) in the periphery are cut.

The Results text now cites these papers as motivation for the experiments of Figure 3.

(6/7) Same comments in (1) in whisking parameters and anatomy.

As we discussed in (1), we are using the conventional parameterization of others. Histological examples are now included in Figure 1 – figure supplement 1.

Reviewer #3 (Public Review):

Previous studies in urethane-anesthetized rats (PMID 16605304) proposed that POm cells code whisker movements. This was observed using "artificial whisking" procedures (stimulating the motor nerve to produce a whisking-like movement). It has been clear for some time now that there are substantial (obvious) differences between this procedure and natural whisking. In addition, under urethane-anesthesia animals are in a sleep-like state that is very dissimilar to waking (although some work has tested the effect of network state on artificial whisking responses in both primary thalamus and cortex; see 25505118). In the present study, the authors measured activity in POm cells during whisking in awake (head-fixed) mice to determine if they code whisking movement. However, this seems to have already been done previously. For instance, Moore et al (2015; 26393890) found that coding of whisking in the ascending paralemniscal pathway, including POm, is "relatively poor" (as stated in the abstract), which is the same conclusion reached in the present study. The authors should clarify the main differences observed in whisking coding between their study and previous work.

The authors then focused on the idea that POm codes behavioral state. However, many studies have previously determined that state has a great impact on thalamocortical dynamics; thalamic cells are very sensitive to state including cells in primary whisker thalamic nuclei, such as VPM, and these effects can be produced by neuromodulators (see work by Castro-Alamancos' group, for example, 16306412). There is nothing special about VPM in this regard; other thalamic sensory nuclei are also sensitive to behavioral state and neuromodulators. Therefore, the observation that POm and LP cells are sensitive to state is unsurprising. It is also known that these thalamic state changes have a great impact on the state of the cortex (see 20053845), which seems very relevant to the main conclusion. The POm has to be doing something different than coding behavioral state since most thalamic nuclei do this. The study did not identify the role of POm, which certainly has to be different from LP (otherwise, why would these nuclei be differentiated?). POm is unlikely to be specialized for monitoring state since this is done by most of the thalamus -including VPM, which projects to the same cortical region. Thus, while it is interesting that most of the whisker-related activity in POm is state-dependent, the study does not clarify the role of POm.

We have added the references we did not already include to our text and improved our discussion.

Prior studies (such as Moore et al 2015 and Urbain et al 2015) have previously characterized the encoding of whisker motion in POm. Indeed, we note the consistency between our results and such studies in both the introduction and conclusion. Here we expand upon prior studies to directly test two prominent hypotheses about the role of the paralemniscal pathway: that it encodes sensory reafference, and that it inherits a motor efference copy from cortical and subcortical regions. We present the impact of several manipulations of the vibrissal system (facial paralysis, cortical silencing, and lesion of superior colliculus) on thalamic activity that, to our knowledge, have not been previously reported. Moreover, we leveraged a novel comparison of POm and LP to test whether movement‐correlations of POm reflected true motor modulation or rather state dependency. We have provided evidence that the coupling of POm activity to whisking reflects state rather than motor signals. We never suggested that POm is a unique monitor of behavioral state. We suggest instead that secondary thalamic nuclei may be state‐modulated and have specific impacts on response gain and plasticity in their respective cortical areas. While our work is consistent with previous studies, we believe these results are novel extensions of past work.

The main strength of the study is that it was performed in awake mice with behavioral state monitoring, which contributes to the current understanding of active whisking coding in the complex network of the vibrissa system.

In our opinion, the main strength of our study is its multiple manipulations to test the sources of modulation and the leveraging of a POm‐LP comparison. We have revised the text to reinforce these points.

Author Response

Reviewer #1 (Public Review):

Bice et al. present new work using an optogenetics-based stimulation to test how this affects stroke recovery in mice. Namely, can they determine if contralateral stimulation of S1 would enhance or hinder recovery after a stroke? The study provides interesting evidence that this stimulation may be harmful, and not helpful. They found that contralesional optogenetic-based excitation suppressed perilesional S1FP remapping, and this caused abnormal patterns of evoked activity in the unaffected limb. They applied a network analysis framework and found that stimulation prevented the restoration of resting-state functional connectivity within the S1FP network, and resulted in limb-use asymmetry in the mice. I think it's an important finding. My suggestions for improvement revolve around quantitative analysis of the behavior, but the experiments are otherwise convincing and important.

Thank you for the positive feedback regarding our work.

Other comments - Data and paper presentation:

1) Figure 1A is misleading; it appears as if optogenetic stimulation is constant (which indeed would be detrimental to the tissue). Also, the atlas map overlaps color-wise with conditions; at a glance it looks like the posterior cortex might be stimulated; consider making greyscale?

We have updated Figure 1A to address these concerns.

Reviewer #2 (Public Review):

These studies test the effect of stimulation of the contralateral somatosensory cortex on recovery, evoked responses, functional interconnectivity and gene expression in a somatosensory cortex stroke. Using transgenic mice with ChR2 in excitatory neurons, these neurons are stimulated in somatosensory cortex from days 1 after stroke to 4 weeks. This stimulation is fairly brief: 3min/day. Mice then received behavioral analysis, electrical forepaw stimulation and optical intrinsic signal mapping, and resting state MRI. The core finding is that this ChR2 stimulation of excitatory neurons in contralateral somatosensory cortex impairs recovery, evoked activity and interconnectivity of contralateral (to the stimulation, ipsilateral to the stroke) cortex in this localized stroke model. This is a surprising result, and resonates with some clinical findings, and a robust clinical discussion, on the role of the contralateral cortex in recovery. This manuscript addresses several important topics. The issue of brain stimulation and alterations in brain activity that the studies explore are also part of human brain stimulation protocols, and pre-clinical studies. The finding that contralateral stimulation inhibits recovery and functional circuit remapping is an important one. The rsMRI analysis is sophisticated.

Thank you for the supportive comments regarding our manuscript

Concerns:

1) The gene expression data is to be expected. Stimulation of the brain in almost any context alters the expression of genes.

We agree with the reviewer that stimulation of the brain is expected to broadly alter gene expression. However, in this set of studies, we examined a subset of genes that are of particular interest in neuroplasticity, and compared expression in ipsi-lesional vs. contra-lesional cortex in the presence or absence of contralesional stimulation during the post stroke recovery period. Genes like Arc, for example, have been shown by our group to be necessary for perilesional plasticity and recovery (Kraft, et al., Science Translational Medicine, 2018). The finding that validated plasticity genes are suppressed by contralesional stimulation is consistent with the central finding that contralesional stimulation suppresses the recovery of normal patterns of brain organization and activity. Importantly, there were also genes associated with spontaneous recovery that were unaltered or increased by contra-lesional brain stimulation. While these data do not provide causal associations, they may prove to be useful for developing hypotheses regarding molecular mechanisms involved in spontaneous brain repair for future studies.

In light of the reviewer’s comment, we have altered text throughout to not focus on specific directionality of transcripts. Instead, we indicate that relevant transcript changes are those that are altered in association with spontaneous recovery, and which are altered in the opposite direction with contralesional brain stimulation.

Minor points.

1) Was the behavior and the functional imaging done while the brain was being stimulated?

We have updated the methods (page 17) to clarify that the only experiments during which the photostimulus occurred during neuroimaging are reported in new Figure 6, and to clarify that photostimulation did not occur during the behavioral tests of asymmetry.

2) It would be useful to understand what is being stimulated. The stimulation method is not described. Is an entire cortical width of tissue stimulated, and this is what is feeding back onto the contralateral cortex? Or is this stimulation mostly affecting excitatory (CaMKII+) cells in upper or lower layers? This will be important to be able to compare to the Chen et al study that gave rise to the stimulation approach here. This gets to the issue of the circuitry that is important in recovery, or in inhibiting recovery. One might answer this question by doing the stimulation and staining tissue for immediate early gene activation, to see the circuits with evoked activity. Also, the techniques used in this study could be applied with OIS or rsMRI during stimulation, to determine the circuits that are activated.

We have clarified the stimulation protocol in response to Essential point 2.2. Due to light scattering and appreciable attenuation of 473nm in brain tissue, only ~1% of photons penetrate to a depth of 600 microns. Experimentally, this provides superficial-layer specificity to Layer 2/3 Camk2a cells (https://doi.org/10.1016/j.neuron.2011.06.004)

To answer the question of what circuits are affecting recovery, we performed 2 sets of additional experiments – Experiment 1: OISI during photostimulation before and after photothrombosis, and Experiment 2: tissue staining for IEG expression (cFOS). We describe each below:

Experiment 1 New results are included from 16 Camk2a-ChR2 mice (Results, page 10-11; Methods, page 18) and reported as new Figure 6. Similar to the previously reported experiments, all mice were subject to photothrombosis of left S1FP, half of which received interventional optogenetic photostimulation beginning 1 day after photothrombosis (+Stim) while the other half recovered spontaneously (-Stim). To visualize in real time whether contralesional photostimulation differentially affected global cortical activity in these 2 groups, concurrent awake OISI during acute contralesional photostimulation was performed in +Stim and –Stim groups before, 1, and 4 weeks after photothrombosis. At baseline, all mice exhibited focal increases in right S1FP activity during photostimulation that spread to contralateral (left) S1FP and other motor regions approximately 8-10 seconds after stimulus onset. While activity increases within the targeted circuit, subtle inhibition of cortical activity can also be observed in surrounding non-targeted cortices. Thus, activity both increases and decreases in different cortical regions during and after optogenetic stimulation of the right S1FP circuit. Of note, regions that are inhibited by S1FP stimulation show more pronounced decreases in activity in +Stim mice at 1 and 4 weeks compared to baseline and were significantly larger in +Stim mice compared to –Stim mice. We conclude that focal stimulation of contralesional cortex results in significant, widespread inhibitory influences that extend well beyond the targeted circuit.

Experiment 2 For experiment 2, we hypothesized that IEG expression would increase in photostimulated regions, cortical regions functionally connected to targeted areas, and potentially deeper brain regions. For the IEG experiments, healthy ChR2 naïve animals (C57 mice) or CamK2a-ChR2 mice were acclimated to the head-restraint apparatus described in the manuscript used for photostimulation treatment. Once trained, awake mice were subject to the same photostimulus protocol as described in the manuscript applied to forepaw somatosensory cortex in the right hemisphere. After stimulation, mice were sacrificed, perfused, and brains were harvested for tissue slicing and immunostaining for cFOS. Tissue slices containing right and left primary forepaw somatosensory cortex and primary and secondary motor cortices (+0.5mm A/P) or visual cortex (-2.8mm A/P) were examined for cFOS staining and compared across groups.

Below is a summary table of our findings, and representative tissue slices. While c-FOS IHC was successful, results are not consistent with expectations from the mouse strains used. Only 1 ChR2+ mouse exhibited staining patterns consistent with local S1FP photostimulation, while expression in ChR2- mice was more variable, and in some instances exhibits higher expression in targeted circuits compared to ChR2+ mice. It is possible that awake behaving mice already exhibit high activity in sensorimotor cortex at rest, which might obscure changes specific to optogenetic photostimulation. Regardless, because the tissue staining experiments were inconclusive in healthy animals, we did not proceed with further experiments in the stroke groups, and do not report these findings in the manuscript.

3) Also, it is possible that contralateral stimulation is impairing recovery, not through an effect on the contralateral cortex (the site of the stroke), but on descending projections, or theoretically even through evoking activity or subclinical movement of the contralateral limb (ipsilateral to the stroke). By more carefully mapping the distribution of the activity of the stimulated brain region, and what exactly is being stimulated, these issues can be explored.

The reviewer raises an excellent point. We have added to the “Limitations and Future work” section of the Discussion on pages 15-16

Author Response

Reviewer #2 (Public Review):

Activation of TEAD-dependent transcription by YAP/TAZ has been implicated in the development and progression of a significant number of malignancies. For example, loss of function mutations in NF2 or LATS1/2 (known upstream regulators that promote YAP phosphorylation and its retention and degradation in the cytoplasm) promote YAP nuclear entry and association with TEAD to drive oncogenic gene transcription and occurs in >70% of mesothelioma patients. High levels of nuclear YAP have also been reported for a number of other cancer cell types. As such, the YAP-TEAD complex represents a promising target for drug discovery and therapeutic intervention. Based on the recently reported essential functional role for TEAD palmitoylation at a conserved cysteine site, several groups have successfully targeted this site using both reversible binding non-covalent TEAD inhibitors (i.e., flufenamic acid (FA), MGH-CP1, compound 2 and VT101~107), as well as covalent TEAD inhibitors (i.e., TED-347, DC-TEADin02, and K-975), which have been demonstrated to inhibit YAP-TEAD function and display antitumor activity in cells and in vivo.

Here, Fan et al. disclose the development of covalent TEAD inhibitors and report on the therapeutic potential of this class of agents in the treatment of TEAD-YAP-driven cancers (e.g., malignant pleural mesothelioma (MPM)). Optimized derivatives of a previously reported flufenamic acid-based acrylamide electrophilic warhead-containing TEAD inhibitor (MYF-01-37, Kurppa et al. 2020 Cancer Cell), which display improved biochemical- and cell-based potency or mouse pharmacokinetic profiles (MYF-03-69 and MYP-03-176) are described and characterized.

Strengths:

All of the authors' claims and conclusions are very well supported and justified by the data that is provided. Clear improvements in biochemical- and cell-based potencies have been made within the compound series. Cell-based selective activities in the HIPPO pathway defective versus normal/control cell types are established. Transcriptional effects and the regulation of BMF proapoptotic mRNA levels are characterized. A 1.68 A X-Ray co-crystal structure of MYF-03-69 covalently bound to TEAD1 via Cys359 is provided. In vivo efficacy in a relevant xenograft is demonstrated, using a 30 mg/kg, BID PO dose.

We thank the reviewer for appreciating and highlighting the strengths of our study.

Weaknesses:

Beyond the impact on BMF gene regulation, new biological insights reported here for this compound series are moderate. Progress and differentiation with respect to activity and/or ADME PK profiles relative to the very closely related and previously described (Keneda et al. 2020 Am J Cancer Res 10:4399. PMID 33415007) acrylamide-based covalent TEAD inhibitor K-975 (identical 11 nM cell-based potencies when compared head-to-head and identical reported in vivo efficacy doses of 30 mg/kg) is not entirely clear. Demonstration of on-target in vivo activity is lacking (e.g., impact on BMF gene expression at the evaluated exposure levels).

We thank the reviewer’s question. We have compared mouse liver microsome stability and hepatocyte stability of K-975 and MYF-03-176 and found that K-975 is metabolically less stable.

Consistently, when NCI-H226 cells derived xenograft mice were dosed with 30 mg/kg K-975 twice daily, the tumors kept growing and reach more than 1.5-fold volume on 14th day. While with the same dosage, MYF-03-176 showed a significant tumor regression. K-975 did not reach such efficacy even with 100 or 300 mg/kg twice daily, either in NCI-H226 or MSTO-211H CDX mouse model according to the paper (Keneda et al. 2020 Am J Cancer Res 10:4399).

To demonstrate the on-target in vivo activity, we tested expression of the TEAD downstream genes and BMF in tumor sample after 3-day BID treatment (PD study) and we observed reduction of CTGF, CYR61, ANKRD1 and an increase of BMF, which indicates an on-target activity in vivo.

Author Response

Reviewer #1 (Public Review):

This study provides further detailed analysis of recently published Fly Atlas datasets supplemented with newly generated single cell RNA-seq data obtained from 6,000 testis cells. Using these data, the authors define 43 germline cell clusters and 22 somatic cell clusters. This work confirms and extends previous observations regarding changing gene expression programs through the course of germ cell and somatic cell differentiation.

This study makes several interesting observations that will be of interest to the field. For example, the authors find that spermatocytes exhibit sex chromosome specific changes in gene expression. In addition, comparisons between the single nucleus and single cell data reveal differences in active transcription versus global mRNA levels. For example, previous results showed that (1) several mRNAs remain high in spermatids long after they are actively transcribed in spermatocytes and (2) defined a set of post-meiotic transcripts. The analysis presented here shows that these patterns of mRNA expression are shared by hundreds of genes in the developing germline. Moreover, variable patterns between the sn- and sc-RNAseq datasets reveals considerable complexity in the post-transcriptional regulation of gene expression.

Overall, this paper represents a significant contribution to the field. These findings will be of broad interest to developmental biologists and will establish an important foundation for future studies. However, several points should be addressed.

In figure 1, I am struck by the widespread expression of vasa outside of the germ cell lineage. Do the authors have a technical or biological explanation for this observation? This point should be addressed in the paper with new experiments or further explanation in the text.

Thank you for pointing this out. We found that our single cell dataset shows a similar (low) level of vasa expression outside the germline, suggesting that this is not due to single nucleus versus single cell RNA-seq (cluster 1, red in the lefthand umap).

Analyzing the single nucleus RNA-seq in more detail revealed that, compared to the germline, both the fraction of cells in a cluster expressing vasa and the level at which they express it are very low. This analysis is included in a new Figure 1 – figure supplement 1. It is likely that much of this is due to a technical artifact, such as ambient RNA. Finally, we note in the resubmission that vasa is in fact expressed in embryonic somatic cells, and thus some of the vasa expression we observe may be real (Renault. Biol Open 2012; https://doi.org/10.1242/bio.20121909).

Plots in the original submission drew undue attention to the few somatic cells that exhibited vasa signal, due to the fact that expressing cell points were forced to the front of the plot. Given our new analysis reporting the low levels and fraction of cells exhibiting vasa expression (Figure 1 – figure supplement 1), we have modified the panels of Figure 1, changing point size to more faithfully reflect the small proportion of somatic cells with some vasa expression.

The proposed bifurcation of the cyst cells into head and tail populations is interesting and worth further exploration/validation. While the presented in situ hybridization for Nep4, geko, and shg hint at differences between these populations, double fluorescent in situs or the use of additional markers would help make this point clearer. Higher magnification images would also help in this regard.

We thank the reviewer for their suggestions on clarifying the differences between HCC and TCC populations. As suggested, we have repeated the FISH experiments of Nep4 and geko with higher resolution, and included the additional marker Coracle that demarcates the junction between HCC and TCC (Figure 6O,Q,S,T). These panels replaced previous Nep4 and geko FISH images (see previous Figure 6Q,U,U’). FISH for Nep4 validated the split, and the enrichment of geko strongly suggests that this arm represents one cell type (HCCs). We have not yet identified a gene reciprocally enriched to the other arm. Therefore, in the revised submission, we call the assignment of TCC identity, and to a lesser extent, HCC identity ‘tentative’, but point out that genes predicted to be enriched to one or the other arm represent fertile candidates for the field to test.

Reviewer #2 (Public Review):

In this manuscript the authors explain in greater detail a recent testis snRNAseq dataset that many of these authors published earlier this year as part of the Fly Cell Atlas (FCA) Li et al. Science 2022. As part of the current effort additional collaborators were recruited and about 6,000 whole cell scRNAseq cells were added to the previous 42,000 nuclei dataset. The authors now describe 65 snRNseq clusters, each representing potential cell types or cell states, including 43 germline clusters and 22 somatic clusters. The authors state that this analysis confirms and extends previously knowledge of the testis in several important areas.

“However, in areas where testis biology is well studied, such as the development of germ cells from GSC to the onset of spermatocyte differentiation, the resolution seems less than current knowledge by considerable margins. No clusters correspond to GSCs, or specific mitotic spermatogonia, and even the major stages of meiotic prophase are not resolved. Instead, the transitions between one state and the next are broad and almost continuous, which could be an intrinsic characteristic of the testis compared to other tissues, of snRNAseq compared to scRNAseq, or of the particular experimental and software analysis choices that were used in this study.”

Note that the referee raises the same issue later in their review also. To respond succinctly, we placed the relevant sentence from a later portion of this referee’s comment here

“Support for the view that the problems are mostly technical, rather than a reflection of testis biology, comes from studies of scRNAseq in the mouse, where it has been possible to resolve a stem cell cluster, and germ cell pathways that follow known germ cell differentiation trajectories with much more discrete steps than were reported here (for example, Cao et al. 2021 cited by the authors).”

Respectfully, we have a different interpretation of other work as cited by this referee. Our data, as well as that from others, supports the notion that transitions are generally broad and continuous and are indeed a feature of testis biology. As we report here, data from both single cell and single nucleus RNAseq exhibit transitions from one cluster to the next. Thus, this feature cannot be due to the choice of method (single cell versus single nucleus).

In fact, prior scRNA-seq results on systems containing a continuously renewing cell population, such as is the case in the testis, do indeed exhibit a contiguous trajectory rather than discrete, well-separated cell states in gene expression space (that is, in a UMAP presentation). For example, this is the case from single-cell or single-nucleus sequencing from spermatogenesis in mouse (Cao et al 2021), human (Sohni et al 2019), and zebrafish (Qian et al 2022).

Along differentiation trajectories in these tissues, successive clusters are defined by their aggregate, transcript repertoire. Indeed, differentially-expressed genes can be identified for clusters, with expression enriched in a given cluster. However, expression is rarely restricted to a cluster. For instance, Cao et al. subcluster spermatogonia into four subgroups, termed SPG1-4. They state clearly that these SPG1-4 “follow a continuous differentiation trajectory,” as can be inferred by marker expression across cells in this lineage. Similar to our findings, while the spermatogonia can fall into discrete clusters, gene expression patterns are contiguous. For example, the “undifferentiated” marker used in Cao et al, Crabp1, clearly shows expression in SPG1-3, annotated as spermatogonial stem cells, undifferentiated spermatogonia, and early differentiated spermatogonia, respectively. Likewise, markers for the “SPG3” state spermatogonia have detectable expression in SPG2 and SPG4, and likewise for markers of the “SPG4” state (with expression found also in SPG3). <br /> Analogous study of human spermatogenesis arrives at a similar conclusion. In that work, although clusters are named as “spermatogonial stem cell (SSC)”, the authors are careful to specifically point out that, “…while we refer to the SSC-1 and SSC-2 cell clusters as ‘‘SSCs,’’ scRNA-seq is not a functional assay and thus we do not know the percentage of cells in these clusters with SSC activity. These subsets almost certainly contain other A-SPG cells [A type spermatogonia], including SPG progenitors that have committed to differentiate.” (Sohi et al 2019)

Thus, the work in several disparate systems, all involving renewing lineages, finds that discrete clusters, such as a “stem cell cluster” are not identified. In the Drosophila testis, germline differentiation flows in a continuous-like manner similar to spermatogenesis in several other organisms studied by scRNA-seq, and our finding is not a function of the methodology, but rather a facet of the biology of the organ.

Operating in parallel with continuous differentiation, we did find evidence of, and extensively discussed in concert with Figure 4, huge and dramatic shifts in transcriptional state in spermatocytes compared to spermatogonia, in early spermatids compared to spermatocytes, and in late spermatid elongation. Lastly, as we describe further below, new data in this resubmission identify four distinct genes with stage-selective expression as predicted by our analysis (new Figure 2 - figure supplement 1), illustrating the utility of our study for the field to find new markers and new genes to test for function.

A goal of the study was to identify new rare cell types, and the hub, a small apical somatic cell region, was mentioned as a target region, since it regulates both stem cell populations, GSCs and CySCs, is capable of regeneration, and other fascinating properties. However the analysis of the hub cluster revealed more problems of specificity. 41 or 120 cells in the cluster were discordant with the remaining 79 which did express markers consistent with previous studies. Why these cells co-clustered was not explained and one can only presume that similar problems may be found in other clusters.

Our writing seems not to have been clear enough on this point and we thank the reviewer. We have revised the section. In addition, we have added new data (Figure 7 - figure supplement 2). We had already stated that only 79 of these 120 nuclei were near to each other in 2D UMAP space, while other members of original cluster 90 were dispersed. Thus the 79 hub nuclei in fact clustered together on the UMAP. Other nuclei that mapped at dispersed positions were initially ‘called’ as part of this cluster in the original Fly Cell Atlas (FCA) paper (Li et al., 2022), making it obvious that a correction to that assignment was necessary, which we carried out. To our eye, no other called cluster was represented by such dispersed groupings. For the hub, we definitively established the 79 nuclei to represent hub cells by marker gene analysis, including the identification of a new maker, tup, that was included in the 79 annotated hub nuclei but excluded from the 41 other nuclei (Figure 7). In this resubmission, to independently verify the relationship of the 79 nuclei to each other, we subjected the 120 nuclei from the original cluster 90 defined by the FCA study to hierarchical clustering using only genes that are highly expressed and variable in these nuclei (Figure 7 - figure supplement 2). This computationally distinct approach strongly supported our identification of the 79 definitive hub nuclei.

Indeed, many other indications of specificity issues were described, including contamination of fat body with spermatocytes, the expression of germline genes such as Vasa in many somatic cell clusters like muscle, hemocytes, and male gonad epithelium, and the promiscuous expression of many genes, including 25% of somatic-specific transcription factors, in mid to late spermatocytes. The expression of only one such genes, Hml, was documented in tissue, and the authors for reasons not explained did not attempt to decisively address whether this phenomenon is biologically meaningful.

We discussed the question of vasa expression in somatic clusters in some detail above, in response to referee #1, and included new analysis in the resubmission.

With respect to the observation of ‘somatic gene’ expression in spermatocytes, we are also intrigued. We do not believe this is due to “contamination,” but rather a spermatocyte expression program that includes expression of somatic genes. First, these somatic markers were not observed in other germline clusters, which would be expected if this was due to general transcript contamination. Second, we observed expression of somatic markers in spermatocytes independently in the single-cell and single-nucleus data, making it unlikely to be an artifact of preparation of isolated nuclei. Finally, in the resubmission, in addition to Hml, we validated ‘somatic’ marker expression in spermatocytes by FISH of a somatic, tail cyst cell marker, Vsx1. Vsx1 is predicted to be expressed at low levels in spermatocytes in our dataset and is clearly visible in germline cells by FISH (Figure 3 – figure supplement 2G,H). We also refer the referee to Figure 6K, where the mRNA for the somatic cyst cell marker eya was observed by FISH at low levels in spermatocytes.

A truly interesting question mentioned by the authors is why the testis consistently ranks near the top of all tissues in the complexity of its gene expression. In the Li et al. (2022) paper it was suggested that this is due an inherently greater biological complexity of spermiogenesis than other tissues. It seems difficult to independently and rationally determine "biological complexity," but if a conserved characteristic of testis was to promiscuously express a wide range of (random?) genes, something not out of the question, this would be highly relevant and important.

We agree that the massive transcriptional program found in spermatocytes is, indeed, truly interesting. There are many speculations as to why spermatocytes are so highly transcriptional, including the possibility of “transcriptional scanning” (e.g., Xia et al. 2020) regulating the evolution of new genes. Testing such models is beyond the scope of this paper. However, one must also keep in mind that spermatogenesis involves one of the most dramatic cellular transformations in biology, where cellular components spanning from nuclei to chromatin to Golgi, cell cycle, extensive membrane addition, changes in cell shape, and building of a complex swimming organelle all must occur and be temporally coordinated. Small wonder that many genes must be expressed to accomplish these tasks.

Unfortunately, the most likely problems are simply technical. Drosophila cells are small and difficult to separate as intact cells. The use of nuclei was meant to overcome this inherent problem, but the effectiveness of this new approach is not yet well-documented. Support for the view that the problems are mostly technical, rather than a reflection of testis biology, comes from studies of scRNAseq in the mouse, where it has been possible to resolve a stem cell cluster, and germ cell pathways that follow known germ cell differentiation trajectories with much more discrete steps than were reported here (for example, Cao et al. 2021 cited by the authors).

We respectfully disagree with the referee about this collection of statements. First, the use of snRNASeq has been extensively characterized and compared to scRNA-seq in brain tissue by McLaughlin et al., 2021 (cited in the original submission) and was shown to be effective (McLaughlin, et al. eLife 2021;10:e63856. DOI: https://doi.org/10.7554/eLife.63856). snRNA-seq has a distinct advantage when dealing with long, thin cells, such as neurons or cyst cells (as featured in this work), where cytoplasm can easily be sheared off during cell isolation. Second, in a previous portion of our response to this referee, we discussed how our interpretation of Cao et al., 2021 differs from that expressed by this referee. Lastly, as requested in ‘Essential revision’ 2, we adjusted clustering methods and selected four genes, two predicted to be markers for early stage germline cells, and two for mid-spermatocyte stage development. FISH analysis demonstrates that expression for each of these maps to the appropriate stages (new Figure 2 - figure supplement 1). This confirms that the datasets we present in this manuscript can be mined to identify unique, diagnostic markers for various stages.

The conclusions that were made by the authors seem to either be facts that are already well known, such as the problem that transcriptional changes in spermatocytes will be obscured by the large stored mRNA pool, or promises of future utility. For example, "mining the snRNA-seq data for changes in gene expression as one cluster advances to the next should identify new sub-stage-specific markers." If worthwhile new markers could be identified from these data, surely this could have been accomplished and presented in a supplemental Table. As it currently stands, the manuscript presents the dataset including a fair description of its current limitations, but very little else of novel biological interest is to be found.

“In sum, this project represents an extremely worthwhile undertaking that will eventually pay off. However, some currently unappreciated technical issues, in cell/nuclear isolation, and certainly in the bioinformatic programs and procedures used that mis-clustered many different cells, has created the current difficulties.

Most scRNAseq software is written to meet the needs of mammalian researchers working with cultured cells, cellular giants compared to Drosophila and of generally similar size. Such software may not be ideal for much smaller cells, but which also include the much wider variation in cell size, properties and biological mechanisms that exist in the world of tissues.”

We appreciate the referee’s acknowledgement that this ‘undertaking will eventually pay off’. It was not our intention to address ‘function’ for this study, but rather to make the system accessible to the broadest community possible. We are uncertain if there is any remaining reservation held by this referee. A brief summary of what we covered in the manuscript may help allay any residual concern. Obviously, study of the Drosophila testis and spermatogenesis benefits from the knowledge of a large number of established cell-type and stage-selective markers. Thus, we extensively used the community’s accepted markers to assign identity to clusters in both the sn- and sc-RNA-seq UMAPs. We believe that effort well establishes the validity and reliability of the dataset . Furthermore, we identified upwards of a dozen new markers out of the cluster analysis, and verified their expression by FISH or reporter line in various figures throughout (tup, amph, piwi, geko, Nep4, CG3902, Akr1B, loqs, Vsx1, Drep2, Pxt, CG43317, Vha16-5, l(2)41Ab). To our mind, these contributions, coupled with annotation of the datasets, suggest strongly that they will serve the community well. This is especially true as we provide users with objects that they can feed into commonly used software algorithms such as Seurat and Monocle to explore the datasets to their purposes. Rather than simply relying on default settings within some of the applications, we also adjusted parameters for various clusterings as called for; some of which were in response to astute comments from referees, and included in the resubmission. Of course, it is possible that rare issues may arise in the datasets as these are further studied, but that is the case with all scRNA-seq data, and is not specific to work on this model organism.

Reviewer #3 (Public Review):

In this study, the authors use recently published single nucleus RNA sequencing data and a newly generated single cell RNA sequencing dataset to determine the transcriptional profiles of the different cell types in the Drosophila ovary. Their analysis of the data and experimental validation of key findings provide new insight into testis biology and create a resource for the community. The manuscript is clearly written, the data provide strong support for the conclusions, and the analysis is rigorous. Indeed, this manuscript serves as a case study demonstrating best practices in the analysis of this type of genomics data and the many types of predictions that can be made from a deep dive into the data. Researchers who are studying the testis will find many starting points for new projects suggested by this work, and the insightful comparison of methods, such as between slingshot and Monocle3 and single cell vs single nucleus sequencing will be of interest beyond the study of the Drosophila testis.

We greatly appreciate the reviewer’s comments.

Reviewer #4 (Public Review):

This is an extraordinary study that will serve as key resource for all researchers in the field of Drosophila testis development. The lineages that derive from the germline stem cells and somatic stem cells are described in a detail that has not been previously achieved. The RNAseq approaches have permitted the description of cell states that have not been inferred from morphological analyses, although it is the combination of RNAseq and morphological studies that makes this study exceptional. The field will now have a good understanding of interactions between specific cell states in the somatic lineage with specific states in the germ cell lineage. This resource will permit future studies on precise mechanisms of communication between these lineages during the differentiation process, and will serve as a model for studies of co-differentiation in other stem cell systems. The combination of snRNAseq and scRNAseq has conclusively shown differences in transcriptional activation and RNA storage at specific stages of germ cell differentiation and is a unique study that will inform other studies of cell differentiation.

Could the authors please describe whether genes on the Y chromosome are expressed outside of the male germline. For example, what is represented by the spots of expression within the seminal vesicle observed in Figure 3D?

Prior work demonstrated that proteins encoded by Y-linked genes are not expressed outside of the germline (Zhang et al. Genetics 2020. https://doi.org/10.1534/genetics.120.303324). In our snRNAseq dataset, we find that genes on the Y chromosome are not highly expressed outside of the male germline (on the order of ~100-fold lower in other tissues). In fact, we observe Y chromosome transcripts at this level in many nuclei across tissues collected for the Fly Cell Atlas project, including the ovary. Since we have not followed up on the Fly Cell Atlas observations directly using FISH to examine Y chromosome transcript expression outside the germline, we cannot rule out the possibility that such low level expression is real. However, the detection across several tissues argues that this is likely technical artifact. With regard to ‘spots of expression within the seminal vesicle’ (Figure 3D), a spot is colored red if the average expression level of genes on the Y chromosome is greater in that cell than in an average cell on our plot. These red spots are likely due to ambient RNA being carried over.

I would appreciate some discussion of the "somatic factors" that are observed to be upregulated in spermatocytes (e.g. Mhc, Hml, grh, Syt1). Is there any indication of functional significance of any of these factors in spermatocytes?

This is an excellent question. Although we validated expression for several (Hml, Vsx1 and eya), we did not test for their function here and this issue remains to be studied. This is now directly stated in the main text.

In the discussion of cyst cell lineage differentiation following cluster 74 the authors state that neither the HCC or TCC lineages were enriched for eya (Figure 6V). It seems in this panel that cluster 57 shows some enrichment for eya - is this regarded as too low expression to be considered enriched?

We thank the reviewer for their insightful comment and we agree with their conclusions. We have modified the text to reflect the low, but present, expression of eya in the HCC and TCC lineages. The text now reads as follows at line (insert line # here): “Enrichment of eya was dramatically reduced in the clusters along either late cyst cell branch compared to those of earlier lineage nuclei (Figure 6J,U).”

Author Response

Reviewer #1 (Public Review):

1) Comment: To determine the effect of diseased monocytes on retinal health, light-injured mouse retinas were injected with monocytes isolated from AMD patients (Figure 1 - figure supplement 1). This resulted in a reduction in photoreceptor number and ERG b-wave amplitude. However, the light-injured control eye was injected with PBS only, so no cells were present. The reasoning for using this control was not provided. The appropriate injection control would include monocytes isolated from non-AMD patients. This control should be performed side-by-side with cells from AMD patients.

We thank the reviewer for this important comment. The purpose of the current study was to identify the macrophage subtype that may be associated with cell death in aAMD. We have previously reported that macrophages from AMD patient demonstrate a different phenotype compared with healthy patient in the rodent model for laser induced CNV (Hagbi-Levi S et al, 2016). Per the reviewer comment, we have performed additional experiments to assess the effect of monocytes from healthy controls in the photic retinal injury model. Results showed that monocytes from AMD and healthy patients exert different impact on the retina in this rodent model for aAMD. Interestingly, we found that monocytes from healthy patients were more neurotoxic to photoreceptors compared with monocytes from AMD patients. These results are included in the revised ms. as Figure 1- figure supplement 1H. A possible explanation for these findings is discussed in lines 179-190 of the revised manuscript. This finding reinforces the idea that the use of monocytes from AMD patients in the experiments is required to obtain a comprehensive understanding of their involvement in the progression of the disease.

2) Comment: The authors hypothesize, from the experiments presented in Figure 1 - figure supplement 1, that the injected monocytes generated macrophages in the retina, which were responsible for the observed neurotoxicity (Lines 143-145). However, no direct evidence was presented. This idea should be tested in vivo. This could be done by injecting tracer-labeled human AMD-derived monocytes into light-injured mouse retinas. If the authors' hypothesis is true, collected retinas should contain tracer-labeled cells that express macrophage markers. Tracer-labeled M2a macrophage cells should be present since subsequent experiments identify this subclass as being associated with retinal cell death.

Thank you for this important comment. To address the reviewers comment, retinal section from mice exposed to photic-retinal injury and injected with Dio-tracer labelled monocytes were stained with two M2a macrophages markers, CD206 (mannose receptor) and VEGF (Kadomoto, S et al, 2022; Jayasingam SD et al, 2019). Interestingly, we found co-localization of Dio-tracer staining (representing the injected human macrophages) with CD206 and VEGF markers in monocytes localized in different retinal layers, but not in monocytes remaining in the vitreous cavity. These data indicate that M2a markers are expressed during the polarization of monocytes into M2a phenotype which is maintained only upon entry into the retina tissue. These results were included in Figure 1- figure supplement 1K-S and discussed in the revised manuscript in lines 179-182.

3) Comment: Photoreceptor number and b-wave amplitudes were measured in light-injured retinas injected with one of four macrophage cell types generated from human AMD-derived monocytes. The authors conclude that only injection of M2a cells reduced photoreceptor number and b-wave amplitudes (Figure 1C, E). This may be true, but it is difficult for the reader to make a conclusion (especially in Fig. 1E) due to the large error bars and five different traces overlapping each other. To make these results easier to interpret, graph control cells with only one experimental sample (cell type) at a time.

Thank you for this comment. Per the reviewer comment, the graphs were modified in the revised ms. (Figure 1, panel H-K).

4) Comment: Most injected macrophages were located in the vitreous. In the case of M2a cells, the authors note that "several of the cells migrated across the retinal layers reaching the subretinal space" (Lines 167,168). One possible explanation for why M0, M1, and M2c macrophages did not induce retinal degeneration is that they did not migrate to the subretinal space and around the optic nerve head. Supplementary figures should be added to demonstrate that this is not the case.

Thank you for this comment. To address the reviewer comment we compared the migration patterns of the different macrophage phenotypes following intravitreal injection in mice exposed to photic-injury. Our results indicated that M0, M1 and M2c macrophages, similarly to M2a macrophages, migrated to the subretinal space and around the optic nerve. Thus, the neurotoxic effect of M2a is not explained by their capacity to infiltrate the retinal tissues. These results was included in Figure 1- figure supplement 2 E-H of the revised manuscript. These results are supported by our ex-vivo experiments, showing that co-culture of M2a macrophages with a retinal explants was associated with increased photoreceptor cells death compared to M1 macrophages. The results are presented and discussed in the revised manuscript in lines 200-203.

5) Comment: Figure 1 - figure supplement 2: Panel A, B cells were stained with CD206 to demonstrate the presence of M2a macrophages (panel B). The authors conclude that panel A contains M1 and panel B contains M2a cells. The lack of CD206 expression illustrates that panel A cells are not M2a macrophages but do not demonstrate they are M1 macrophages. A control using an M1 cell marker is necessary to show that panel A cells are M1 and M1 cells are not detected in M2a cultures.

Thank you for this comment. We have validated the phenotype of each macrophages subtype by qPCR (Figure 1 panel A). To further address the reviewer comment, we have performed additional immunocytochemistry for M1 macrophages using anti-CD80 antibody which is utilized as M1 macrophages marker (Bertani FR et al.2017). Results of the staining confirmed the identity of the M1 macrophages. These new results were included in Figure 1- figure supplement 2A, and are discussed in lines 168-170.

6) Comment: Ex vivo, apoptotic photoreceptor and RPE cells are observed when cultured with M2a macrophages (Figure 2). Do injected M2a cells also induce apoptosis of RPE cells in vivo? This is important to establish that retinal explants are a good model for in vivo experiments.

Thank you for this comment. To address the reviewer comment, we assessed RPE apoptosis (using TUNEL, Caspase 3 staining and RPE65 marker) after M2A cells delivery, in the in-vivo photic injury model. We could not detect apoptotic signal in the RPE layers 7 days after photic injury and therefore could not evaluate the effect of M2a macrophages on the RPE cells in-vivo (see Author response image 1). One possible explanation is that RPE cells that have undergone apoptosis are rapidly removed from the damaged tissue and are no longer detectable unlike photoreceptors. Furthermore, a study that investigated the impact of bright light on RPE cells in-vivo, showed that although RPE cells undergone structural and chemical modifications after photic-injury, TUNEL signal was not detected because RPE cell die by necrosis mechanism and not apoptosis (Jaadane I et al, 2017). Other studies validated that blue light induces RPE necrosis (Song W et al, 2022; Mohamed A et al, 2022). Taken together, it seems that ex-vivo retinal explant and in-vivo photic injury both simulate the mechanism of retinal cell death. However, the use of ex-vivo model allows for establishing the direct impact of M2a macrophages on retina in non-inflammatory context.

Author responnse image 1.

7) Comment: Reactive oxygen species (ROS) production was measured to determine if M2a cell-mediated neurotoxicity was due to oxidative stress. It is concluded that a ROS increase is partly responsible (Line 218). The data do not support this conclusion. ROS was detected in cultured M2a macrophages. More importantly, however, there was no increase in oxidative damage in vivo. The in vivo and cell culture results contradict each other so no conclusion can be made. The lack of in vivo confirmation weakens the argument that ROS drives M2a neurotoxicity. Text suggesting a role for ROS in neurotoxicity should be appropriately edited (Lines including 218, 244, 401,406,481).

Thank you for this comment. The manuscript was revised according to the reviewer suggestion (Lines 250-256).

8) Comment: The authors ask if the photoreceptor cell death is cytokine-mediated. Multiple cytokines were enriched in M2a-conditioned media. Of particular interest were CCR1 ligands MPIF1 and MCP4. The implication is that these two ligands mediate the M2a macrophages to photoreceptor cell death through CCR1. However, there is no attempt to show that either MPIF1 or MCP4 are present in vivo, or are sufficient to induce the retinal response observed. This could be demonstrated by injection of MPIF1 or MCP4. Evidence that either ligand phenocopies M2a macrophage injection would be direct evidence that CCR1 ligands activate the retinal response. Furthermore, co-injection with BX174 should block the effect of these ligands if they work through CCR1.

Thank you for this comment. The identification of CCR1 ligands expression from M2a polarized macrophages directed our decision to study CCR1 in the context of atrophic AMD. We do not claim that these specific CCR1 ligands are sufficient to activate CCR1 and exert retinal injury. The mechanism is likely more complex. Yet, to address the reviewer comment, we have performed the experiments suggested by the reviewer. Mice were exposed to photic injury and immediately injected in one eye with MPIF1, MCP-4, or a combination of both and in second eye with PBS as vehicle. Intravitreal cytokines delivery was repeated two days later (following the half-life time of these cytokines) and ERG were recorded two days after the last injection. Injection of cytokines at a concentration of 300 ng per eye did not exacerbated photoreceptor death. Then, the same experiment was repeated with two higher concentrations of cytokine, 1.2 ug/eye and 2 ug/eye, but no changes are observed between the cytokines treated-eyes and the vehicle treated-eyes. Based on previous studies reporting the physiological concentration of different cytokines in eyes of un/healthy individuals and on experiments in which different cytokines are injected in rodent eye (Estevao C et al, 2021. Zeng Y et al, 2019; Roybal CN et al, 2018; Mugisho OO et al, 2018), the cytokine concentrations used in our experiment are in the range in which effect on the retina is expected.

It is likely that a synergistic effect of M2a-secreted proteins in a particular microenvironment is necessary to increase the level of retinal damage (Bartee E et al, 2013). It is also likely that in the photic retinal injury model there is upregulation of cytokines that may mask additional delivery of exogenous cytokines. Comprehensive understanding of the complex interactions of these cytokines during retinal degeneration is beyond the scope of the current manuscript which is not focus on identifying ligand-induced CCR1 activation and its consequences. Additionally, we suggest that due to cytokine redundancy (Nicola NA; 1994), demonstrating that MPIF-4 or MCP-3 can increase photoreceptor death is not required for proving CCR1 receptor involvement.

Author Response

Reviewer #1 (Public Review):

In this work George et al. describe RatInABox, a software system for generating surrogate locomotion trajectories and neural data to simulate the effects of a rodent moving about an arena. This work is aimed at researchers that study rodent navigation and its neural machinery.

Strengths:

• The software contains several helpful features. It has the ability to import existing movement traces and interpolate data with lower sampling rates. It allows varying the degree to which rodents stay near the walls of the arena. It appears to be able to simulate place cells, grid cells, and some other features.

• The architecture seems fine and the code is in a language that will be accessible to many labs.

• There is convincing validation of velocity statistics. There are examples shown of position data, which seem to generally match between data and simulation.

Weaknesses:

• There is little analysis of position statistics. I am not sure this is needed, but the software might end up more powerful and the paper higher impact if some position analysis was done. Based on the traces shown, it seems possible that some additional parameters might be needed to simulate position/occupancy traces whose statistics match the data.

Thank you for this suggestion. We have added a new panel to figure 2 showing a histogram of the time the agent spends at positions of increasing distance from the nearest wall. As you can see, RatInABox is a good fit to the real locomotion data: positions very near the wall are under-explored (in the real data this is probably because whiskers and physical body size block positions very close to the wall) and positions just away from but close to the wall are slightly over explored (an effect known as thigmotaxis, already discussed in the manuscript).

As you correctly suspected, fitting this warranted a new parameter which controls the strength of the wall repulsion, we call this “wall_repel_strength”. The motion model hasn’t mathematically changed, all we did was take a parameter which was originally a fixed constant 1, unavailable to the user, and made it a variable which can be changed (see methods section 6.1.3 for maths). The curves fit best when wall_repel_strength ~= 2. Methods and parameters table have been updated accordingly. See Fig. 2e.

• The overall impact of this work is somewhat limited. It is not completely clear how many labs might use this, or have a need for it. The introduction could have provided more specificity about examples of past work that would have been better done with this tool.

At the point of publication we, like yourself, also didn’t know to what extent there would be a market for this toolkit however we were pleased to find that there was. In its initial 11 months RatInABox has accumulated a growing, global user base, over 120 stars on Github and north of 17,000 downloads through PyPI. We have accumulated a list of testimonials[5] from users of the package vouching for its utility and ease of use, four of which are abridged below. These testimonials come from a diverse group of 9 researchers spanning 6 countries across 4 continents and varying career stages from pre-doctoral researchers with little computational exposure to tenured PIs. Finally, not only does the community use RatInABox they are also building it: at the time of writing RatInABx has received logged 20 GitHub “Issues” and 28 “pull requests” from external users (i.e. those who aren’t authors on this manuscript) ranging from small discussions and bug-fixes to significant new features, demos and wrappers.

Abridged testimonials:

● “As a medical graduate from Pakistan with little computational background…I found RatInABox to be a great learning and teaching tool, particularly for those who are underprivileged and new to computational neuroscience.” - Muhammad Kaleem, King Edward Medical University, Pakistan

● “RatInABox has been critical to the progress of my postdoctoral work. I believe it has the strong potential to become a cornerstone tool for realistic behavioural and neuronal modelling” - Dr. Colleen Gillon, Imperial College London, UK

● “As a student studying mathematics at the University of Ghana, I would recommend RatInABox to anyone looking to learn or teach concepts in computational neuroscience.” - Kojo Nketia, University of Ghana, Ghana

● “RatInABox has established a new foundation and common space for advances in cognitive mapping research.” - Dr. Quinn Lee, McGill, Canada

The introduction continues to include the following sentence highlighting examples of past work which relied of generating artificial movement and/or neural dat and which, by implication could have been done better (or at least accelerated and standardised) using our toolbox.

“Indeed, many past[13, 14, 15] and recent[16, 17, 18, 19, 6, 20, 21] models have relied on artificially generated movement trajectories and neural data.”

• Presentation: Some discussion of case studies in Introduction might address the above point on impact. It would be useful to have more discussion of how general the software is, and why the current feature set was chosen. For example, how well does RatInABox deal with environments of arbitrary shape? T-mazes? It might help illustrate the tool's generality to move some of the examples in supplementary figure to main text - or just summarize them in a main text figure/panel.

Thank you for this question. Since the initial submission of this manuscript RatInABox has been upgraded and environments have become substantially more “general”. Environments can now be of arbitrary shape (including T-mazes), boundaries can be curved, they can contain holes and can also contain objects (0-dimensional points which act as visual cues). A few examples are showcased in the updated figure 1 panel e.

To further illustrate the tools generality beyond the structure of the environment we continue to summarise the reinforcement learning example (Fig. 3e) and neural decoding example in section 3.1. In addition to this we have added three new panels into figure 3 highlighting new features which, we hope you will agree, make RatInABox significantly more powerful and general and satisfy your suggestion of clarifying utility and generality in the manuscript directly.

On the topic of generality, we wrote the manuscript in such a way as to demonstrate how the rich variety of ways RatInABox can be used without providing an exhaustive list of potential applications. For example, RatInABox can be used to study neural decoding and it can be used to study reinforcement learning but not because it was purpose built with these use-cases in mind. Rather because it contains a set of core tools designed to support spatial navigation and neural representations in general. For this reason we would rather keep the demonstrative examples as supplements and implement your suggestion of further raising attention to the large array of tutorials and demos provided on the GitHub repository by modifying the final paragraph of section 3.1 to read:

“Additional tutorials, not described here but available online, demonstrate how RatInABox can be used to model splitter cells, conjunctive grid cells, biologically plausible path integration, successor features, deep actor-critic RL, whisker cells and more. Despite including these examples we stress that they are not exhaustive. RatInABox provides the framework and primitive classes/functions from which highly advanced simulations such as these can be built.”

Reviewer #3 (Public Review):

George et al. present a convincing new Python toolbox that allows researchers to generate synthetic behavior and neural data specifically focusing on hippocampal functional cell types (place cells, grid cells, boundary vector cells, head direction cells). This is highly useful for theory-driven research where synthetic benchmarks should be used. Beyond just navigation, it can be highly useful for novel tool development that requires jointly modeling behavior and neural data. The code is well organized and written and it was easy for us to test.

We have a few constructive points that they might want to consider.

• Right now the code only supports X,Y movements, but Z is also critical and opens new questions in 3D coding of space (such as grid cells in bats, etc). Many animals effectively navigate in 2D, as a whole, but they certainly make a large number of 3D head movements, and modeling this will become increasingly important and the authors should consider how to support this.

Agents now have a dedicated head direction variable (before head direction was just assumed to be the normalised velocity vector). By default this just smoothes and normalises the velocity but, in theory, could be accessed and used to model more complex head direction dynamics. This is described in the updated methods section.

In general, we try to tread a careful line. For example we embrace certain aspects of physical and biological realism (e.g. modelling environments as continuous, or fitting motion to real behaviour) and avoid others (such as the biophysics/biochemisty of individual neurons, or the mechanical complexities of joint/muscle modelling). It is hard to decide where to draw but we have a few guiding principles:

1. RatInABox is most well suited for normative modelling and neuroAI-style probing questions at the level of behaviour and representations. We consciously avoid unnecessary complexities that do not directly contribute to these domains.

2. Compute: To best accelerate research we think the package should remain fast and lightweight. Certain features are ignored if computational cost outweighs their benefit.

3. Users: If, and as, users require complexities e.g. 3D head movements, we will consider adding them to the code base.

For now we believe proper 3D motion is out of scope for RatInABox. Calculating motion near walls is already surprisingly complex and to do this in 3D would be challenging. Furthermore all cell classes would need to be rewritten too. This would be a large undertaking probably requiring rewriting the package from scratch, or making a new package RatInABox3D (BatInABox?) altogether, something which we don’t intend to undertake right now. One option, if users really needed 3D trajectory data they could quite straightforwardly simulate a 2D Environment (X,Y) and a 1D Environment (Z) independently. With this method (X,Y) and (Z) motion would be entirely independent which is of unrealistic but, depending on the use case, may well be sufficient.

Alternatively, as you said that many agents effectively navigate in 2D but show complex 3D head and other body movements, RatInABox could interface with and feed data downstream to other softwares (for example Mujoco[11]) which specialise in joint/muscle modelling. This would be a very legitimate use-case for RatInABox.

We’ve flagged all of these assumptions and limitations in a new body of text added to the discussion:

“Our package is not the first to model neural data[37, 38, 39] or spatial behaviour[40, 41], yet it distinguishes itself by integrating these two aspects within a unified, lightweight framework. The modelling approach employed by RatInABox involves certain assumptions:

1. It does not engage in the detailed exploration of biophysical[37, 39] or biochemical[38] aspects of neural modelling, nor does it delve into the mechanical intricacies of joint and muscle modelling[40, 41]. While these elements are crucial in specific scenarios, they demand substantial computational resources and become less pertinent in studies focused on higher-level questions about behaviour and neural representations.

2. A focus of our package is modelling experimental paradigms commonly used to study spatially modulated neural activity and behaviour in rodents. Consequently, environments are currently restricted to being two-dimensional and planar, precluding the exploration of three-dimensional settings. However, in principle, these limitations can be relaxed in the future.

3. RatInABox avoids the oversimplifications commonly found in discrete modelling, predominant in reinforcement learning[22, 23], which we believe impede its relevance to neuroscience.

4. Currently, inputs from different sensory modalities, such as vision or olfaction, are not explicitly considered. Instead, sensory input is represented implicitly through efficient allocentric or egocentric representations. If necessary, one could use the RatInABox API in conjunction with a third-party computer graphics engine to circumvent this limitation.

5. Finally, focus has been given to generating synthetic data from steady-state systems. Hence, by default, agents and neurons do not explicitly include learning, plasticity or adaptation. Nevertheless we have shown that a minimal set of features such as parameterised function-approximator neurons and policy control enable a variety of experience-driven changes in behaviour the cell responses[42, 43] to be modelled within the framework.

• What about other environments that are not "Boxes" as in the name - can the environment only be a Box, what about a circular environment? Or Bat flight? This also has implications for the velocity of the agent, etc. What are the parameters for the motion model to simulate a bat, which likely has a higher velocity than a rat?

Thank you for this question. Since the initial submission of this manuscript RatInABox has been upgraded and environments have become substantially more “general”. Environments can now be of arbitrary shape (including circular), boundaries can be curved, they can contain holes and can also contain objects (0-dimensional points which act as visual cues). A few examples are showcased in the updated figure 1 panel e.

Whilst we don’t know the exact parameters for bat flight users could fairly straightforwardly figure these out themselves and set them using the motion parameters as shown in the table below. We would guess that bats have a higher average speed (speed_mean) and a longer decoherence time due to increased inertia (speed_coherence_time), so the following code might roughly simulate a bat flying around in a 10 x 10 m environment. Author response image 1 shows all Agent parameters which can be set to vary the random motion model.

Author response image 1.

• Semi-related, the name suggests limitations: why Rat? Why not Agent? (But its a personal choice)

We came up with the name “RatInABox” when we developed this software to study hippocampal representations of an artificial rat moving around a closed 2D world (a box). We also fitted the random motion model to open-field exploration data from rats. You’re right that it is not limited to rodents but for better or for worse it’s probably too late for a rebrand!

• A future extension (or now) could be the ability to interface with common trajectory estimation tools; for example, taking in the (X, Y, (Z), time) outputs of animal pose estimation tools (like DeepLabCut or such) would also allow experimentalists to generate neural synthetic data from other sources of real-behavior.

This is actually already possible via our “Agent.import_trajectory()” method. Users can pass an array of time stamps and an array of positions into the Agent class which will be loaded and smoothly interpolated along as shown here in Fig. 3a or demonstrated in these two new papers[9,10] who used RatInABox by loading in behavioural trajectories.

• What if a place cell is not encoding place but is influenced by reward or encodes a more abstract concept? Should a PlaceCell class inherit from an AbstractPlaceCell class, which could be used for encoding more conceptual spaces? How could their tool support this?

In fact PlaceCells already inherit from a more abstract class (Neurons) which contains basic infrastructure for initialisation, saving data, and plotting data etc. We prefer the solution that users can write their own cell classes which inherit from Neurons (or PlaceCells if they wish). Then, users need only write a new get_state() method which can be as simple or as complicated as they like. Here are two examples we’ve already made which can be found on the GitHub:

Author response image 2.

Phase precession: PhasePrecessingPlaceCells(PlaceCells)[12] inherit from PlaceCells and modulate their firing rate by multiplying it by a phase dependent factor causing them to “phase precess”.

Splitter cells: Perhaps users wish to model PlaceCells that are modulated by recent history of the Agent, for example which arm of a figure-8 maze it just came down. This is observed in hippocampal “splitter cell”. In this demo[1] SplitterCells(PlaceCells) inherit from PlaceCells and modulate their firing rate according to which arm was last travelled along.

• This a bit odd in the Discussion: "If there is a small contribution you would like to make, please open a pull request. If there is a larger contribution you are considering, please contact the corresponding author3" This should be left to the repo contribution guide, which ideally shows people how to contribute and your expectations (code formatting guide, how to use git, etc). Also this can be very off-putting to new contributors: what is small? What is big? we suggest use more inclusive language.

We’ve removed this line and left it to the GitHub repository to describe how contributions can be made.

• Could you expand on the run time for BoundaryVectorCells, namely, for how long of an exploration period? We found it was on the order of 1 min to simulate 30 min of exploration (which is of course fast, but mentioning relative times would be useful).

Absolutely. How long it takes to simulate BoundaryVectorCells will depend on the discretisation timestep and how many neurons you simulate. Assuming you used the default values (dt = 0.1, n = 10) then the motion model should dominate compute time. This is evident from our analysis in Figure 3f which shows that the update time for n = 100 BVCs is on par with the update time for the random motion model, therefore for only n = 10 BVCs, the motion model should dominate compute time.

So how long should this take? Fig. 3f shows the motion model takes ~10-3 s per update. One hour of simulation equals this will be 3600/dt = 36,000 updates, which would therefore take about 72,000*10-3 s = 36 seconds. So your estimate of 1 minute seems to be in the right ballpark and consistent with the data we show in the paper.

Interestingly this corroborates the results in a new inset panel where we calculated the total time for cell and motion model updates for a PlaceCell population of increasing size (from n = 10 to 1,000,000 cells). It shows that the motion model dominates compute time up to approximately n = 1000 PlaceCells (for BoundaryVectorCells it’s probably closer to n = 100) beyond which cell updates dominate and the time scales linearly.

These are useful and non-trivial insights as they tell us that the RatInABox neuron models are quite efficient relative to the RatInABox random motion model (something we hope to optimise further down the line). We’ve added the following sentence to the results:

“Our testing (Fig. 3f, inset) reveals that the combined time for updating the motion model and a population of PlaceCells scales sublinearly O(1) for small populations n > 1000 where updating the random motion model dominates compute time, and linearly for large populations n > 1000. PlaceCells, BoundaryVectorCells and the Agent motion model update times will be additionally affected by the number of walls/barriers in the Environment. 1D simulations are significantly quicker than 2D simulations due to the reduced computational load of the 1D geometry.”

And this sentence to section 2:

“RatInABox is fundamentally continuous in space and time. Position and velocity are never discretised but are instead stored as continuous values and used to determine cell activity online, as exploration occurs. This differs from other models which are either discrete (e.g. “gridworld” or Markov decision processes) or approximate continuous rate maps using a cached list of rates precalculated on a discretised grid of locations. Modelling time and space continuously more accurately reflects real-world physics, making simulations smooth and amenable to fast or dynamic neural processes which are not well accommodated by discretised motion simulators. Despite this, RatInABox is still fast; to simulate 100 PlaceCell for 10 minutes of random 2D motion (dt = 0.1 s) it takes about 2 seconds on a consumer grade CPU laptop (or 7 seconds for BoundaryVectorCells).”

• Regarding the Geometry and Boundary conditions, would supporting hyperbolic distance might be useful, given the interest in alternative geometry of representations (ie, https://www.nature.com/articles/s41593-022-01212-4)?

Whilst this would be very interesting it would likely represent quite a significant edit, requiring rewriting of almost all the geometry-handling code. We’re happy to consider changes like these according to (i) how simple they will be to implement, (ii) how disruptive they will be to the existing API, (iii) how many users would benefit from the change. If many users of the package request this we will consider ways to support it.

• In general, the set of default parameters might want to be included in the main text (vs in the supplement).

We also considered this but decided to leave them in the methods for now. The exact value of these parameters are subject to change in future versions of the software. Also, we’d prefer for the main text to provide a low-detail high-level description of the software and the methods to provide a place for keen readers to dive into the mathematical and coding specifics.

• It still says you can only simulate 4 velocity or head directions, which might be limiting.

Thanks for catching this. This constraint has been relaxed. Users can now simulate an arbitrary number of head direction cells with arbitrary tuning directions and tuning widths. The methods have been adjusted to reflect this (see section 6.3.4).

• The code license should be mentioned in the Methods.

We have added the following section to the methods:

6.6 License RatInABox is currently distributed under an MIT License, meaning users are permitted to use, copy, modify, merge publish, distribute, sublicense and sell copies of the software.

Author Response:

Reviewer #1 (Public Review):

The authors present a system that allows the measurement of OCR on diverse tissues. Using two optopes, one before the tissue under examination, and one after, allows the OCR to be measured as the difference between the concentration of O2 in the in-flow gas and the concentration of O2 in the out-flow gas. The system maintains the tissue at a set concentration of dissolved O2 so that experiments can be performed over a long period of time. The authors have provided ample data and full methods and their conclusions are most likely reliable.

Currently, we know that O2 is critical for diverse physiological processes, however it is rarely as well controlled for as well as non-gas solutes such as glucose, as we lack methods to control its delivery and infer its consumption. By addressing this need, the authors contribute something valuable to the field, which will hopefully be built on by others. The authors have already begun to show the utility of their system by exploring the complicated biology of H2S. As delivering this gas in a controlled manner is hard, often people use NaHS instead. In line with previous studies (well cited by the authors), differences are observed.

Specific points

1) The gas control system is used with islets, INS-1 832/12 cells, retinas, and liver tissue, demonstrating its broad applicability.

2) The system as a platform can have diverse extra measurement modalities attached to it, for example visible-wavelength absorbance and fluorescence. Metabolite concentrations in the tissue culture outflow could also be measured.

3) The reduction state of cyt c and cyt c oxidase are measured from the second derivative of absorbance at 550 and 605 nm. Ideally, to reliably decompose these signals full spectra around 550-605 nm would be collected. As the authors are only using cytochrome reduction state as a qualitative measure and appear careful to avoid over-interpretation this method should be fine. However, the authors ought to show a representative time course including the fully oxidised and reduced states demonstrating this approach as making these measurements is demanding and will depend on the exact spectroscopic set-up. Without this information it is hard to judge the reliability of the paper.

We appreciate giving us the latitude for a less robust measurement. However, we actually did do what you have suggested should be done. That is, with the Ocean Optics spectrophotometer, we measure the full light spectrum from 400 to 650. Using this spectral data, we calculate the first and second derivatives of the absorption. We have previously published our approach to spectral analysis, as well as the inclusion of the fully oxidized and reduced states (Sweet IR, G Khalil, AR Wallen, M Steedman, KA Schenkman, JA Reems, SE Kahn, JB Callis. Continuous measurement of oxygen consumption by pancreatic islets. Diabetes Tech. Ther. 4: 661-672, 2002; Sweet IR, Cook DL, DeJulio E, Wallen AR, Khalil G, Callis JB, Reems JA: Regulation of ATP/ADP in pancreatic islets. Diabetes 53:401-409. 2004), so we did not include all the details. In order to ensure that our description is clear, we have added a more thorough explanation that we used spectral analysis and not just data obtained as single wavelengths.

Reviewer #2 (Public Review):

The present project is an extension of prior work from this work group in which they describe a technological advancement to their published flow-culture system. Such improvements now incorporate technology that allows for metabolic characterization of mammalian tissues while precisely controlling the concentration of abundant gases (e.g., O2), as well as trace gases (e.g., H2S). The present article demonstrates the utility of this system in the context of hypoxia/re-oxygenation experiments, as well as exposure to H2S. Although the methodology described herein is clearly capable of detecting nuanced metabolic changes in response to variations in O2 or H2S, the lack of a head-to-head comparison with other techniques makes it difficult to discern the potential impact of the technology.

We understand the benefit of comparing compare a new method with the currently utilized methods. However, the novelty of our methodology is that it is able to control the exposure of tissue to levels of both abundant and trace dissolved gas composition, functions that neither of these existing instruments provide. In addition, continuous flow of media allows maintenance and assessment of tissue models that cannot be accommodated by static or spinner systems. Since we are the first to report an entirely novel technology, the direct comparison to benchmarks is not possible. In the past, however, we have tested liver slices and retina in a Seahorse and the tissue died within 120 minutes presumably due to the lack of flow/reoxygenation in the tissue. In addition, islets placed in spinner systems such as the Oxygraph become fragmented and broken very rapidly. So, a head to head comparison on the tissue OCR response to changes in gas composition cannot be meaningfully carried out for the facets of our method that we highlighted. The methodology we present has capabilities that do not exist in any other commercially available system. We have stated this latter point in the last line of the second paragraph of the Introduction. Regarding the general reliability of the O2 consumption measurement: the unprecedented accuracy and stability of the O2 detectors and the ability of our flow system to maintain tissue for days while generating accurate and reproducible measurements of O2 consumption has previously been established (Sweet IR, Gilbert M, Sabek O, Fraga DW, Gaber AO, Reems JA. Glucose Stimulation of Cytochrome c Reduction and Oxygen Consumption as Assessment of Human Islet Quality. Transplantation 80: 1003- 1011, 2005; Neal AS, Rountree AM, Philips CW, Kavanagh TJ, Williams DP, Newham P, Khalil G, Cook DL, Sweet IR. Quantification of low-level drug effects using real-time, in vitro measurement of oxygen consumption rate. Toxicological Sciences 148: 594-602, 2015).

In addition, diffusion gradients both in the bath, as well as the tissue itself likely impact the accuracy of the metabolic measurements. This is likely relevant for the liver slices experiments.

We agree that there are certainly concentration gradients within tissue, and these are increased in the absence of capillary flow. Nonetheless, the gradients will certainly be less than what occurs in static systems. In general, optimal size of tissue pieces are a trade-off between potential for hypoxia if the tissue is too large, and a lack of untraumatized tissue if it is too small. We have added text to address this concern that these effects are to be considered when choosing the size and shape of the liver slices or other tissue models to place into the flow system.

Following resection, liver tissue can be mechanically permeabilized (PMID: 12054447). In the present experiments, no controls were put in place to discern if the tissue was permeabilized. This could be checked by adding in adenylates and additional carbon substrates and assessing the impact on OCR. Similar controls likely need to be implemented for the islet and retina experiments.

As we have used flow systems in the past to maintain islets and liver for 24 hours and more (Neal AS, Rountree AM, Kernan K, Van Yserloo B, Zhang H, Reed BJ, Osborne W, Wang W, Sweet IR. Real time imaging of intracellular hydrogen peroxide in pancreatic islets. Biochem. J. 473:4443-4456, 2016; Neal AS, Rountree AM, Philips CW, Kavanagh TJ, Williams DP, Newham P, Khalil G, Cook DL, Sweet IR. Quantification of low-level drug effects using real-time, in vitro measurement of oxygen consumption rate. Toxicological Sciences 148: p. 594-602, 2015) and based on stable OCR we concluded that the tissue is viable. However, it is possible that the membranes of some of the tissue would become permeabilized which would affect the responses to test compounds. We considered this issue from two perspectives. 1. Whether established models that we used to test the BaroFuse were prone to high cell permeability; and 2. Whether loading and maintenance of the tissue models in the fluidics system resulted in increased permeability. We did do experiments measuring the ADP responses in OCR by islets and retina within the fluidics system. Effects were observable but small. However, these results are not definitive, because it was difficult to know what the response in permeabilized tissue was (and permeabilizing tissue slices was difficult). We then used Propidium Iodide staining to visualize and quantify the level of permeability. In islets, the fluorescence in isolated islets before and after perifusion was negligible compared to that in islets permeabilized by H2O2 treatment (see below).

Fig. 1. Staining of isolated rat islets with the indicator of cell membrane integrity propidium iodide. Islets were stained either before or after a 3-hour perifusion. As a positive control for PI staining, islets were treated with 500 uM H2O2 for 30 minutes and incubated overnight. Each data point was the average +/- SE for an n of 3.

There was some fluorescence in retina and liver however, but it was difficult to interpret this data in terms of a fraction of the tissue that is permeabilized due to the fact that dye close to the surface of the tissue is preferentially imaged. So, we finally assessed the amount of permeabilized tissue in retina and liver by comparing uptake of 3H H2O and an extracellular marker C14 sucrose.

Fig. 2. Fraction of tissue water space that is accessible to the extracellular marker sucrose. Left: Mouse retina. Right: Rat liver slice. Each data point was the average +/- SE for an n of 3.

Extracellular water in liver and retina is well established to be about 25%, close to the volume of distribution of sucrose. Thus, we cannot rule out that there are a small percentage of cells that are permeabilized, but the vast majority are not.

Additional comments are detailed below:

-The experiments with H2S are particularly interesting, as this system does seem well suited to investigate the metabolic effects of H2S.

Thanks! We are excited by the potential for this method to assess the effects of H2S and other trace gases.

-The authors state the transient rise in O2 consumption was surprising; however, accumulation of succinate during ischemia and rapid oxidation upon reperfusion has been previously demonstrated (PMID: 32863205).

This is an interesting paper which describes findings that speak to the role of succinate in supplying fuel that could drive the transient changes in O2 consumption observed following hypoxia. It would be an interesting experiment to perform our hypoxia-reoxygenation experiment in the absence and presence of the permeable malonate to see if the spike in O2 consumption following reoxygenation was absent in the presence of the drug. We have removed the word surprising and cited this paper.

-In the paper, Zaprinast was used to block pyruvate uptake. However, the rationale to use this compound, as opposed to the more specific MPC inhibitor UK5099 is unclear.

We could have used UK5099, but we had used Zaprinast in past studies (Du J, Cleghorn WM, Contreras L, Lindsay K, Rountree AM, Chertov AO, Turner SJ, Sahaboglu A, Linton J, Sadilek M, Satrústegui I, Sweet IR, Paquet-Durand F, Hurley JB. Inhibition of mitochondrial pyruvate transport by Zaprinast causes massive accumulation of aspartate at the expense of glutamate in retinas. J Biol. Chem, 288:36129-40, 2013) and so we knew that in our hands that it blocked pyruvate mitochondrial uptake and would therefore be a good test of the rapid transfer of pyruvate across the plasma membrane.

-Throughout the paper, the authors list 'COVID-19' as a potential application. It is not clear how this technology could be used in the context of COVID-19.

Reference to COVID-19 has been removed.

Author Response:

Reviewer #2 (Public Review):

The molecular mechanisms as well as the cellular players of colonization of the adult thymus are incompletely understood. In this manuscript, the authors investigate the role of the SIRPa-CD47 ligand pair in seeding of bone-marrow derived progenitors to the adult murine thymus. The study is based on the authors' earlier characterization of thymic portal endothelial cells, which have a role in mediating progenitor homing to the thymus (Shi et al., 2016). The authors show that loss of SIRPa or CD47 results in reduced frequencies and numbers of early T lineage progenitors (ETPs), but no substantial alterations in thymocyte numbers at later developmental stages and of bone-marrow precursors. Short-term homing assays suggest impaired colonization of the thymus. The authors further characterize cell biology and biochemistry of the SIRPa-CD47 system using peripheral lymphocyte co-cultures with genetically engineered MS1 endothelial cells. Finally, they assess the role of SIRPa-CD47 in thymus regeneration in combination with growth of a model tumor.

Strengths:

The authors describe a clear phenotype, consistent with the moderate effect size in ETP loss upon deletion of other homing mediators, such as PSGL-1 or individual chemokine receptors, such as CCR7, CCR9 or CXCR4.

The authors use multiple genetic models, including both, SIRPa and CD47 deficient mouse strains, to support their findings. Using the Tie2Cre model for endothelial cell-specific deletion is particularly informative and could have been used more extensively. Some data are further strengthened by the complementary use of inhibitory SIRPa-Ig fusion proteins.

In vitro analysis of the molecular mechanism and the role of signaling mediators using MS1 cells is well executed and conclusive.

Weaknesses:

Short-term homing assays suffer from the problem that the system is overwhelmed by an excessive number of donor cells (millions), whereas at steady state only a few hundred HPCs capable of colonizing the thymus circulate in peripheral blood, questioning the physiological relevance of this approach. The short-term nature of the experiments also precludes analysis, whether homed cells do in fact constitute T cell progenitors. More suitable experiments comprise mixed competitive bone marrow chimeras using congenically discernible donor cells or, even better, transfers into non-irradiated recipients of defined age as pioneered by the Goldschneider and Petrie labs. Thus, the conclusion that the SIRPa-CD47 system mediates homing of thymus seeding progenitors is not fully justified.

a) Thank you for the comments. To overcome the disadvantage of total bone marrow transfer, we sorted progenitor-containing lineage- bone marrow cells, which takes about 3% of the total bone marrow cells, by MACS enrichment followed by FACS. The amount of donor cells needed for transfer was therefore reduced from 5×10^7 total bone marrow cells per mouse to less than 1×10^6 lineagecells per mouse. This would prevent the overwhelming effect in the previous method. Result of short-term homing assay with 1×10^6 lineage- bone marrow cells confirmed the homing defect in the thymus of Sirpα^-/- mice (new Figure 2I), but not in the spleen (new Figure2—figure supplement 2J).

b) To track whether immigrated lineage^- progenitors actually develop into thymocytes, we conducted adoptive transfer of congenically marked (CD45.1) WT lineage^- into naïve non-irradiated WT or Sirpα^-/- (CD45.2) recipients. 3 weeks later, donor-derived cell subsets were detected. Significant defect of donor-derived thymocyte development, particularly at DN and DP stages, was found in Sirpα^-/- mice as shown in new Figure 2J,K. Therefore, the defective thymic homing of progenitor cells in Sirpα^-/- mice indeed influence following T cell development.

c) Mixed bone marrow chimera or mixed congenically discernible WT and CD47KO progenitor cell transfer into non-irradiated WT recipients is not applicable as has been explained in details in response to the 2nd point of Summary of Essential Revisions. This is probably due to rapid clearance of CD47-null cells from the system by phagocytosis(Jaiswal et al., 2009). Therefore, it currently remains a technical difficulty to address the role of CD47 on progenitor cells for thymic homing using mixed competitive bone marrow chimeras or mixed progenitor cell transfer in non-irradiated hosts. Instead, we have used cleaner in vitro transwell assay to confirm the role of CD47 on progenitor cells during TEM (new Figure 4F), as explained in more details just below.

While technically elegant and mechanistically conclusive, the in vitro studies using MS1 cells and peripheral lymphocytes are somewhat isolated from the original focus of the paper addressing the role of SIRPa-CD47 specifically in thymus seeding. It should be considered devising similar assays replacing lymphocytes with bone-marrow derived progenitors.

Major in vitro transendothelial migration assays have been repeated with FACS sorted lineage^- bone marrow progenitor cells (Lin^- BMCs). Lin^- BMCs showed significant defect of TEM on Sirpα^-/- ECs compared to that on WT ECs (new Figure 3F); Cd47^-/- Lin^- BMCs also showed significant defect of TEM compared with WT Lin^- BMCs (new Figure 4F). Therefore, the conclusion that progenitor CD47 - endothelial SIRPα signaling is required for TEM remains unchanged.

Analysis of thymus regeneration is interesting, but a number of open questions remain for this experimental setup, also in part raised by the authors in the discussion section. Most notably, during regeneration, the reduction in ETPs is accompanied by reduced numbers in more mature thymocyte subsets and peripheral T cells. Such a reduction was not observed at steady-state in KO models and it cannot be concluded from this experiment, that these observations are caused by a defect in thymus colonization. Notably, SL-TBI is associated with massive cell death and alterations in phagocytosis and many other factors may come into play here as well.

We agree with these comments. CV-1 treatment during SL-TBI induced thymic injury and regeneration is a complicated scenario. To make it cleaner, we did SL-TBI directly on Sirpα^-/- mice and control mice. Congenically marked bone marrow cells were also adoptively transferred for better monitoring. At 4 weeks after transfer, donor derived DN thymocyte subset was found defective in Sirpα^-/- recipients compared to that in control hosts (Figure R1). However, DP, SP subsets did not show difference, probably due to compensation effect.

Figure R1. Reconstitution of bone marrow-derived progenitors in Sirpα^-/- *mice. (A) Schematic view of the experiment. (B,C) Statistics of proportion (B) and cell number (C) of donor derived cells in the thymus 4 weeks after SL-TBI and adoptive transfer. n=6 in each group, unpaired t-test applied. *: p <0.01*

As the reviewer indicated, SB-TBI is associated with massive changes on many aspects. Therefore, we also tested the role of SIRPα on thymic homing and thymocyte development in steady state. First, we conducted short-term homing assay using sorted lineage- bone marrow progenitor cells instead of total bone marrow cells to avoid the overwhelming effect of massive number of cells used. Short-term homing assay with 1×10^6 lineage^- bone marrow progenitor cells showed similarly significant defect in Sirpα^-/- recipient thymus (new Figure 2I), but not in the spleen (new Figure2—figure supplement 2J). Second, we also examined following T cell development in this scenario. At 3 weeks after adoptive transfer of lineage^- bone marrow progenitor cells, significantly reduced population of donor-derived thymocytes (mainly DP subset) was found in Sirpα^-/- mice (new Figure 2J,K). However, it should be noted that, later stage of thymocyte development, such as SP, was not significantly impaired, although there is a trend to be reduced in Sirpα^-/- mice.

Thus, our data suggest that while SIRPα deficiency results in impaired thymic homing of progenitor cells and is accompanied with reduced ETP population and impaired early thymocyte development, later thymocyte development is less affected probably due to compensation effect. Whether this effect might be amplified at certain scenarios remains an intriguing open question.

Taken together, the study in its presents form contains the description of an interesting new phenotype, consistent with a role of the CD47-SIRPa interaction in colonization of the thymus by bone-marrow derived progenitors. However, at present, homing experiments lack sufficient rigor and experiments on thymus regeneration, while showing an interesting additional finding, do not justify to conclude homing as mechanistic explanation.

Thank you for the comment. With these new data, hopefully the role of SIRPα on thymic progenitor homing, T cell development during steady state and T cell regeneration at SL-TBI scenario has been made clearer. We agree that the causal relationship between thymic progenitor homing and thymus regeneration is still indirect and inconclusive, which may require further investigation in future. In this study, we would like to emphasize more on the novel role of CD47-SIRPα in controlling thymic progenitor homing, and the underlying molecular and biochemical mechanism. We hope these have been validated.

Reviewer #3 (Public Review):

The manuscript by Ren et al. seeks to describe a role for endothelial cell (EC) expression of Sirpα playing a role in the importation of hematopoietic progenitors from the circulation into the thymus. Specifically, the authors demonstrate that there is a reduction in the number of the earliest T lineage progenitors (ETPs) in the thymus in mice deficient for Sirpa or CD47 (its ligand), and through a series of elegant in vitro transendothelial migration studies, identify that intracellular Sirpα signaling mediates this process by regulating VE-Cadherin expression and thus EC tight junctions. In particular, the use of transwell assays modified to study TEM is particularly well utilized to tease apart the mechanisms. Overall, I found this to be an excellent manuscript. In fact, every time I had a critique developing in my head, the authors quickly dispensed of it by producing some follow up data that addressed my concern! My biggest concern with the manuscript is that it was difficult to determine exactly how many repeats of each experiment have been performed and what data is being presented in the figures (and being statistically analyzed). This should not change the conclusions of the manuscript but will make reading the figures and matching them with the legends easier. The following are a some major and minor concerns that should be addressed to strengthen the manuscript:

Major:

• My main concern is that there needs to be greater care taken with highlighting the number of repeats done for each individual study as it is not always clear. For instance, in Figure 2 the data are presented as being representative of three independent experiments with an n of 3 in each experiment but in 2B, D, and F there are 4 data points for the Sirpa-/- group. This is likely explained by there being 4 mice in that particular experiment, but that is why the numbers should be presented for each experiment rather than a general statement at the end. Another example of this is that in Figure 2 S1 the authors would like to claim that the only differences are in the DN1 subsets which contains the ETPs. However, it is likely this is just due to low numbers as it seems like there is a real decrease in the number of DN2, DN3, DN4 and even DP thymocytes (as well as total cellularity).

1. This should not change any conclusions of the paper but will aid in reader interpretation.

Thank you for your advice and we apologize for the negligence and have rechecked all figure legends and reported sample size for each panel individually. Furthermore, we repeated those experiments with too few samples in the group. For mouse experiments, we used littermates for detection which were not always have equal number of individual mouse in each group, now mouse used have been labeled specifically in each experiment. For thymic subset detection in Sirpα^-/- mice, we have increased sample size (n=5 for both Sirpα^-/- and control group as shown in Figure 2—figure supplement 1AE) and indeed found significant decrease of DN2, DN3 and DN4 subsets in Sirpα KO mice, though total cellularity was still not significantly changed. Overall, the conclusion of defective early thymocyte development in Sirpα^-/- mice retains valid.

2. In this manuscript the authors show that Sirpa expression by TPECs is critical for their capacity to guide the importation of HPCs, and in their previous work they have shown that lymphotoxin can regulate the importation capacity of these same TPECs. Therefore, it would be extremely interesting to know if LT signaling is regulating the expression of Sirpa. Furthermore, it would be important to at least comment on what may be influencing Sirpa expression. For instance, we know from the work of Petrie and others that DN niche availability can influence the ability of the thymus to import of progenitors. Similarly, after TBI the "gates" are let open and the capacity of the thymus to import progenitors increases. Do the authors know (or could they comment) on what happens to Sipra expression after TBI in ECs?

Thank you for your suggestion. It is an interesting and important question how SIRPα expression is regulated on TPECs. As the reviewer suggested, we examined SIRPα expression in different settings. Given the important role of LT-LTβR signaling on TPEC development and maintenance, we first tested whether LT-LTβR signal would be required for SIRPα expression. However, the remaining TPECs in Ltbr^-/- mice showed similar level of SIRPα expression compared to that in WT mice (new Figure 1—figure supplement 1C). Thymic stromal niche is another factor regulating thymic settling of progenitor cells (Krueger, 2018; Prockop and Petrie, 2004). Increased thymic stromal niche was found during irradiation (Zlotoff et al., 2011). We also detected SIRPα expression on TEPC at Day 14 after 5.5Gy total body sublethal irradiation and found no significant change in SIRPα expression (new Figure 1—figure supplement 1D). Whether SIRPα expression on TPECs is a constitutive event or regulatable upon thymic microenvironmental change remains to be tested in future.

3. The use of the in vitro TEM assays in transwell plates are a nifty way of interrogating and manipulating the effect of Sirpa in these conditions, however, the caveat is that these all use EC cell lines that do not correspond to the TPECs being described in vivo. This caveat should be acknowledged in the text.

Thank you for the advice, EC cell line we used is a pancreatic islet endothelial cell line (MS1), which is not derived from or corresponding to TPECs. We have mentioned this caveat in the text.

4. I am a little confused as to the interpretation of the final experiment looking at tumor clearance. The authors show that this could be clinically relevant as blockade of the CD47-Sirpa axis is becoming an increasingly attractive immunotherapy option but its use could preclude thymic recovery after damage and thus contribute toward poorer T cell responses against tumors. This last study is very interesting but also very hard to interpret given the likely positive effect of Sirpa-CD47 blockade on tumor clearance, in opposition to its potential effects hindering thymic repair. While it is notable that there is reduced clearance of tumor in mice treated with CV1, it is unclear why there does not seem to be any positive effect of CV1 on tumor clearance (is this because there are fewer T cells in the periphery as it is still early after damage?). On the thymic repair and reconstitution front, perhaps a cleaner way would be to look in Sirpa or CD47 deficient mice and without tumors.

We agree that the findings regarding tumor immunotherapy need further explanation on detailed mechanism, therefore this part of results was removed from this project. CV1 treatment in our approach is ahead of tumor inoculation, therefore, CV1 mediated blockaded of CD47 (which is the case in CV1 mediated tumor clearance) would not occur on tumor cells. However, we did not test for the mechanism behind, which is quite interesting and would be done in future study.

As to the suggestion of testing thymic regeneration in straightforward Sirpα or CD47 deficient mice, we have done this in Sirpα deficient mice. We conducted SL-TBI directly on Sirpα-/- mice and control mice. Congenically marked bone marrow cells were also adoptively transferred for better monitoring. At 4 weeks after transfer, donor derived DN thymocyte subset was found defective in Sirpα-/- recipients compared to that in control hosts (Figure R1). However, DP, SP subsets did not show difference, probably due to compensation effect. (Figure R1).

Minor Comments:

• In Fig. 2I (and Fig. 2S2I-J), it is difficult to determine how long after the chimera transplant the homing assays were performed. However, this approach has limitations as the process of creating those chimeras (conditioning such as irradiation etc.) will change the function and possibly the mechanisms of progenitor entry into the thymus. There is clearly still an effect of Sirpa in this context but it is possible (even likely) that the importation mechanisms in the thymus change after damage such as that caused by the conditioning required in the initial chimera generation.

For the study of short-term homing in bone marrow chimeric mice, we have updated legends for the related figure (which is now Figure 2G in the article). The homing assays were performed at 8 weeks after the chimeric reconstruction. Meanwhile, it is indeed possible that the changes of the thymic homing mechanisms may give rise to the abnormal progenitor cells entry. In order to exclude this potential effect, we conducted homing assays without irradiation. In this experiment, we also observed impaired shortterm homing (new Figure 2I) and following T cell development (new Figure 2J,K)

Furthermore, although using the Tie2-Cre strain will distinguish Sirpa on ECs and TECs, it will not distinguish between expression on other cells such as DCs (Tie2 will delete expression in both endothelial and hematopoietic lineages). Although the optimal experiment to address these concerns would be to delete Sirpa from ECs specifically (such as with Cdh5-CreERT2 mice), I am convinced by the preponderance of in vitro data that there is an EC-specific effect and therefore it is not necessary to perform this time-consuming, albeit interesting, potential experiment. However, these limitations should be acknowledged in the discussion or text.

Thank you for your kind suggestion, we have discussed this limitation in the text.

• As a technical note I am surprised that there was considerable reconstitution of naive T cells at day 21 after TBI (Fig.7G-H). In our experience that is very early for naïve T cells in the periphery which generally take about 4 weeks to start reconstituting in a real sense. Is it possible there are direct effects of this treatment on residual radio-resistant peripheral T cell numbers?

Thank you very much for sharing your information. Indeed, we cannot exclude the possibility of residual radio-resistant peripheral T cells. To better clarify this, we have performed SL-TBI (6 Gy) followed by adoptive transfer of congenically marked WT (CD45.1) total bone marrow cells into Sirpα^-/- or control mice (CD45.2) for better monitoring. In this situation, we found that at day 28, more that 97% of thymocytes were donor-derived in both groups and the thymus had been completely reconstituted (Figure R2). In addition, as have been shown in Figure R1, donor-derived DN thymocyte subset was found significantly reduced in Sirpα^-/- mice compared to that in control mice. However, no defect was found at later development stages of thymocytes.

Given the complication of the original experimental design, and as suggested by the reviewers, the original Fig. 7 was removed. The new data described above are hopeful informative to understand the role of SIRPα in a thymic regeneration scenario.

Figure R4. Chimerism detection at day 28 in host transferred with bone marrow cells. (A) Chimerism of thymic subsets, chimerism=CD45.1^+%/(CD45.1+ %+CD45.2^+ %). (B) Representative FACS of donor (CD45.1) and host (CD45.2) cells in total thymocyte (single and live cell gated). n=6 in each group, unpaired t-test applied. **: p<0.01

Author Response

Reviewer #1 (Public Review):

Point 1) There is affluent evidence that the cortical activity in the waking brain, even in head restrained mice, is not uniform but represents a spectrum of states ranging from complete desynchronization to strong synchronization, reminiscent of the up and down states observed during sleep (Luczak et al., 2013; McGinley et al., 2015; Petersen et al., 2003). Moreover, awake synchronization can be local, affecting selective cortical areas but not others (Vyazovskiy et al., 2011). State fluctuations can be estimated using multiple criteria (e.g., pupil diameter). The authors consider reduced glutamatergic drive or long-range inhibition as potential sources of the voltage decrease but do not attempt to address this cortical state continuum, which is also likely to play a role. For example: does the voltage inactivation following ripples reflect a local downstate? The authors could start by detecting peaks and troughs in the voltage signal and investigate how ripple power is modulated around those events.

Our study is correlational, and hence, we cannot speak as to any casual role that the awake hippocampal ripples may play in the post-ripple hyperpolarization observed in aRSC. It is indeed possible that the post-awake-ripple neocortical hyperpolarization is independent of ripples and reflects other mechanisms that our experiments have possibly been blind to. One such mechanism is neocortical synchronization in the awake state. As reviewer 1 pointed out, it is possible that a proportion of hippocampal ripples occur before neocortical awake down-states. To test this hypothesis, we triggered the ripple power signal by the troughs (as proxies of awake down-states) and peaks (as proxies of awake up-states) of the voltage signals, captured from different neocortical regions, during periods of high ripple activity when the probability of neocortical synchronization is highest (McGinley et al., 2015; Nitzan et al., 2020). According to this analysis (see the figure below), the ripple power was, on average, higher before troughs of aRSC voltage signal than before those of other regions. On the other hand, the ripple power, on average, was not higher after the peaks of aRSC voltage signal than after those of other regions. This observation supports the hypothesis that a local awake down-state could occur in aRSC after the occurrence of a portion of hippocampal ripples. However, a recent work whose preprint version was cited in our submission (Chambers et al., 2022, 2021) reported that, out of 33 aRSC neurons whose membrane potentials were recorded, only 1 showed up-/down-states transitions (bimodal membrane potential distribution). Still, a portion (10 out of 30) of the remaining neurons showed an abrupt post-ripple hyperpolarization. In addition, they reported a modest post-ripple modulation of aRSC neurons’ membrane potential (~ %20 of the up/down-states transition range). Hence, these results suggest that the post-ripple aRSC hyperpolarization is not necessarily the result of down-states in aRSC. A paragraph discussing this point was added to the discussion lines 262-279.

Mean ripple power triggered by troughs and peaks of voltage signal captured from aRSC, V1, and FLS1. Zero time represents the timestamp of neocortical troughs/peaks. The shading represents SEM (n = 6 animals).

Point 2) Ripples are known to be heterogeneous in multiple parameters (e.g., power, duration, isolated events/ ripple bursts, etc.), and this heterogeneity was shown to have functional significance on multiple occasions (e.g. Fernandez-Ruiz et al., 2019 for long-duration ripples; Nitzan et al., 2022 for ripple magnitude; Ramirez-Villegas et al., 2015 for different ripple sharp-wave alignments). It is possible that the small effect size shown here (e.g. 0.3 SD in Fig. 2a) is because ripples with different properties and downstream effects are averaged together? The authors should attempt to investigate whether ripples of different properties differ in their effects on the cortical signals.

The seeming small effect size (e.g. 0.3 SD in Fig. 2a) is because the individual peri-ripple voltage/glutamate traces were z-scored against a peri-non-ripple distribution and then averaged. Alternatively, the peri-ripple traces could have been averaged first, and the averaged trace could have been z-scored against a sampling distribution constructed from the abovementioned peri-non-ripple distribution where the sample size would have been the number of ripples detected for a specific animal. In the latter case, the standard deviation of the sampling distribution would have been used as the divisor in the z-scoring process as opposed to the former case where the standard deviation of the original peri-non-ripple distribution would have been used. Since the standard deviation of the sampling distribution is smaller than the standard deviation of the original distribution by a factor of √(sample size), the final z-scored values in the latter would be higher than those in the former case by a factor of √(sample size). For instance, if the sample size in Fig. 2A (number of ripples) was 100, the mean z-scored value would be 0.3*10 = 3. In any case, it is of interest to investigate the relationship between the ripple and neocortical activity features.

To investigate the relationship between the hippocampal ripple power and the peri-ripple neocortical voltage activity, we focused on the agranular retrosplenial cortex (aRSC) as it showed the highest level of modulation around ripples. To get an idea of what features of the aRSC voltage activity might be correlated with the ripple power, the ripples were divided into 8 subgroups using 8-quantiles of their power distribution, and the corresponding aRSC voltage traces were averaged for each subgroup (similar to the work of Nitzan et al. (Nitzan et al., 2022)). The results of this analysis are summarized in the figure below.

Left: peri-ripple aRSC voltage trace was triggered on ripples in the odd-numbered ripple power subgroups for each animal and then averaged across 6 animals. The standard errors of the mean were not shown for the sake of simplicity. Right: the same as the left panel but for only lowest and highest power subgroups. The shading represents the standard error of the mean.

These results suggested that there might be a positive correlation between the ripple power and the pre-ripple and post-ripple aRSC voltage amplitude. To test this possibility, Pearson’s correlation between the ripple power and pre-/post-ripple aRSC amplitude was calculated for each animal separately. The ripple power for each detected ripple was defined as the average of the ripple-band-filtered, squared, and smoothed hippocampal LFP trace from -50 ms to +50ms relative to the ripple's largest trough timestamp (ripple center). The pre- and post-ripple aRSC amplitude for each ripple was calculated as the average of the aRSC voltage trace over the intervals [-200ms, 0] and [0, 200ms], respectively. The results come as follows.

Top: the scatter plots of the ripple power and pre-ripple aRSC voltage amplitude for individual animals. The black lines in each graph represent the linear regression line. The blue circles in each graph are associated with one ripple. The Pearson’s correlation values (ρ) and the p-value of their corresponding statistical significance are represented on top of each graph. Bottom: the same as top graphs but for post-ripple aRSC amplitude.

According to this analysis, 4 out of 6 animals showed a weak positive correlation (ρ = 0.0806 ± 0.0115; mean ± std), 1 animal showed a negative correlation (ρ = -0.20183), and 1 animal did not show a statistically significant correlation (p-value > 0.05) between ripple power and pre-ripple aRSC voltage amplitude. Moreover, 2 out of 6 animals showed a negative correlation (ρ = -0.1 and -0.14), and 4 animals did not show a statistically significant correlation (p-value > 0.05) between ripple power and post-ripple aRSC voltage amplitude.

To check that the correlation results were not influenced by the extreme values of the ripple power and aRSC voltage, we repeated the same correlation analysis after removing the ripples associated with top and bottom %5 of the ripple power and aRSC voltage values. According to this analysis, 1 out of 6 animals showed a negative correlation (ρ = -0.13), and 5 animals did not show a statistically significant correlation (p-value > 0.05) between ripple power and pre-ripple aRSC voltage amplitude. Moreover, 2 out of 6 animals showed a negative correlation (same animals that showed negative correlation before removing the extreme values; ρ = -0.12 and -0.14), 1 animal showed a positive correlation (ρ = 0.1), and 3 animals did not show a statistically significant correlation (p-value > 0.05) between ripple power and post-ripple aRSC voltage amplitude.

Based on these results, we cannot conclude that there is a meaningful correlation between the ripple power and amplitude of aRSC voltage activity before and after the ripples. It is noteworthy to mention that Nitzan et al. (see Fig S6 in (Nitzan et al., 2022)) did not report a statistically significant correlation between ripple power octile number (by discretizing a continuous-valued random variable into 8 subgroups) and pre-ripple firing rate of the mouse visual cortex. However, they reported a statistically significant negative correlation (ρ = -0.13) between the ripple power octile number and post-ripple firing rate of the mouse visual cortex. It appears that their reported negative correlation was influenced by the disproportionately larger values of the firing rate associated with the first ripple power octile compared to the other octiles. Therefore, repeating their analysis after removing the first octile would probably lead to a weak correlation value close to 0.

Next, we investigated the relationship between ripple duration and aRSC voltage activity. To get an idea of what features of the aRSC voltage activity might be correlated with the ripple duration, the ripples were divided into 8 subgroups using 8-quantiles of their duration distribution, and the corresponding aRSC voltage traces were averaged for each subgroup. The results of this analysis are summarized in the figure below.

Left: peri-ripple aRSC voltage trace was triggered on ripples in the odd-numbered ripple duration subgroups for each animal and then averaged across 6 animals. The standard errors of the mean were not shown for the sake of simplicity. Right: the same as the left panel but for only lower and highest duration subgroups. The shading represents standard error of the mean.

These results do not reveal a qualitative difference between the patterns of aRSC peri-ripple voltage modulation and ripple duration. However, the same correlation analysis performed for the ripple power was also conducted for the ripple duration. Only 1 animal out of 6 showed a statistically significant correlation (ρ = 0.08) between pre-ripple aRSC voltage amplitude and ripple duration.

Moreover, only 1 animal out of 6 showed a statistically significant correlation (ρ = -0.08) between post-ripple aRSC voltage amplitude and ripple duration. In conclusion, there does not seem to be a meaningful linear relationship between peri-ripple aRSC voltage amplitude and ripple duration.

Next, we investigated whether the peri-ripple aRSC voltage modulation differs depending on whether a single or a bundled ripple occurs in the dorsal hippocampus. The bundled ripples were detected following the method described in our previous work (Karimi Abadchi et al., 2020). We found that 9.4 ± 3.5 (mean ± std across 6 animals) percent of the ripples occurred in bundles. Then, the aRSC voltage trace was triggered by the centers of the single as well as centers of the first/second ripples in the bundled ripples, averaged for each animal, and averaged across 6 animals. The results of this analysis are represented in the following figure.

Left: animal-wise average of mean peri-ripple aRSC voltage trace triggered by centers of the single and centers of the first ripple in the bundled ripples. Right: Same to the left but triggered by the centers of the second ripple in the bundled ripples.

These results suggest that the amplitude of aRSC voltage activity is larger before bundled than single ripples, and the timing of aRSC voltage activity is shifted to the later times for bundled versus single ripples. The pre-ripple larger depolarization might signal the occurrence of a bundled ripple (similar to larger pre-bundled- than pre-single-ripple deactivation observed during sleep (Karimi Abadchi et al., 2020)).

Point 3) The differences between the voltage and glutamate signals are puzzling, especially in light of the fact that in the sleep state they went hand in hand (Karimi Abadchi et al., 2020, Fig. 2). It is also somewhat puzzling that the aRSC is the first area to show voltage inactivation but the last area to display an increase in glutamate signal, despite its anatomical proximity to hippocampal output (two synapses away). The SVD analysis hints that the glutamate signal is potentially multiplexed (although this analysis also requires more attention, see below), but does not provide a physiologically meaningful explanation. The authors speculate that feed-forward inhibition via the gRSC could be involved, but I note that the aRSC is among the two major targets of the gRSC pyramidal cells (the other being homotypical projections) (Van Groen and Wyss, 2003), i.e., glutamatergic signals are also at play. To meaningfully interpret the results in this paper, it would be instrumental to solve this discrepancy, e.g., by adding experiments monitoring the activity of inhibitory cells.

Observing that glutamate and voltage signals do not go hand-in-hand in awake versus sleep states was surprising for us as well, and it was the main reason that SVD analysis was performed. Especially that a portion of aRSC excitatory neurons showed elevated calcium activity despite the reduction of voltage and delayed elevation of glutamate signals in aRSC at the population level. At the time of initial submission, pre-ripple reduction and post-ripple elevation of calcium activity in a portion of three subclasses of the superficial aRSC inhibitory neurons were reported (Chambers et al., 2022, 2021), and it was the basis of our speculation on the potential involvement of feed-forward inhibition in the post-ripple voltage reduction. We speculated that the source of this potential feed-forward inhibition could stem from gRSC excitatory neurons, as the reviewer 1 pointed out, or from other neocortical or subcortical regions projecting to aRSC. It is also possible that feedback inhibition would be involved where the principal aRSC neurons that are excited by gRSC (as reviewer 1 pointed out) or any other region, including aRSC itself, excite aRSC inhibitory neurons.

Point 4) I am puzzled by the ensemble-wise correlation analysis of the voltage imaging data: the authors point to a period of enhanced positive correlation between cortex and hippocampus 0-100 ms after the ripple center but here the correlation is across ripple events, not in time. This analysis hints that there is a positive relationship between CA1 MUA (an indicator for ripple power) and the respective cortical voltage (again an incentive to separate ripples by power), i.e. the stronger the ripple the less negative the cortical voltage is, but this conclusion is contradictory to the statements made by the authors about inhibition.

A closer look at Figure 2B iv reveals that elevation of the cross-correlation function between peri-ripple aRSC voltage and hippocampal MUA starts with a short delay (~20 ms) and peaks around 75 ms after the ripple centers. It means the maximum correlation between the two signals occurs at point (75ms, 75ms) on the MUA time-voltage time plane whose origin (i.e. the point (0, 0)) is the ripple centers in the hippocampal MUA and corresponding imaging frame in the voltage signal. Reviewer 1’s interpretation would be correct if the maximum correlation occurred at the point (0, 0) not at the point (75ms, 75 ms). It is because the MUA value at the time of ripple centers (t = 0) is the indicator of the ripple power not at the time t = 75ms. Figure 2B iii shows that the amplitude of hippocampal MUA is more than 2 dB less at t = 75ms than at t = 0 which is a reflection of the fact that ripples are often short-duration events. Instead, if the maximum correlation occurred at the point (0, 100ms) where the ripples had maximum power and aRSC voltage was at its trough (Figure 2B iii), it could have been concluded that “the stronger the ripple the less negative the cortical voltage”.

Point 5) Following my previous point, it is difficult to interpret the ensemble-wise correlation analysis in the absence of rigorous significance testing. The increased correlation between the HPC and RSC following ripples is equal in magnitude to the correlation between pre-ripple HPC MUA and post-ripple cortical activity. How should those results be interpreted? The authors could, for example, use cluster-based analysis (Pernet et al., 2015) with temporal shuffling to obtain significant regions in those plots. In addition, the authors should mark the diagonal of those plots, or even better compute the asymmetry in correlation (see Steinmetz et al., 2019 Extended Fig. 8 as an example), to make it easier for the reader to discern lead/lag relationships.

The purpose of calculating the ensemble-wise correlation coefficient was to provide further information about the relationship between the two random processes peri-ripple HPC MUA and peri-ripple neocortical activity. In general, the correlation between the two random processes cannot be inferred from the temporal relationship between their mean functions. In other words, there are infinitely many options for the shape of the correlation function between two random processes with given mean functions. Moreover, the point was to compare the correlation of peri-ripple neocortical activity and HPC MUA across neocortical regions. The fact that mean peri-ripple activity in, for example, RSC and FLS1 are different does not necessarily mean their correlation functions with peri-ripple HPC MUA are also different.

As requested, we performed cluster-based significant testing via temporal shuffling for each individual VSFP (n = 6), iGluSnFR Ras (n = 4), and iGluSnFR EMX (n = 4) animals. The following figures summarize the number of animals showing significant regions in their correlation functions between peri-ripple HPC MUA and different neocortical regions. The diagonal of the correlation functions is marked; however, the temporal lead/lag should not be inferred from these results mainly because the temporal resolution of the two signals, one electrophysiological and one optical, are not the same.

Point 6) For the single cell 2-photon responses presented in Fig. 3, how should the reader interpret a modulation that is at most 1/20 of a standard deviation? Was there any attempt to test for the significance of modulation (e.g., by comparing to shuffle)? If yes, what is the proportion of non-modulated units? In addition, it is not clear from the averages whether those cells represent bona fide distinct groups or whether, for instance, some cells can be upmodulated by some ripples but downmodulated by others. Again, separation of ripples based on objective criteria would be useful to answer this question.

As explained in response to point 2, the seeming small modulation size (e.g. 0.05 SD in Fig. 3b) is because the individual peri-ripple calcium traces were z-scored against a peri-non-ripple distribution and then averaged. Alternatively, the peri-ripple traces could have been averaged first, and the averaged trace could have been z-scored against a sampling distribution constructed from the abovementioned peri-non-ripple distribution where the sample size would have been the number of ripples detected for a specific animal. In this latter case, the standard deviation of the sampling distribution would have been used as the divisor in the z-scoring process as opposed to the former case where the standard deviation of the original peri-non-ripple distribution would have been used. Since the standard deviation of the sampling distribution is smaller than that of the original distribution by a factor of √(sample size), the final z-scored values in the latter would be higher than those in the former case by a factor of √(sample size).

As suggested by the reviewer and to make our results more comparable with those of electrophysiological studies, we deconvolved the calcium traces and tested for the significance of the modulation of each neuron by comparing its mean peri-ripple deconvolved trace with a neuron-specific shuffled distribution (see the methods section for details). We found %8.46 ± 3 (mean ± std across 11 mice) of neurons were significantly modulated over the interval [0, 200ms] and %81.08 ± 8.91 (mean ± std across 11 mice) of which were up-modulated. If the criterion of being distinct is being significantly up- or down-modulated, these two groups could be considered distinct groups. The following figures show mean peri-ripple calcium and deconvolved traces, averaged across up- or down-modulated neurons for each mouse and then averaged across 11 mice.

Point 7) Fig. 3: The decomposition-based analysis of glutamate imaging using SVD needs to be improved. First, it is not clear how much of the variance is captured by each component, and it seems like no attempt has been made to determine the number of significant components or to use a cross-validated approach. Second, the authors imply that reconstructing the glutamate imaging data using the 2nd-100th components 'matches' the voltage signal but this statement holds true only in the case of the aRSC and not for other regions, without providing an explanation, raising questions as to whether this similarity is genuine or merely incidental.

The first 100 components explained about %99.9 of the variance in the concatenated stack of peri-ripple neocortical glutamate activity for each animal which is practically equivalent to the entire variance in the data. Our goal was not to obtain a low-rank approximation of the data for which the number of significant components had to be determined. Instead, we decomposed the data into the activity along the first principal component for which there was no noticeable topography among neocortical regions and the activity along the rest of the components for which there was a noticeable topography among neocortical regions. The first component explained %83.11 ± 6.75 (mean ± std across 4 iGluSnFR Ras mice) and %83.3 ± 5.07 (mean ± std across 4 iGluSnFR EMX mice) of variance in the concatenated stack of peri-ripple neocortical glutamate activity.

As we discussed in the discussion section of the manuscript, SVD is agnostic about brain mechanisms and only cares about capturing maximum variance. Specifically, it is not designed to capture the maximum similarity between glutamate and voltage activity in the brain. Therefore, the only thing we can say with certainty comes as follows: when the activity along the axis with maximum co-variability (1st principal component) across the neocortical regions’ glutamate activity is removed, only aRSC, and no other regions, show a post-ripple down-modulation, whose timing matches that of aRSC post-ripple voltage down-modulation. Moreover, the timing of activity of 1st principal component matches better with that of calcium activity among the up-modulated portion of aRSC neurons. Even though the genuineness of these results is not guaranteed, the similarity between the timing of SVD output in aRSC glutamatergic activity with that in two independently collected signals in aRSC, i.e. voltage and calcium, could support the idea that peri-ripple aRSC glutamatergic activity is likely a mixture of up- and down-modulated components.

Point 8) The estimation of deep pyramidal cells' glutamate activity by subtracting the Ras group (Fig. 4d) is not very convincing. First, the efficiency of transgene expression can vary substantially across different mouse lines. Second, it is not clear to what extent the wide field signal reflects deep cells' somatic vs. dendritic activity due to non-linear scattering (Ma et al., 2016), and it is questionable whether a simple linear subtraction is appropriate. The quality of the manuscript would improve substantially if the authors probe this question directly, either by using deep layer specific line/ 2-P imaging of deep cells or employing available public datasets.

Simulation studies have suggested that the signal, captured by wide-field imaging of voltage-sensitive dye, can be modeled as a weighted sum of voltage activity across neocortical layers (Chemla and Chavane, 2010; Newton et al., 2021). Hence, modeling the glutamate signal as a weighted sum of the glutamate activity across neocortical layers is a good starting point. Future studies would be needed to improve this starting point by imaging glutamate activity in a cohort of mice with iGluSnFR expression in only deep layers’ neurons. Moreover, Ma et al. (Ma et al. 2016) stated that “This means that signal detected at the cortical surface (in the form of a two-dimensional image) represents a superficially weighted sum of signals from shallow and deeper layers of the cortex”.

Reviewer #2 (Public Review):

Point 1) The authors throughout the manuscript compare the correlation between hippocampal MUA and the imaged cortical ensemble activity (Example: Lines 120-122). There is a potential time lag in signal detection with regard to the two detection methods. While the time lag using electrophysiological recording is at the scale of milliseconds, the glutamate-sensitive imaging might take several 100s of ms to be detected. It is not clear in the manuscript how the authors considered this problem during the analysis.

The ensemble-wise correlation analysis characterizes the relationship between two random processes, peri-ripple HPC MUA and peri-ripple neocortical activity (please see the response to reviewer 1’s major point 5). Although it is a valid point that the temporal resolution of the two signals is not the same which could introduce an error in the exact timing of the relationship between the two processes, we did not draw any conclusion based on the exact timing of the elevated correlation between the two processes. Moreover, we smoothed (equivalent to low-pass filtering) and down-sampled the MUA signal (please see the methods section) to bring the temporal scale of the two processes closer to each other. We also want to clarify that the temporal resolution of voltage and glutamate imaging is in the range of 10s of ms (Xie et al., 2016).

Point 2) In the results section "The peri-ripple glutamatergic activity is layer dependent", are the Ras and EMX expressed in two different experimental animal groups? If yes, and there was a time lag between the two groups, is it valid to estimate the deeper layer activity using a scaled version of the Ras from the EMX signal?

This comment is addressed in response to reviewer 1’s major point 8.

Point 3) The authors did not discuss the results adequately in the discussion section. Since there is no behavioral paradigm and no behavioral read-out to induce or correlate it with possible planning and future decision-making process, the significance of the paper will be enhanced by discussing the possible underlying circuitry mechanism that might cause the reported observations. With no planning periods in the task (instead just sitting on a platform), it is actually quite unclear what the purpose of wake ripples should be. For example, the authors discuss the superficial and deep layer responses and their relation to the memory index theory. However, the RSC possesses different groups of excitable neurons in different layers. Specifically, three excitable neurons are found within the different layers of the RSC; the intrinsically bursting neurons (IB), regular spiking (RS), and low-rheobase (LR) neurons. These neurons are distributed heterogeneously within the RSC cortical layer. Although the RS are abundant in the deeper layers of the RSC, they occupy 40% of the total amount of excitable neurons found in layers II/III. On the other hand, the LR is the dominant excitable neuron in the superficial layers. It will add to the significance of the work if the authors discussed the results in the context of the cellular structure of the RSC and how would that impact the observed inhibition in the peri-ripple time window. It would be helpful for the readers and the reviewers to add a schematic diagram to the discussion section.

The goal of our study was to characterize the patterns of neocortical activity around hippocampal ripples in the awake state and not shed light on the function (purpose) of awake ripples. However, we speculated about what our results could mean in the discussion section. To address the reviewer’s comment on the differences across RSC layers, the following paragraph was added to the discussion section lines 342-353.

“Our results suggest that dendrites of deep pyramidal neurons, arborized in the superficial layers of the neocortex, receive glutamatergic modulation earlier than those of the superficial ones. However, the results do not provide a mechanistic explanation of the phenomenon. It is possible that the observed layer-dependency of the glutamatergic modulation would partially result from the heterogeneity of the excitatory as well as inhibitory neurons across aRSC layers. But, the question is how this heterogeneity may lead to the above-mentioned layer-dependency to which our data does not provide an answer. It could be speculated that the difference in the dendritic morphology and firing type of different types of RSC excitatory neurons (Yousuf et al., 2020) or the difference in connectivity of different RSC layers with other brain regions would play a role (Sugar et al., 2011; van Groen and Wyss, 1992; Whitesell et al., 2021). This is a complicated problem and could only be resolved by conducting experiments specifically designed to address this problem.”

Point 4. A general issue (in addition to the missing behaviour), is the mix of the methods. On one side this makes the article very interesting since it highlights that with different methods you actually observe different things. But on the other side, it makes it very difficult to follow the results. It would be a major improvement of the article if the authors could include (as mentioned above) a schematic of the results and their theory, especially highlighting how the different methods would capture different parts of the mechanism. Finally, the authors should not use calcium signals as a direct measure of neuronal firing. Calcium influx is only seen in bursts of firing, not with individual spikes. It is a plasticity signal and therefore should be treated and discussed as such. Just recently it was shown by Adamantidis lab that the calcium signal changes between wake and sleep and this change does not parallel changes in neuronal firing/spikes.

We agree with the reviewer that the calcium signal is biased toward burst of spikes (Huang et al., 2021). To address this concern, the term “spiking activity” was replaced with “calcium activity” throughout the manuscript. Moreover, the calcium signal was deconvoled to get a better estimate of the spiking activity (please refer to our response to the reviewer 1’s point 6).

Point 5. In the discussion section, the authors focus their discussion on the connectivity between the CA1 area and the RSC. Although it is an important point, since the authors are examining the peri-ripple cortical dynamics, it is critical to discuss other possible connectivity effects. Furthermore, the hippocampal input preferentially targets the granular RSC, how would that impact the results and the interpretation of the authors? Additionally, a previous study reported the suppression of the thalamic activity during hippocampal ripples (Yang et al., 2019). Importantly, the thalamic inputs to the RSC target the superficial layers. It will add to the value of the paper if the authors expanded the discussion section and elaborated further on the possible interpretation of the results.

At the time of our initial submission, pre-ripple reduction and post-ripple elevation of calcium activity in a portion of three subclasses of the superficial aRSC inhibitory neurons were reported (Chambers et al., 2022, 2021), and it was the basis of our speculation on the potential involvement of feed-forward inhibition in the post-ripple voltage reduction. We speculated that the source of this potential feed-forward inhibition could stem from gRSC excitatory neurons or other neocortical or subcortical regions projecting to aRSC (please see the discussion section). However, the source being from the thalamus is less likely because multiple studies have observed the suppression of the majority of thalamic neurons during awake ripples (Logothetis et al., 2012; Nitzan et al., 2022; Yang et al., 2019). Moreover, peri-awake-ripple suppression of thalamic axons projecting to the first layer of aRSC is reported (Chambers et al., 2022). On the other hand, it is also possible that feedback inhibition would be involved where the excitatory aRSC neurons that are excited by gRSC (as reviewer 1 pointed out) or any other region, including aRSC itself, excite aRSC inhibitory neurons which in turn inhibit pyramidal cells. To address this comment, the following paragraph was added to the discussion section in lines 323-328.

“Thalamus is another source of axonal projections to aRSC (Van Groen and Wyss, 1992). However, it is less likely that thalamic projections contribute to the peri-awake-ripple aRSC activity modulation because multiple studies have observed the suppression of the majority of thalamic neurons during awake ripples (Logothetis et al., 2012; Nitzan et al., 2022; Yang et al., 2019). Moreover, peri-awake-ripple suppression of thalamic axons projecting to the first layer of aRSC is reported (Chambers et al., 2022).”

Author Response

Reviewer #3 (Public Review):

The authors showed that D2R antagonism did not affect the initial dip amplitude but shortened the temporal length of the dip and the rebound ACh levels. In addition, by using both ACh and DA sensors, the authors showed DA levels correlate with ACh dip length and rebound level, not the dip amplitude. Both pieces of evidence support their conclusion that DA does not evoke the dip but controls the overall shape of ACh dip. Overall the current study provides solid data and interpretation. The combination of D2R antagonist and CIN-specific Drd2 KO further support a causal relationship between DA and ACh dip. Overall, the experiments are well-designed, carefully conducted and the manuscript is well-written.

At the behavioral level, the author found a positive correlation between total AUC (of ACh signal dip) and press latency in Figure 10, indicating cholinergic levels contributes to the motivation. The next logic experiment would be to compare the press latency between control and ChAT-Drd2KO mice, since KO mice have smaller AUC while not affecting DA. However, this piece of information was missing in the manuscript. The author instead showed the correlation between AUC and latency was disrupted, which is indirectly related to the conclusion and hard to interpret. Figure 10 showed that eticlopride elongates the press latency, in a dose-dependent manner. However, it is not clear what this press latency means and how it was measured in this CRF task (Since there is no initial cue in the CRF test, how can we define the press latency?).

We did compare the press latency between control and ChATDrd2KO mice (Figure 10B). At baseline (saline), there is no difference between press latency between these two groups. We measured press latency as the time to press the lever after the lever has been extended. When the lever extends, it makes a sound (cue), which signals to the mice that a new trial has started. The fact that press latency is not enhanced in ChATDrd2KO mice was surprising to us. It is possibly due to compensation via other neuronal mechanisms that regulate press latency (see discussion to comment 6 of public review).

Pearson r<0.5 is normally defined as a weak correlation. It is better to state r values and discuss that in the manuscript.

A valid comment. We clarified our correlation analyses in the methods section (line 717):

“We used a variance explained statistical analysis (R2) to determine the % of variance in our correlation analyses (example: a correlation of 0.5 means 0.52 X 100= 25% of the variance in Y is “explained” or predicted by the X variable. When comparing correlation values, Fisher’s transformation was used to convert Pearson correlation coefficients to z-scores.”

We also added this to the result section: e.g., line 256: “which accounts for 22% of the variance in the ACh decrease explained by the DA peak.

Is there any correlation between ACh AUC and other behavior indexes such as press speed or the time between press and reward licking?

We don’t have the ability to measure press speed and there is no press rate because the lever retracts after the first lever press. We quantified the correlation between time to press until head entry (press to reward latency) and ACh AUC and the results are difficult to interpret. For Drd2f/fl control mice we determined a weak negative correlation (the larger the ACh dip the lower the press to reward latency). In contrast, in ChATDrd2KO mice we found a weak positive correlation between ACh AUC and press to reward latency (the smaller the dip, the lower the press to reward latency). Given these conflicting results, it is difficult to determine how the ACh AUC affects press to reward latency.

In figure 2B CS+ group, the author was focusing on the responses at CS+, however, the ACh dips at reward delivery seem to persist even after in this particular example. This might be an interesting phenomenon in which ACh got dissociated from DA signals, which needs further analysis from the author.

We see a persistent signal at reward delivery in both DA and ACh up to the 8 days of testing. However, 1 mouse lost its optical fiber for the GACh signal so the data from Days 6-8 is from 2 mice. We also measured the correlation between DA and ACh at reward delivery for all 8 days of testing (see below). The correlation data is variable with the strongest correlation being observed on Day 2. It is possible that these signals could get dissociated after even more days of testing, but we do not have this data available.

Author Response:

Reviewer #1:

In this ms, Voroslakos et al., describe a customizable and versatile microdrive and head cap system for silicon probe recordings in freely moving rodents (mice and rats). While there are similar designs elsewhere, the added value here is: a) a carefully designed solution to facilitate probe recovery, thus reducing experimental costs and favoring reproducibility; b) flexibility to accommodate several microdrives and additional instrumentation; c) open access design and documentation to favor customization and dissemination. Authors provide detailed description to faccilitate building the system.

Personally, I found this resource very useful to democratize multi-site recordings, not only for standard silicon probes, but also more novel integrated optoelectrodes and neuropixels. While there are other solutions, this design is quite simple and versatile. A potential caveat is whether it could be perceived as just an upgrade, given some similitudes with previous designs (e.g. Chung et al., Sci Rep 2017 doi: 10.1038/s41598-017-03340-5) and concepts (Headley et al., JNP doi: 10.1152/jn.00955.2014). However, the system presented in this paper provides added value and knowledge-based solutions to make silicon probe recordings more accessible.

We thank the reviewer for carefully reading our manuscript and providing useful and constructive comments.

Reviewer #2:

This manuscript provides an updated guide on the procedures for performing chronic recordings with silicon probes in mice and rats in the lab of the senior author, who is one of the leaders in the use of this experimental method. The new set of procedures relies on metal and plastic 3D printed parts, and represents a major improvement over the older methodology (i.e. Vandecasteele et al. 2012).

The manuscript is clearly written and the technical instructions (in the Methods section) seem rather detailed. The main concerns I had are as follows.

We thank the reviewer for carefully reading our manuscript and providing useful and constructive comments.

1) The present design is an improvement over Chung et al. (the most similar previously published explantable microdrive design, as far as I am aware) in terms of the footprint and travel distance. However, a main disadvantage of the system in its present form is that (apparently) it does not support Neuropixels probes. While such probes might not be suitable for some uses (e.g. to record from large populations in dorsal hippocampus), Neuropixels probes are of considerable interest to many labs.

Our microdrive and head cap system can also support Neuropixels probes. Since our initial submission, we have implanted a Neuropixels probe in the intermediate hippocampus of a rat using our recoverable, plastic microdrive. At the end of the experiment, the Neuropixels probe was successfully recovered, cleaned, and implanted again in a new rat. In addition, we designed a new arm for our metal microdrive which can support Neuropixels probes (Figure 2) and implanted another rat (Figure 3 and 4). We have also created a video showing how to attach Neuropixels probe to a metal microdrive (Suppl. Video 3).

Figure 2. Metal microdrive adapter for Neuropixels probe. A Arm design for 64-channel silicon probes. 45o, front, side and top views are shown (from left to right). All dimensions are in mm. B Changing the overall length (from 7.35 mm to 10 mm) and width (from 4 mm to 5.4 mm) of the 64-channel arm makes our metal microdrive compatible with Neuropixels probe. Note, that only three dimensions of the 64-channel arm were modified (red numbers). 45-degree, front, side and top views are shown (from left to right). All dimensions are in mm. C Photograph of the different arm designs of the metal, recoverable microdrive (top shows an arm designed for a 64-ch silicon probe, bottom shows an arm designed for Neuropixels probe).

Figure 3. Recording of unit firing with Neuropixels probe attached to a metal microdrive in freely moving rat. A Metal microdrive for Neuropixels probe (a – stereotax attachment, b – drive holder, c – metal microdrive, d – Neuropixels probe and e – Neuropixels headstage). B Photo of Neuropixels probe attached to a metal microdrive (a-e same as in A). C Location of probe implantation (Bregma - 4.8 mm, mediolateral + 4.6 mm, 11-degree angle). D High pass filtered traces (1s) from a freely moving rat implanted with Neuropixels probe. Note the single unit activity in the cellular layer of cortex (top) and hippocampus (bottom).

Figure 4. Implantation of Neuropixels probe in a rat using metal microdrive and rat cap system. A The base of the rat cap is attached to the skull. Reference (ref) and ground (gnd) screws are placed over the cerebellum. Neuropixels probe is mounted on a metal microdrive. The microdrive is held by the drive holder and attached to a stereotax arm using the stereotax attachment. For more details, see video: Neuropixels_attachment.mp4. B Once the probe is inserted to its final depth (left), the base of the microdrive is cemented to the skull (zoomed in photograph on the right). C The surface of the brain is kept wet using saline during probe insertion and during cementing the base of the microdrive. D After the base is cemented, the craniotomy is sealed with bone wax. E Releasing the drive from the drive holder. Once it is released the stereotax arm is moved upwards. F Neuropixels headstage is removed from the male header of the stereotax attachment (soldering joint) and placed on the animals back. G The walls of the cap system are attached to the base. Ground and reference wires are soldered to the probe (not shown). H The male header of the headstage is secured to the walls. The headstage and its cable are oriented to allow easy access to the screw head of the microdrive. Note, that there is enough room for custom connectors inside the rat cap.

2) The total weight of the mouse implant seems quite high (together with the headstage, I estimate it is >= 4gr). Could the authors provide the exact value, and describe whether this has any impact on the way the animal moves? Also, the authors should describe how the animals are housed (e.g. do they carry the headstage even when not being recorded). The authors say that a mouse can be implanted with more than one microdrive. The authors should clarify whether they actually have an experience with such implants, or is this just a suggestion based on their educated estimate?

The total weight of the metal microdrive, including the base, body and arm is 0.87 gram. Additional weight is the metabond and dental acrylic cement. The amount of cement that is used during surgery can vary between researchers and the type of surgery. The overall weight of the assembly also depends on the silicon probe with Omnetics connector(s) that is used for the surgery, e.g.: 32-channel micro-LED probe is 1.11g (NeuroLight Technologies LTD.), 64-channel 4-shank probe is 0.96g (ASSY E-1, Cambridge NeuroTech), 64-channel 5-shank probe is 1.05g (A5x12- 16-Buz-Lin-5mm, NeuroNexus Ltd.) and a 128-channel 4-shank probe with integrated Intan chips is 0.94g (P128-5, Diagnostic Biochips). In addition, the overall weight of the entire assembly can change if optic fibers are used in optogenetic studies or if any custom connectors are implanted (e.g., connector and wires for brain stimulation). That is the reason why we reported the overall weight of each system (metal microdrive, mouse cap and rat cap) individually.

The implanted mice are single housed, and they do not carry the headstage while in the vivarium. During recordings, the headstage is attached and a counterbalanced pulley system ensures that the animal is not carrying the extra weight of the headstage. We have quantitatively compared running speed with traditional and the new head caps in both rats and mice (Fig. 6).

The small footprint of the metal microdrive enables researchers to perform more than one silicon probe implantation in freely moving mice. For this purpurse, larger mice (>35 g) are selected (Figure 5).

Figure 5. Metal microdrive enables double silicon probe recordings in freely moving mice. A Intraoperative photograph of double silicon probe implantation. Note that the metal microdrive on the left had been secured to the skull and the second drive is being implanted using the stereotaxic attachment and drive holder. The probe PCBs are placed on the copper mesh. B Photograph focused on the metal microdrives.

3) There is no information in the results section on the number of implants performed, the duration the animals were implanted, the quality of the recordings obtained, number of successes or failures failures. The figures merely provide examples of one successful recording in a mouse and in a rat. All these details should be provided, along with details of how many probes were reused and how many times (a brief mention of one case, lines 252-253 and 359-360, is not sufficient).

We have added a Supplementary Table explaining all the details of our implants. We would like to refer the Reviewer to response #1 to Reviewer 1.

Adapting new technology is challenging. To date, we have extensive experience with the rat cap system only (n=3 users in the lab, n = 25 rats implanted). Two lab members have started to adapt our mouse cap and implanted 3 mice since our submission. We included their maze running behavioral data for comparison between the copper mesh and cap system.

Prior to the development of the metal microdrive, we have conducted an internal lab survey comparing the hand-made microdrive (Vandecasteele et al., 2012) and our recoverable, plastic microdrive. Six lab members who had extensive experience with both types participated (Figure 6). Our questions were:

1) On a scale 1-10, how would you compare the plastic, recoverable drive to the Vandecasteele et. al. 2012 one in terms of: a) ease of building a drive, b) size and c) ease of recovery.

Figure 6. Internal lab survey using recoverable, plastic microdrives. A User feedback based on four criteria: ease of building, ease of implantation, size, ease of recovery. The 3D printed microdrive surpasses the manually built drive (Vandecasteele et. al., 2012) on every parameter except the size. B 24 silicon probes were used with the recoverable plastic microdrive. On average each probe was recovered two times. Out of these 48 recovery attempts 5 failed only. There were 2 total losses during recovery and in three cases different number of shanks broke during the recovery process making the recovery partially successful. One major limitation of reusability is the sudden increase in impedance over time (we have to discard 30% of the successfully recovered probes due this reason). Researchers in our lab spend on average 30 minutes to recover a silicon probe.

Overall, the success rate of recovery is much higher using a recoverable microdrive system, but the size of the plastic, recoverable microdrive is limits certain experiments. This was one of the main motivations to develop the metal, recoverable microdrive.

4) In fig. 2, spike waveforms are classified as pyramidal, wide or narrow interneurons. I did not find any description of how this classification was performed.

We have removed the single cell putative cell types from the manuscript as this issue is not relevant to the current manuscript. Figure 2 has been simplified and a new figure 5 is dedicated to the single cell quantification.

5) Also in fig. 2, refractory period violations are reported in percent (permille in fact). First, it is not clear how refractory period was defined. Second, such quantification is incorrect in principle: we use refractory period violations to infer the rate of false positives. Yet the relationship between fraction of ISI violations and false positive rate depends on the firing rate of the neuron. For example, 0.1% of ISI violations is quite good for a unit spiking at 10 spikes/s, is so so for a unit spiking at 1 spike/s, and is very bad if the firing rate is 0.1 spike/s (see Hill et al. JNeurosci. 2011 for derivation). Alternatively, the authors can follow an approach described in an old paper by the same lab (Harris et al., JNeuropsysiol. 2000), quantifying the violations in spike autocorrelogram relative to its asymptotic height.

We have removed this panel from Fig. 2 and dedicated a new figure (Fig. 5) to the single cell quantification. Refractory violations can be used as an alarm for poor cluster quality. Absence of refractory violations alone does not guarantee good separation for the reasons the Reviewer mentioned.

6) Line 477: the authors write that the probes were mounted on a plastic microdrive. This seems to contradict the key claim of the manuscript (namely that the microdrives were from stainless steel).

We apologize if this description was not clear in the original manuscript. In the revised version, we have added a table (Suppl. Table 1) explaining all details of each animal subject (species, strain, weight, cap type), type of silicon probe and microdrive used. As we explained in Response 3, our main goal was to test each system individually and once all components have been verified, we combined everything into one surgery.

The plastic and metal microdrives are based on the same principles. The implantation/recovery tools are also identical in design concepts. Based on our own experience, users dol not recognize any changes in terms of ease of use, ease of implantation and ease of recovery when changing from plastic recoverable microdrives to metal ones. The advantage of metal drives is size reduction, their multiple reusability and stability.

7) I believe that the work of Luo & Bondy et al. (eLife 2020) and should be references and compared to.

We reference Luo et. al. (2020) in our revised manuscript. One of the main advantages of using a microdrive system is the ability to move the recording probe inside the brain tissue and sample new sets of neurons. This is not the case in Luo & Bondy et al. (eLife 2020).

Author response:

Reviewer #1 (Public Review):

Reviewer #1, comment #1: The study is thorough and systematic, and in comparing three well-separated hypotheses about the mechanism leading from grid cells to hexasymmetry it takes a neutral stand above the fray which is to be particularly appreciated. Further, alternative models are considered for the most important additional factor, the type of trajectory taken by the agent whose neural activity is being recorded. Different sets of values, including both "ideal" and "realistic" ones, are considered for the parameters most relevant to each hypothesis. Each of the three hypotheses is found to be viable under some conditions, and less so in others. Having thus given a fair chance to each hypothesis, nevertheless, the study reaches the clear conclusion that the first one, based on conjunctive grid-by-head-direction cells, is much more plausible overall; the hypothesis based on firing rate adaptation has intermediate but rather weak plausibility; and the one based on clustering of cells with similar spatial phases in practice would not really work. I find this conclusion convincing, and the procedure to reach it, a fair comparison, to be the major strength of the study.

Response: Thanks for your positive assessment of our manuscript.

Reviewer #1, comment #2: What I find less convincing is the implicit a priori discarding of a fourth hypothesis, that is, that the hexasymmetry is unrelated to the presence of grid cells. Full disclosure: we have tried unsuccessfully to detect hexasymmetry in the EEG signal from vowel space and did not find any (Kaya, Soltanipour and Treves, 2020), so I may be ranting off my disappointment, here. I feel, however, that this fourth hypothesis should be at least aired, for a number of reasons. One is that a hexasymmetry signal has been reported also from several other cortical areas, beyond entorhinal cortex (Constantinescu et al, 2016); true, also grid cells in rodents have been reported in other cortical areas as well (Long and Zhang, 2021; Long et al, bioRxiv, 2021), but the exact phenomenology remains to be confirmed.

Response: Thank you for the suggestion to add the hypothesis that the neural hexasymmetry observed in previous fMRI and intracranial EEG studies may be unrelated to grid cells. Following your suggestion, we have now mentioned at the end of the fourth paragraph of the Introduction that “the conjunctive grid by head-direction cell hypothesis does not necessarily depend on an alignment between the preferred head directions with the grid axes”. Furthermore, at the end of section “Potential mechanisms underlying hexadirectional population signals in the entorhinal cortex” (in the Discussion) we write: “However, none of the three hypotheses described here may be true and another mechanism may explain macroscopic grid-like representations. This includes the possibility that neural hexasymmetry is completely unrelated to grid-cell activity, previously summarized as the ‘independence hypothesis' (Kunz et al., 2019). For example, a population of head-direction cells whose preferred head directions occur at offsets of 60 degrees from each other could result in neural hexasymmetry in the absence of grid cells. The conjunctive grid by head-direction cell hypothesis thus also works without grid cells, which may explain why grid-like representations have been observed (using fMRI) in regions outside the entorhinal cortex, where rodent studies have not yet identified grid cells (Doeller et al., 2010; Constantinescu et al., 2016). In that case, however, another mechanism would be needed that could explain why the preferred head directions of different head-direction cells occur at multiples of 60 degrees. Attractor-network structures may be involved in such a mechanism, but this remains speculative at the current stage.” We now also mention the results from Long and Zhang (second paragraph of the Introduction): “Surprisingly, grid cells have also been observed in the primary somatosensory cortex in foraging rats (Long and Zhang, 2021).”

Regarding your EEG study, we have added a reference to it in the manuscript and state that it is an example for a study that did not find evidence for neural hexasymmetry (end of first paragraph of the Discussion): “We note though that some studies did not find evidence for neural hexasymmetry. For example, a surface EEG study with participants “navigating” through an abstract vowel space did not observe hexasymmetry in the EEG signal as a function of the participants’ movement direction through vowel space (Kaya et al., 2020). Another fMRI study did not find evidence for grid-like representations in the ventromedial prefrontal cortex while participants performed value-based decision making (Lee et al., 2021). This raises the question whether the detection of macroscopic grid-like representations is limited to some recording techniques (e.g., fMRI and iEEG but not surface EEG) and to what extent they are present in different tasks.”

Reviewer #1, comment #3: Second, as the authors note, the conjunctive mechanism is based on the tight coupling of a narrow head direction selectivity to one of the grid axes. They compare "ideal" with "Doeller" parameters, but to me the "Doeller" ones appear rather narrower than commonly observed and, crucially, they are applied to all cells in the simulations, whereas in reality only a proportion of cells in mEC are reported to be grid cells, only a proportion of them to be conjunctive, and only some of these to be narrowly conjunctive. Further, Gerlei et al (2020) find that conjunctive grid cells may have each of their fields modulated by different head directions, a truly surprising phenomenon that, if extensive, seems to me to cast doubts on the relation between mass activity hexasymmetry and single grid cells.

Response: We have revised the manuscript in several ways to address the different aspects of this comment.

Firstly, we agree with the reviewer that our “Doeller” parameter for the tuning width is narrower than commonly observed. We have therefore reevaluated the concentration parameter κ_c in the ‘realistic’ case from 10 rad-2 (corresponding to a tuning width of 18o) to 4 rad-2 (corresponding to a tuning width of 29o). We chose this value by referring to Supplementary Figure 3 of Doeller et al. (2010). In their figure, the tuning curves usually cover between one sixth and one third of a circle. Since stronger head-direction tuning contributes the most to the resulting hexasymmetry, we chose a value of κ_c=4 for the tuning parameter, which corresponds to a tuning width (= half width) of 29o (full width of roughly one sixth of a circle). Regarding the coupling of the preferred head directions to the grid axes, the specific value of the jitter σc = 3 degrees that quantifies the coupling of the head-direction preference to the grid axes was extracted from the 95% confidence interval given in the third row of the Table in Supplementary Figure 5b of Doeller et al. 2010. We now better explain the origin of these values in our new Methods section “Parameter estimation” and provide an overview of all parameter values in Table 1.

Furthermore, in response to your comment, we have revised Figure 2E to show neural hexasymmetries for a larger range of values of the jitter (σc from 0 to 30 degrees), going way beyond the values that Doeller et al. suggested. We have also added a new supplementary figure (Figure 2 – figure supplement 1) where we further extend the range of tuning widths (parameter κ_c) to 60 degrees. This provides the reader with a comprehensive understanding of what parameter values are needed to reach a particular hexasymmetry.

Regarding your comments on the prevalence of conjunctive grid by head-direction cells, we have revised the manuscript to make it explicit that the actual percentage of conjunctive cells with the necessary properties may be low in the entorhinal cortex (first paragraph of section “A note on our choice of the values of model parameters” of the Discussion): “Empirical studies in rodents found a wide range of tuning widths among grid cells ranging from broad to narrow (Doeller et al., 2010; Sargolini et al., 2006). The percentage of conjunctive cells in the entorhinal cortex with a sufficiently narrow tuning may thus be low. Such distributions (with a proportionally small amount of narrowly tuned conjunctive cells) lead to low values in the absolute hexasymmetry. The neural hexasymmetry in this case would be driven by the subset of cells with sufficiently narrow tuning widths. If this causes the neural hexasymmetry to drop below noise levels, the statistical evaluation of this hypothesis would change.” In addition, in Figure 5, we have applied the coupling between preferred head directions and grid axes to only one third of all grid cells (parameter pc= ⅓ in Table 1), following the values reported by Boccara et al. 2010 and Sargolini et al. 2006. To strengthen the link between Figure 5 and Figure 2, we now state the hexasymmetry when using pc= ⅓ along with a ‘realistic’ tuning width and jitter for head-direction modulated grid cells in Figure 2H. Additionally, we performed new simulations where we observed a linear relationship (above the noise floor) between the proportion of conjunctive cells and the hexasymmetry. This shall help the reader understand the effect of a reduced percentage of conjunctive cells on the absolute hexasymmetry values. We have added these results as a new supplementary figure (Figure 2 – figure supplement 2).

Finally, regarding your comment on the findings by Gerlei et al. 2020, we now reference this study in our manuscript and discuss the possible implications (second paragraph of section “A note on our choice of the values of model parameters” of the Discussion): “Additionally, while we assumed that all conjunctive grid cells maintain the same preferred head direction between different firing fields, conjunctive grid cells have also been shown to exhibit different preferred head directions in different firing fields (Gerlei et al., 2020). This could lead to hexadirectional modulation if the different preferred head directions are offset by 60o from each other, but will not give rise to hexadirectional modulation if the preferred head directions are randomly distributed. To the best of our knowledge, the distribution of preferred head directions was not quantified by Gerlei et al. (2020), thus this remains an open question.”

Reviewer #1, comment #4: Finally, a variant of the fourth hypothesis is that the hexasymmetry might be produced by a clustering of head direction preferences across head direction cells similar to that hypothesized in the first hypothesis, but without such cells having to fire in grid patterns. If head direction selectivity is so clustered, who needs the grids? This would explain why hexasymmetry is ubiquitous, and could easily be explored computationally by, in fact, a simplification of the models considered in this study.

Response: We fully agree with you. We now explain this possibility in the Introduction where we introduce the conjunctive grid by head-direction cell hypothesis (fourth paragraph of the Introduction) and return to it in the Discussion (section “Potential mechanisms underlying hexadirectional population signals in the entorhinal cortex”). There, we now also explain that in such a case another mechanism would be needed to ensure that the preferred head directions of head-direction cells exhibit six-fold rotational symmetry.

Reviewer #2 (Public Review):

Reviewer #2, comment #1: Grid cells - originally discovered in single-cell recordings from the rodent entorhinal cortex, and subsequently identified in single-cell recordings from the human brain - are believed to contribute to a range of cognitive functions including spatial navigation, long-term memory function, and inferential reasoning. Following a landmark study by Doeller et al. (Nature, 2010), a plethora of human neuroimaging studies have hypothesised that grid cell population activity might also be reflected in the six-fold (or 'hexadirectional') modulation of the BOLD signal (following the six-fold rotational symmetry exhibited by individual grid cell firing patterns), or in the amplitude of oscillatory activity recorded using MEG or intracranial EEG. The mechanism by which these network-level dynamics might arise from the firing patterns of individual grid cells remains unclear, however.

In this study, Khalid and colleagues use a combination of computational modelling and mathematical analysis to evaluate three competing hypotheses that describe how the hexadirectional modulation of population firing rates (taken as a simple proxy for the BOLD, MEG, or iEEG signal) might arise from the firing patterns of individual grid cells. They demonstrate that all three mechanisms could account for these network-level dynamics if a specific set of conditions relating to the agent's movement trajectory and the underlying properties of grid cell firing patterns are satisfied.

The computational modelling and mathematic analyses presented here are rigorous, clearly motivated, and intuitively described. In addition, these results are important both for the interpretation of hexadirectional modulation in existing data sets and for the design of future experiments and analyses that aim to probe grid cell population activity. As such, this study is likely to have a significant impact on the field by providing a firmer theoretical basis for the interpretation of neuroimaging data. To my mind, the only weakness is the relatively limited focus on the known properties of grid cells in rodent entorhinal cortex, and the network level activity that these firing patterns might be expected to produce under each hypothesis. Strengthening the link with existing neurobiology would further enhance the importance of these results for those hoping to assay grid cell firing patterns in recordings of ensemble-level neural activity.

Response: Thank you very much for reviewing our manuscript and your positive assessment. Following your comments, we have revised the manuscript to more closely link our simulations to known properties of grid cells in the rodent entorhinal cortex.

Reviewer #3 (Public Review):

Reviewer #3, comment #1: This is an interesting and carefully carried out theoretical analysis of potential explanations for hexadirectional modulation of neural population activity that has been reported in the human entorhinal cortex and some other cortical regions. The previously reported hexadirectional modulation is of considerable interest as it has been proposed to be a proxy for the activation of grid cell networks. However, the extent to which this proposal is consistent with the known firing properties of grids hasn't received the attention it perhaps deserves. By comparing the predictions of three different models this study imposes constraints on possible mechanisms and generates predictions that can be tested through future experimentation.

Overall, while the conclusions of the study are convincing, I think the usefulness to the field would be increased if null hypotheses were more carefully considered and if the authors' new metric for hexadirectional modulation (H) could be directly contrasted with previously used metrics. For example, if the effect sizes for hexadirectional modulation in the previous fMRI and EEG data could be more directly compared with those of the models here, then this could help in establishing the extent to which the experimental hexadirectional modulation stands out from path hexasymmetry and how close it comes to the striking modulation observed with the conjunctive models. It could also be helpful to consider scenarios in which hexadirectional modulation is independent of grid firing, for example perhaps with appropriate coordination of head direction cell firing.

Response: Thanks for reviewing our manuscript and for the overall positive assessment. The new Methods section “Implementation of previously used metrics” starts with the following sentences: “We applied three previously used metrics to our framework: the Generalized Linear Model (GLM) method by Doeller et al. 2010; the GLM method with binning by Kunz et al. 2015; and the circular-linear correlation method by Maidenbaum et al. 2018.” We have created a new supplementary figure (Figure 5 – figure supplement 4) in which we compare the results from these other methods to the results of our new method. Overall, the results are highly similar, indicating that all these methods are equally suited to test for a hexadirectional modulation of neural activity.

In section “Implementation of previously used metrics” we then explain: “In brief, in the GLM method (e.g. used in Doeller et al., 2010), the hexasymmetry is found in two steps: the orientation of the hexadirectional modulation is first estimated on the first half of the data by using the regressors and on the time-discrete fMRI activity (Equation 9), with θt being the movement direction of the subject in time step t. The amplitude of the signal is then estimated on the second half of the data using the single regressor , where . The hexasymmetry is then evaluated as .

The GLM method with binning (e.g. used in Kunz et al., 2015) uses the same procedure as the GLM method for estimating the grid orientation in the first half of the data, but the amplitude is estimated differently on the second half by a regressor that has a value 1 if θt is aligned with a peak of the hexadirectional modulation (aligned if , modulo operator) and a value of -1 if θt is misaligned. The hexasymmetry is then calculated from the amplitude in the same way as in the GLM method.

The circular-linear correlation method (e.g. used in Maidenbaum et al., 2018) is similar to the GLM method in that it uses the regressors β1 cos(6θ_t) and β2 on the time-discrete mean activity, but instead of using β1 and β2 to estimate the orientation of the hexadirectional modulation, the beta values are directly used to estimate the hexasymmetry using the relation .”

For each of the three previously used metrics and our new method, we estimated the resulting hexasymmetry (new Figure 5 – figure supplement 4 in the manuscript). In the Methods section “Implementation of previously used metrics” we then continue with our explanations: “Regarding the statistical evaluation, each method evaluates the size of the neural hexasymmetry differently. Specifically, the new method developed in our manuscript compares the neural hexasymmetry to path hexasymmetry to test whether neural hexasymmetry is significantly above path hexasymmetry. For the two generalized linear model (GLM) methods, we compare the hexasymmetry to zero (using the Mann-Whitney U test) to establish significance. Hexasymmetry values can be negative in these approaches, allowing the statistical comparison against 0. Negative values occur when the estimated grid orientation from the first data half does not match the grid orientation from the second data half. Regarding the statistical evaluation of the circular-linear correlation method, we calculated a z-score by comparing each empirical observation of the hexasymmetry to hexasymmetries from a set of surrogate distributions (as in Maidenbaum et al., 2018). We then calculate a p-value by comparing the distribution of z-scores versus zero using a Mann-Whitney U test. We use the z-scores instead of the hexasymmetry for the circular-linear correlation method to match the procedure used in Maidenbaum et al. (2018). We obtained the surrogate distributions by circularly shifting the vector of movement directions relative to the time dependent vector of firing rates. For random walks, the vector is shifted by a random number drawn from a uniform distribution defined with the same length as the number of time points in the vector of movement directions. For the star-like walks and piecewise linear walks, the shift is a random integer multiplied by the number of time points in a linear segment. Circularly shifting the vector of movement directions scrambles the correlations between movement direction and neural activity while preserving their temporal structure.”

The results of these simulations, i.e. the comparison of our new method to previously used metrics, are summarized in Figure 5 – figure supplement 4 and show qualitatively identical findings when using the different methods. We have added this information also to the manuscript in the third paragraph of section “Quantification of hexasymmetry of neural activity and trajectories” of the Methods: “Empirical (fMRI/iEEG) studies (e.g. Doeller et al., 2010; Kunz et al., 2015; Maidenbaum et al., 2018) addressed this problem of trajectories spuriously contributing to hexasymmetry by fitting a Generalized Linear Model (GLM) to the time discrete fMRI/iEEG activity. In contrast, our new approach to hexasymmetry in Equation (12) quantifies the contribution of the path to the neural hexasymmetry explicitly, and has the advantage that it allows an analytical treatment (see next section). Comparing our new method with previous methods for evaluating hexasymmetry led to qualitatively identical statistical effects (Figure 5 – figure supplement 4).” We have also added a pointer to this new supplementary figure in the caption of Figure 5 in the manuscript: “For a comparison between our method and previously used methods for evaluating hexasymmetry, see Figure 5 – figure supplement 4.”

Author Response

Reviewer #1 (Public Review):

This manuscript will interest cognitive scientists, neuroimaging researchers, and neuroscientists interested in the systems-level organization of brain activity. The authors describe four brain states that are present across a wide range of cognitive tasks and determine that the relative distribution of the brain states shows both commonalities and differences across task conditions.

The authors characterized the low-dimensional latent space that has been shown to capture the major features of intrinsic brain activity using four states obtained with a Hidden Markov Model. They related the four states to previously-described functional gradients in the brain and examined the relative contribution of each state under different cognitive conditions. They showed that states related to the measured behavior for each condition differed, but that a common state appears to reflect disengagement across conditions. The authors bring together a state-of-the-art analysis of systemslevel brain dynamics and cognitive neuroscience, bridging a gap that has long needed to be bridged.

The strongest aspect of the study is its rigor. The authors use appropriate null models and examine multiple datasets (not used in the original analysis) to demonstrate that their findings replicate. Their thorough analysis convincingly supports their assertion that common states are present across a variety of conditions, but that different states may predict behavioural measures for different conditions. However, the authors could have better situated their work within the existing literature. It is not that a more exhaustive literature review is needed-it is that some of their results are unsurprising given the work reported in other manuscripts; some of their work reinforces or is reinforced by prior studies; and some of their work is not compared to similar findings obtained with other analysis approaches. While space is not unlimited, some of these gaps are important enough that they are worth addressing:

We appreciate the reviewer’s thorough read of our manuscript and positive comments on its rigor and implications. We agree that the original version of the manuscript insufficiently situated this work in the existing literature. We have made extensive revisions to better place our findings in the context of prior work. These changes are described in detail below.

1) The authors' own prior work on functional connectivity signatures of attention is not discussed in comparison to the latest work. Neither is work from other groups showing signatures of arousal that change over time, particularly in resting state scans. Attention and arousal are not the same things, but they are intertwined, and both have been linked to large-scale changes in brain activity that should be captured in the HMM latent states. The authors should discuss how the current work fits with existing studies.

Thank you for raising this point. We agree that the relationship between low-dimensional latent states and predefined activity and functional connectivity signatures is an important and interesting question in both attention research and more general contexts. Here, we did not empirically relate the brain states examined in this study and functional connectivity signatures previously investigated in our lab (e.g., Rosenberg et al., 2016; Song et al., 2021a) because the research question and methodological complexities deserved separate attention that go beyond the scope of this paper. Therefore, we conceptually addressed the reviewer’s question on how functional connectivity signatures of attention are related to the brain states that were observed here. Next, we asked how arousal relates to the brain states by indirectly predicting arousal levels of each brain state based on its activity patterns’ spatial resemblance to the predefined arousal network template (Goodale et al., 2021).

Latent states and dynamic functional connectivity

Previous work suggested that, on medium time scales (~20-60 seconds), changes in functional connectivity signatures of sustained attention (Rosenberg et al., 2020) and narrative engagement (Song et al., 2021a) predicted changes in attentional states. How do these attention-related functional connectivity dynamics relate to latent state dynamics, measured on a shorter time scale (1 second)?

Theoretically, there are reasons to think that these measures are related but not redundant. Both HMM and dynamic functional connectivity provide summary measures of the whole-brain functional interactions that evolve over time. Whereas HMM identifies recurring low-dimensional brain states, dynamic functional connectivity used in our and others’ prior studies captures high-dimensional dynamical patterns. Furthermore, while the mixture Gaussian function utilized to infer emission probability in our HMM infers the states from both the BOLD activity patterns and their interactions, functional connectivity considers only pairwise interactions between regions of interests. Thus, with a theoretical ground that the brain states can be characterized at multiple scales and different methods (Greene et al., 2023), we can hypothesize that the both measures could (and perhaps, should be able to) capture brain-wide latent state changes. For example, if we were to apply kmeans clustering methods on the sliding window-based dynamic functional connectivity as in Allen et al. (2014), the resulting clusters could arguably be similar to the latent states derived from the HMM.

However, there are practical reasons why the correspondence between our prior dynamic functional connectivity models and current HMM states is difficult to test directly. A time point-bytime point matching of the HMM state sequence and dynamic functional connectivity is not feasible because, in our prior work, dynamic functional connectivity was measured in a sliding time window (~20-60 seconds), whereas the HMM state identification is conducted at every TR (1 second). An alternative would be to concatenate all time points that were categorized as each HMM state to compute representative functional connectivity of that state. This “splicing and concatenating” method, however, disrupts continuous BOLD-signal time series and has not previously been validated for use with our dynamic connectome-based predictive models. In addition, the difference in time series lengths across states would make comparisons of the four states’ functional connectomes unfair.

One main focus of our manuscript was to relate brain dynamics (HMM state dynamics) to static manifold (functional connectivity gradients). We agree that a direct link between two measures of brain dynamics, HMM and dynamic functional connectivity, is an important research question. However, due to some intricacies that needed to be addressed to answer this question, we felt that it was beyond the scope of our paper. We are eager, however, to explore these comparisons in future work which can more thoroughly address the caveats associated with comparing models of sustained attention, narrative engagement, and arousal defined using different input features and methods.

Arousal, attention, and latent neural state dynamics

Next, the reviewer posed an important question about the relationship between arousal, attention, and latent states. The current study was designed to assess the relationship between attention and latent state dynamics. However, previous neuroimaging work showed that low-dimensional brain dynamics reflect fluctuations in arousal (Raut et al., 2021; Shine et al., 2016; Zhang et al., 2023). Behavioral studies showed that attention and arousal hold a non-linear relationship, for example, mind-wandering states are associated with lower arousal and externally distracted states are associated with higher arousal, when both these states indicate low attention (Esterman and Rothlein, 2019; Unsworth and Robison, 2018, 2016).

To address the reviewer’s suggestion, we wanted to test if our brain states reflected changes in arousal, but we did not collect relevant behavioral or physiological measures. Therefore, to indirectly test for relationships, we predicted levels of arousal in brain states by applying the “arousal network template” defined by Dr. Catie Chang’s group (Chang et al., 2016; Falahpour et al., 2018; Goodale et al., 2021). The arousal network template was created from resting-state fMRI data to predict arousal levels indicated by eye monitoring and electrophysiological signals. In the original study, the arousal level at each time point was predicted by the correlation between the BOLD activity patterns of each TR to the arousal template. The more similar the whole-brain activation pattern was to the arousal network template, the higher the participant was predicted to be aroused at that moment. This activity pattern-based model was generalized to fMRI data during tasks (Goodale et al., 2021).

We correlated the arousal template to the activity patterns of the four brain states that were inferred by the HMM. The DMN state was positively correlated with the arousal template (r=0.264) and the SM state was negatively correlated with the arousal template (r=-0.303) (Author response image 1). These values were not tested for significance because they were single observations. While speculative, this may suggest that participants are in a high arousal state during the DMN state and a low arousal state during the SM state. Together with our results relating brain states to attention, it is possible that the SM state is a common state indicating low arousal and low attention. On the other hand, the DMN state, a signature of a highly aroused state, may benefit gradCPT task performance but not necessarily in engaging with a sitcom episode. However, because this was a single observation and we did not collect a physiological measure of arousal to validate this indirect prediction result, we did not include the result in the manuscript. We hope to more directly test this question in future work with behavioral and physiological measures of arousal.

Author response image 1.

Changes made to the manuscript

Importantly, we agree with the reviewer that a theoretical discussion about the relationships between functional connectivity, latent states, gradients, as well as attention and arousal was a critical omission from the original Discussion. We edited the Discussion to highlight past literature on these topics and encourage future work to investigate these relationships.

[Manuscript, page 11] “Previous studies showed that large-scale neural dynamics that evolve over tens of seconds capture meaningful variance in arousal (Raut et al., 2021; Zhang et al., 2023) and attentional states (Rosenberg et al., 2020; Yamashita et al., 2021). We asked whether latent neural state dynamics reflect ongoing changes in attention in both task and naturalistic contexts.”

[Manuscript, page 17] “Previous work showed that time-resolved whole-brain functional connectivity (i.e., paired interactions of more than a hundred parcels) predicts changes in attention during task performance (Rosenberg et al., 2020) as well as movie-watching and story-listening (Song et al., 2021a). Future work could investigate whether functional connectivity and the HMM capture the same underlying “brain states” to bridge the results from the two literatures. Furthermore, though the current study provided evidence of neural state dynamics reflecting attention, the same neural states may, in part, reflect fluctuations in arousal (Chang et al., 2016; Zhang et al., 2023). Complementing behavioral studies that demonstrated a nonlinear relationship between attention and arousal (Esterman and Rothlein, 2019; Unsworth and Robison, 2018, 2016), future studies collecting behavioral and physiological measures of arousal can assess the extent to which attention explains neural state dynamics beyond what can be explained by arousal fluctuations.”

2) The 'base state' has been described in a number of prior papers (for one early example, see https://pubmed.ncbi.nlm.nih.gov/27008543). The idea that it might serve as a hub or intermediary for other states has been raised in other studies, and discussion of the similarity or differences between those studies and this one would provide better context for the interpretation of the current work. One of the intriguing findings of the current study is that the incidence of this base state increases during sitcom watching, the strongest evidence to date is that it has a cognitive role and is not merely a configuration of activity that the brain must pass through when making a transition.

We greatly appreciate the reviewer’s suggestion of prior papers. We were not aware of previous findings of the base state at the time of writing the manuscript, so it was reassuring to see consistent findings. In the Discussion, we highlighted the findings of Chen et al. (2016) and Saggar et al. (2022). Both studies highlighted the role of the base state as a “hub”-like transition state. However, as the reviewer noted, these studies did not address the functional relevance of this state to cognitive states because both were based on resting-state fMRI.

In our revised Discussion, we write that our work replicates previous findings of the base state that consistently acted as a transitional hub state in macroscopic brain dynamics. We also note that our study expands this line of work by characterizing what functional roles the base state plays in multiple contexts: The base state indicated high attentional engagement and exhibited the highest occurrence proportion as well as longest dwell times during naturalistic movie watching. The base state’s functional involvement was comparatively minor during controlled tasks.

[Manuscript, page 17-18] “Past resting-state fMRI studies have reported the existence of the base state. Chen et al. (2016) used the HMM to detect a state that had “less apparent activation or deactivation patterns in known networks compared with other states”. This state had the highest occurrence probability among the inferred latent states, was consistently detected by the model, and was most likely to transition to and from other states, all of which mirror our findings here. The authors interpret this state as an “intermediate transient state that appears when the brain is switching between other more reproducible brain states”. The observation of the base state was not confined to studies using HMMs. Saggar et al. (2022) used topological data analysis to represent a low-dimensional manifold of resting-state whole-brain dynamics as a graph, where each node corresponds to brain activity patterns of a cluster of time points. Topologically focal “hub” nodes were represented uniformly by all functional networks, meaning that no characteristic activation above or below the mean was detected, similar to what we observe with the base state. The transition probability from other states to the hub state was the highest, demonstrating its role as a putative transition state.

However, the functional relevance of the base state to human cognition had not been explored previously. We propose that the base state, a transitional hub (Figure 2B) positioned at the center of the gradient subspace (Figure 1D), functions as a state of natural equilibrium. Transitioning to the DMN, DAN, or SM states reflects incursion away from natural equilibrium (Deco et al., 2017; Gu et al., 2015), as the brain enters a functionally modular state. Notably, the base state indicated high attentional engagement (Figure 5E and F) and exhibited the highest occurrence proportion (Figure 3B) as well as the longest dwell times (Figure 3—figure supplement 1) during naturalistic movie watching, whereas its functional involvement was comparatively minor during controlled tasks. This significant relevance to behavior verifies that the base state cannot simply be a byproduct of the model. We speculate that susceptibility to both external and internal information is maximized in the base state—allowing for roughly equal weighting of both sides so that they can be integrated to form a coherent representation of the world—at the expense of the stability of a certain functional network (Cocchi et al., 2017; Fagerholm et al., 2015). When processing rich narratives, particularly when a person is fully immersed without having to exert cognitive effort, a less modular state with high degrees of freedom to reach other states may be more likely to be involved. The role of the base state should be further investigated in future studies.”

3) The link between latent states and functional connectivity gradients should be considered in the context of prior work showing that the spatiotemporal patterns of intrinsic activity that account for most of the structure in resting state fMRI also sweep across functional connectivity gradients (https://pubmed.ncbi.nlm.nih.gov/33549755/). In fact, the spatiotemporal dynamics may give rise to the functional connectivity gradients (https://pubmed.ncbi.nlm.nih.gov/35902649/). HMM states bear a marked resemblance to the high-activity phases of these patterns and are likely to be closely linked to them. The spatiotemporal patterns are typically obtained during rest, but they have been reported during task performance (https://pubmed.ncbi.nlm.nih.gov/30753928/) which further suggests a link to the current work. Similar patterns have been observed in anesthetized animals, which also reinforces the conclusion of the current work that the states are fundamental aspects of the brain's functional organization.

We appreciate the comments that relate spatiotemporal patterns, functional connectivity gradients, and the latent states derived from the HMM. Our work was also inspired by the papers that the reviewer suggested, especially Bolt et al.’s (2022), which compared the results of numerous dimensionality and clustering algorithms and suggested three spatiotemporal patterns that seemed to be commonly supported across algorithms. We originally cited these studies throughout the manuscript, but did not discuss them comprehensively. We have revised the Discussion to situate our findings on past work that used resting-state fMRI to study low-dimensional latent brain states.

[Manuscript, page 15-16] “This perspective is supported by previous work that has used different methods to capture recurring low-dimensional states from spontaneous fMRI activity during rest. For example, to extract time-averaged latent states, early resting-state analyses identified task-positive and tasknegative networks using seed-based correlation (Fox et al., 2005). Dimensionality reduction algorithms such as independent component analysis (Smith et al., 2009) extracted latent components that explain the largest variance in fMRI time series. Other lines of work used timeresolved analyses to capture latent state dynamics. For example, variants of clustering algorithms, such as co-activation patterns (Liu et al., 2018; Liu and Duyn, 2013), k-means clustering (Allen et al., 2014), and HMM (Baker et al., 2014; Chen et al., 2016; Vidaurre et al., 2018, 2017), characterized fMRI time series as recurrences of and transitions between a small number of states. Time-lag analysis was used to identify quasiperiodic spatiotemporal patterns of propagating brain activity (Abbas et al., 2019; Yousefi and Keilholz, 2021). A recent study extensively compared these different algorithms and showed that they all report qualitatively similar latent states or components when applied to fMRI data (Bolt et al., 2022). While these studies used different algorithms to probe data-specific brain states, this work and ours report common latent axes that follow a long-standing theory of large-scale human functional systems (Mesulam, 1998). Neural dynamics span principal axes that dissociate unimodal to transmodal and sensory to motor information processing systems.”

Reviewer #2 (Public Review):

In this study, Song and colleagues applied a Hidden Markov Model to whole-brain fMRI data from the unique SONG dataset and a grad-CPT task, and in doing so observed robust transitions between lowdimensional states that they then attributed to specific psychological features extracted from the different tasks.

The methods used appeared to be sound and robust to parameter choices. Whenever choices were made regarding specific parameters, the authors demonstrated that their approach was robust to different values, and also replicated their main findings on a separate dataset.

I was mildly concerned that similarities in some of the algorithms used may have rendered some of the inter-measure results as somewhat inevitable (a hypothesis that could be tested using appropriate null models).

This work is quite integrative, linking together a number of previous studies into a framework that allows for interesting follow-up questions.

Overall, I found the work to be robust, interesting, and integrative, with a wide-ranging citation list and exciting implications for future work.

We appreciate the reviewer’s comments on the study’s robustness and future implications. Our work was highly motivated by the reviewer’s prior work.

Reviewer #3 (Public Review):

My general assessment of the paper is that the analyses done after they find the model are exemplary and show some interesting results. However, the method they use to find the number of states (Calinski-Harabasz score instead of log-likelihood), the model they use generally (HMM), and the fact that they don't show how they find the number of states on HCP, with the Schaeffer atlas, and do not report their R^2 on a test set is a little concerning. I don't think this perse impedes their results, but it is something that they can improve. They argue that the states they find align with long-standing ideas about the functional organization of the brain and align with other research, but they can improve their selection for their model.

We appreciate the reviewer’s thorough read of the paper, evaluation of our analyses linking brain states to behavior as “exemplary”, and important questions about the modeling approach. We have included detailed responses below and updated the manuscript accordingly.

Strengths:

• Use multiple datasets, multiple ROIs, and multiple analyses to validate their results

• Figures are convincing in the sense that patterns clearly synchronize between participants

• Authors select the number of states using the optimal model fit (although this turns out to be a little more questionable due to what they quantify as 'optimal model fit')

We address this concern on page 30-31 of this response letter.

• Replication with Schaeffer atlas makes results more convincing

• The analyses around the fact that the base state acts as a flexible hub are well done and well explained

• Their comparison of synchrony is well-done and comparing it to resting-state, which does not have any significant synchrony among participants is obvious, but still good to compare against.

• Their results with respect to similar narrative engagement being correlated with similar neural state dynamics are well done and interesting.

• Their results on event boundaries are compelling and well done. However, I do not find their Chang et al. results convincing (Figure 4B), it could just be because it is a different medium that explains differences in DMN response, but to me, it seems like these are just altogether different patterns that can not 100% be explained by their method/results.

We entirely agree with the reviewer that the Chang et al. (2021) data are different in many ways from our own SONG dataset. Whereas data from Chang et al. (2021) were collected while participants listened to an audio-only narrative, participants in the SONG sample watched and listened to audiovisual stimuli. They were scanned at different universities in different countries with different protocols by different research groups for different purposes. That is, there are numerous reasons why we would expect the model should not generalize. Thus, we found it compelling and surprising that, despite all of these differences between the datasets, the model trained on the SONG dataset generalized to the data from Chang et al. (2021). The results highlighted a robust increase in the DMN state occurrence and a decrease in the base state occurrence after the narrative event boundaries, irrespective of whether the stimulus was an audiovisual sitcom episode or a narrated story. This external model validation was a way that we tested the robustness of our own model and the relationship between neural state dynamics and cognitive dynamics.

• Their results that when there is no event, transition into the DMN state comes from the base state is 50% is interesting and a strong result. However, it is unclear if this is just for the sitcom or also for Chang et al.'s data.

We apologize for the lack of clarity. We show the statistical results of the two sitcom episodes as well as Chang et al.’s (2021) data in Figure 4—figure supplement 2 in our original manuscript. Here, we provide the exact values of the base-to-DMN state transition probability, and how they differ across moments after event boundaries compared to non-event boundaries.

For sitcom episode 1, the probability of base-to-DMN state transition was 44.6 ± 18.8 % at event boundaries whereas 62.0 ± 10.4 % at non-event boundaries (FDR-p = 0.0013). For sitcom episode 2, the probability of base-to-DMN state transition was 44.1 ± 18.0 % at event boundaries whereas 62.2 ± 7.6 % at non-event boundaries (FDR-p = 0.0006). For the Chang et al. (2021) dataset, the probability of base-to-DMN state transition was 33.3 ± 15.9 % at event boundaries whereas 58.1 ± 6.4 % at non-event boundaries (FDR-p < 0.0001). Thus, our result, “At non-event boundaries, the DMN state was most likely to transition from the base state, accounting for more than 50% of the transitions to the DMN state” (pg 11, line 24-25), holds true for both the internal and external datasets.

• The involvement of the base state as being highly engaged during the comedy sitcom and the movie are interesting results that warrant further study into the base state theory they pose in this work.

• It is good that they make sure SM states are not just because of head motion (P 12).

• Their comparison between functional gradient and neural states is good, and their results are generally well-supported, intuitive, and interesting enough to warrant further research into them. Their findings on the context-specificity of their DMN and DAN state are interesting and relate well to the antagonistic relationship in resting-state data.

Weaknesses:

• Authors should train the model on part of the data and validate on another

Thank you for raising this issue. To the best of our knowledge, past work that applied the HMM to the fMRI data has conducted training and inference on the same data, including initial work that implemented HMM on the resting-state fMRI (Baker et al., 2014; Chen et al., 2016; Vidaurre et al., 2018, 2017) as well as more recent work that applied HMMs to the task or movie-watching fMRI (Cornblath et al., 2020; Taghia et al., 2018; van der Meer et al., 2020; Yamashita et al., 2021). That is, the parameters—emission probability, transition probability, and initial probability—were estimated from the entire dataset and the latent state sequence was inferred using the Viterbi algorithm on the same dataset.

However, we were also aware of the potential problem this may have. Therefore, in our recent work asking a different research question in another fMRI dataset (Song et al., 2021b), we trained an HMM on a subset of the dataset (moments when participants were watching movie clips in the original temporal order) and inferred latent state sequence of the fMRI time series in another subset of the dataset (moments when participants were watching movie clips in a scrambled temporal order). To the best of our knowledge, this was the first paper that used different segments of the data to fit and infer states from the HMM.

In the current study, we wanted to capture brain states that underlie brain activity across contexts. Thus, we presented the same-dataset training and inference procedure as our primary result. However, for every main result, we also showed results where we separated the data used for model fitting and state inference. That is, we fit the HMM on the SONG dataset, primarily report the inference results on the SONG dataset, but also report inference on the external datasets that were not included in model fitting. The datasets used were the Human Connectome Project dataset (Van Essen et al., 2013), Chang et al. (2021) audio-listening dataset, Rosenberg et al. (2016) gradCPT dataset, and Chen et al. (2017) Sherlock dataset.

However, to further address the concern of the reviewer whether the HMM fit is reliable when applied to held-out data, we computed the reliability of the HMM inference by conducting crossvalidations and split-half reliability analysis.

(1) Cross-validation

To separate the dataset used for HMM training and inference, we conducted cross-validation on the SONG dataset (N=27) by training the model with the data from 26 participants and inferring the latent state sequence of the held-out participant.

First, we compared the robustness of the model training by comparing the mean activity patterns of the four latent states fitted at the group level (N=27) with the mean activity patterns of the four states fitted across cross-validation folds. Pearson’s correlations between the group-level vs. cross-validated latent states’ mean activity patterns were r = 0.991 ± 0.010, with a range from 0.963 to 0.999.

Second, we compared the robustness of model inference by comparing the latent state sequences that were inferred at the group level vs. from held-out participants in a cross-validation scheme. All fMRI conditions had mean similarity higher than 90%; Rest 1: 92.74 ± 5.02 %, Rest2: 92.74 ± 4.83 %, GradCPT face: 92.97 ± 6.41 %, GradCPT scene: 93.27 ± 5.76 %, Sitcom ep1: 93.31 ± 3.92 %, Sitcom ep2: 93.13 ± 4.36 %, Documentary: 92.42 ± 4.72 %.

Third, with the latent state sequences inferred from cross-validation, we replicated the analysis of Figure 3 to test for synchrony of the latent state sequences across participants. The crossvalidated results were highly similar to manuscript Figure 3, which was generated from the grouplevel analysis. Mean synchrony of latent state sequences are as follows: Rest 1: 25.90 ± 3.81%, Rest 2: 25.75 ± 4.19 %, GradCPT face: 27.17 ± 3.86 %, GradCPT scene: 28.11 ± 3.89 %, Sitcom ep1: 40.69 ± 3.86%, Sitcom ep2: 40.53 ± 3.13%, Documentary: 30.13 ± 3.41%.

Author response image 2.

(2) Split-half reliability

To test for the internal robustness of the model, we randomly assigned SONG dataset participants into two groups and conducted HMM separately in each. Similarity (Pearson’s correlation) between the two groups’ activation patterns were DMN: 0.791, DAN: 0.838, SM: 0.944, base: 0.837. The similarity of the covariance patterns were DMN: 0.995, DAN: 0.996, SM: 0.994, base: 0.996.

Author response image 3.

We further validated the split-half reliability of the model using the HCP dataset, which contains data of a larger sample (N=119). Similarity (Pearson’s correlation) between the two groups’ activation patterns were DMN: 0.998, DAN: 0.997, SM: 0.993, base: 0.923. The similarity of the covariance patterns were DMN: 0.995, DAN: 0.996, SM: 0.994, base: 0.996.

Together the cross-validation and split-half reliability results demonstrate that the HMM results reported in the manuscript are reliable and robust to the way we conducted the analysis. The result of the split-half reliability analysis is added in the Results.

[Manuscript, page 3-4] “Neural state inference was robust to the choice of 𝐾 (Figure 1—figure supplement 1) and the fMRI preprocessing pipeline (Figure 1—figure supplement 5) and consistent when conducted on two groups of randomly split-half participants (Pearson’s correlations between the two groups’ latent state activation patterns: DMN: 0.791, DAN: 0.838, SM: 0.944, base: 0.837).”

• Comparison with just PCA/functional gradients is weak in establishing whether HMMs are good models of the timeseries. Especially given that the HMM does not explain a lot of variance in the signal (~0.5 R^2 for only 27 brain regions) for PCA. I think they don't report their own R^2 of the timeseries

We agree with the reviewer that the PCA that we conducted to compare with the explained variance of the functional gradients was not directly comparable because PCA and gradients utilize different algorithms to reduce dimensionality. To make more meaningful comparisons, we removed the data-specific PCA results and replaced them with data-specific functional gradients (derived from the SONG dataset). This allows us to directly compare SONG-specific functional gradients with predefined gradients (derived from the resting-state HCP dataset from Margulies et al. [2016]). We found that the degrees to which the first two predefined gradients explained whole-brain fMRI time series (SONG: 𝑟! = 0.097, HCP: 0.084) were comparable to the amount of variance explained by the first two data-specific gradients (SONG: 𝑟! = 0.100, HCP: 0.086). Thus, the predefined gradients explain as much variance in the SONG data time series as SONG-specific gradients do. This supports our argument that the low-dimensional manifold is largely shared across contexts, and that the common HMM latent states may tile the predefined gradients.

These analyses and results were added to the Results, Methods, and Figure 1—figure supplement 8. Here, we only attach changes to the Results section for simplicity, but please see the revised manuscript for further changes.

[Manuscript, page 5-6] “We hypothesized that the spatial gradients reported by Margulies et al. (2016) act as a lowdimensional manifold over which large-scale dynamics operate (Bolt et al., 2022; Brown et al., 2021; Karapanagiotidis et al., 2020; Turnbull et al., 2020), such that traversals within this manifold explain large variance in neural dynamics and, consequently, cognition and behavior (Figure 1C). To test this idea, we situated the mean activity values of the four latent states along the gradients defined by Margulies et al. (2016) (see Methods). The brain states tiled the two-dimensional gradient space with the base state at the center (Figure 1D; Figure1—figure supplement 7). The Euclidean distances between these four states were maximized in the two-dimensional gradient space, compared to a chance where the four states were inferred from circular-shifted time series (p < 0.001). For the SONG dataset, the DMN and SM states fell at more extreme positions of the primary gradient than expected by chance (both FDR-p values = 0.004; DAN and SM states, FDRp values = 0.171). For the HCP dataset, the DMN and DAN states fell at more extreme positions on the primary gradient (both FDR-p values = 0.004; SM and base states, FDR-p values = 0.076). No state was consistently found at the extremes of the secondary gradient (all FDR-p values > 0.021).

We asked whether the predefined gradients explain as much variance in neural dynamics as latent subspace optimized for the SONG dataset. To do so, we applied the same nonlinear dimensionality reduction algorithm to the SONG dataset’s ROI time series. Of note, the SONG dataset includes 18.95% rest, 15.07% task, and 65.98% movie-watching data whereas the data used by Margulies et al. (2016) was 100% rest. Despite these differences, the SONG-specific gradients closely resembled the predefined gradients, with significant Pearson’s correlations observed for the first (r = 0.876) and second (r = 0.877) gradient embeddings (Figure 1—figure supplement 8). Gradients identified with the HCP data also recapitulated Margulies et al.’s (2016) first (r = 0.880) and second (r = 0.871) gradients. We restricted our analysis to the first two gradients because the two gradients together explained roughly 50% of the entire variance of functional brain connectome (SONG: 46.94%, HCP: 52.08%), and the explained variance dropped drastically from the third gradients (more than 1/3 drop compared to second gradients). The degrees to which the first two predefined gradients explained whole-brain fMRI time series (SONG: 𝑟! = 0.097, HCP: 0.084) were comparable to the amount of variance explained by the first two data-specific gradients (SONG: 𝑟! = 0.100, HCP: 0.086; Figure 1—figure supplement 8). Thus, the low-dimensional manifold captured by Margulies et al. (2016) gradients is highly replicable, explaining brain activity dynamics as well as data-specific gradients, and is largely shared across contexts and datasets. This suggests that the state space of whole-brain dynamics closely recapitulates low-dimensional gradients of the static functional brain connectome.”

The reviewer also pointed out that the PCA-gradient comparison was weak in establishing whether HMMs are good models of the time series. However, we would like to point out that the purpose of the comparison was not to validate the performance of the HMM. Instead, we wanted to test whether the gradients introduced by Margulies et al. (2016) could act as a generalizable lowdimensional manifold of brain state dynamics. To argue that the predefined gradients are a shared manifold, these gradients should explain SONG data fMRI time series as much as the principal components derived directly from the SONG data. Our results showed comparable 𝑟!, both in predefined gradient vs. data-specific PC comparisons and predefined gradient vs. data-specific gradient comparisons, which supported our argument that the predefined gradients could be the shared embedding space across contexts and datasets.

The reviewer pointed out that the 𝑟2 of ~0.5 is not explaining enough variance in the fMRI signal. However, we respectfully disagree with this point because there is no established criterion for what constitutes a high or low 𝑟2 for this type of analysis. Of note, previous literature that also applied PCA to fMRI time series (Author response image 4A and 4B) (Lynn et al., 2021; Shine et al., 2019) also found that the cumulative explained variance of top 5 principal components is around 50%. Author response image 4C shows cumulative variances to which gradients explain the functional connectome of the resting-state fMRI data (Margulies et al., 2016).

Author response image 4.

Finally, the reviewer pointed out that the 𝑟! of the HMM-derived latent sequence to the fMRI time series should be reported. However, there is no standardized way of measuring the explained variance of the HMM inference. There is no report of explained variance in the traditional HMMfMRI papers (Baker et al., 2014; Chen et al., 2016; Vidaurre et al., 2018, 2017). Rather than 𝑟!, the HMM computes the log likelihood of the model fit. However, because log likelihood values are dependent on the number of data points, studies do not report log likelihood values nor do they use these metrics to interpret the goodness of model fit.

To ask whether the goodness of the HMM fit was significant above chance, we compared the log likelihood of the HMM to the log likelihood distribution of the null HMM fits. First, we extracted the log likelihood of the HMM fit with the real fMRI time series. We iterated this 1,000 times when calculating null HMMs using the circular-shifted fMRI time series. The log likelihood of the real model was significantly higher than the chance distribution, with a z-value of 2182.5 (p < 0.001). This indicates that the HMM explained a large variance in our fMRI time series data, significantly above chance.

• Authors do not specify whether they also did cross-validation for the HCP dataset to find 4 clusters

We apologize for the lack of clarity. When we computed the Calinski-Harabasz score with the HCP dataset, three was chosen as the most optimal number of states (Author response image 5A). When we set K as 3, the HMM inferred the DMN, DAN, and SM states (Author response image 5C). The base state was included when K was set to 4 (Author response image 5B). The activation pattern similarities of the DMN, DAN, and SM states were r = 0.981, 0.984, 0.911 respectively.

Author response image 5.

We did not use K = 3 for the HCP data replication because we were not trying to test whether these four states would be the optimal set of states in every dataset. Although the CalinskiHarabasz score chose K = 3 because it showed the best clustering performance, this does not mean that the base state is not meaningful to this dataset. Likewise, the latent states that are inferred when we increase/decrease the number of states are also meaningful states. For example, in Figure 1—figure supplement 1, we show an example of the SONG dataset’s latent states when we set K to 7. The seven latent states included the DAN, SM, and base states, the DMN state was subdivided into DMN-A and DMN-B states, and the FPN state and DMN+VIS state were included. Setting a higher number of states like K = 7 would mean that we are capturing brain state dynamics in a higher dimension than when using K = 4. Because we are utilizing a higher number of states, a model set to K = 7 would inevitably capture a larger variance of fMRI time series than a model set to K = 4.

The purpose of latent state replication with the HCP dataset was to validate the generalizability of the DMN, DAN, SM, and base states. Before characterizing these latent states’ relevance to cognition, we needed to verify that these latent states were not simply overfit to the SONG dataset. The fact that the HMM revealed a similar set of latent states when applied to the HCP dataset suggested that the states were not merely specific to SONG data.

To make our points clearer in the manuscript, we emphasized that we are not arguing for the four states to be the exclusive states. We made edits to Discussion as follows.

[Manuscript, page 16] “Our study adopted the assumption of low dimensionality of large-scale neural systems, which led us to intentionally identify only a small number of states underlying whole-brain dynamics. Importantly, however, we do not claim that the four states will be the optimal set of states in every dataset and participant population. Instead, latent states and patterns of state occurrence may vary as a function of individuals and tasks (Figure 1—figure supplement 2). Likewise, while the lowest dimensions of the manifold (i.e., the first two gradients) were largely shared across datasets tested here, we do not argue that it will always be identical. If individuals and tasks deviate significantly from what was tested here, the manifold may also differ along with changes in latent states (Samara et al., 2023). Brain systems operate at different dimensionalities and spatiotemporal scales (Greene et al., 2023), which may have different consequences for cognition. Asking how brain states and manifolds—probed at different dimensionalities and scales—flexibly reconfigure (or not) with changes in contexts and mental states is an important research question for understanding complex human cognition.”

• One of their main contributions is the base state but the correlation between the base state in their Song dataset and the HCP dataset is only 0.399

This is a good point. However, there is precedent for lower spatial pattern correlation of the base state compared to other states in the literature.

Compared to the DMN, DAN, and SM states, the base state did not show characteristic activation or deactivation of functional networks. Most of the functional networks showed activity levels close to the mean (z = 0). With this flattened activation pattern, relatively low activation pattern similarity was observed between the SONG base state and the HCP base state.

In Figure 1—figure supplement 6, we write, “The DMN, DAN, and SM states showed similar mean activity patterns. We refrained from making interpretations about the base state’s activity patterns because the mean activity of most of the parcels was close to z = 0”.

A similar finding has been reported in a previous work by Chen et al. (2016) that discovered the base state with HMM. State 9 (S9) of their results is comparable to our base state. They report that even though the spatial correlation coefficient of the brain state from the split-half reliability analysis was the lowest for S9 due to its low degrees of activation or deactivation, S9 was stably inferred by the HMM. The following is a direct quote from their paper:

“To the best of our knowledge, a state similar to S9 has not been presented in previous literature. We hypothesize that S9 is the “ground” state of the brain, in which brain activity (or deactivity) is similar for the entire cortex (no apparent activation or deactivation as shown in Fig. 4). Note that different groups of subjects have different spatial patterns for state S9 (Fig. 3A). Therefore, S9 has the lowest reproducible spatial pattern (Fig. 3B). However, its temporal characteristics allowed us to distinguish it consistently from other states.” (Chen et al., 2016)

Thus, we believe our data and prior results support the existence of the “base state”.

• Figure 1B: Parcellation is quite big but there seems to be a gradient within regions

This is a function of the visualization software. Mean activity (z) is the same for all voxels within a parcel. To visualize the 3D contours of the brain, we chose an option in the nilearn python function that smooths the mean activity values based on the surface reconstructed anatomy.

In the original manuscript, our Methods write, “The brain surfaces were visualized with nilearn.plotting.plot_surf_stat_map. The parcel boundaries in Figure 1B are smoothed from the volume-to-surface reconstruction.”

• Figure 1D: Why are the DMNs further apart between SONG and HCP than the other states

To address this question, we first tested whether the position of the DMN states in the gradient space is significantly different for the SONG and HCP datasets. We generated surrogate HMM states from the circular-shifted fMRI time series and positioned the four latent states and the null DMN states in the 2-dimensional gradient space (Author response image 6).

Author response image 6.

We next tested whether the Euclidean distance between the SONG dataset’s DMN state and the HCP dataset’s DMN state is larger than would be expected by chance (Author response image 7). To do so, we took the difference between the DMN state positions and compared it to the 1,000 differences generated from the surrogate latent states. The DMN states of the SONG and HCP datasets did not significantly differ in the Gradient 1 dimension (two-tailed test, p = 0.794). However, as the reviewer noted, the positions differed significantly in the Gradient 2 dimension (p = 0.047). The DMN state leaned more towards the Visual gradient in the SONG dataset, whereas it leaned more towards the Somatosensory-Motor gradient in the HCP dataset.

Author response image 7.

Though we cannot claim an exact reason for this across-dataset difference, we note a distinctive difference between the SONG and HCP datasets. Both datasets largely included resting-state, controlled tasks, and movie watching. The SONG dataset included 18.95% of rest, 15.07% of task, and 65.98% of movie watching. The task only contained the gradCPT, i.e., sustained attention task. On the other hand, the HCP dataset included 52.71% of rest, 24.35% of task, and 22.94% of movie watching. There were 7 different tasks included in the HCP dataset. It is possible that different proportions of rest, task, and movie watching, and different cognitive demands involved with each dataset may have created data-specific latent states.

• Page 5 paragraph starting at L25: Their hypothesis that functional gradients explain large variance in neural dynamics needs to be explained more, is non-trivial especially because their R^2 scores are so low (Fig 1. Supplement 8) for PCA

We address this concern on page 21-23 of this response letter.

• Generally, I do not find the PCA analysis convincing and believe they should also compare to something like ICA or a different model of dynamics. They do not explain their reasoning behind assuming an HMM, which is an extremely simplified idea of brain dynamics meaning they only change based on the previous state.

We appreciate this perspective. We replaced the Margulies et al.’s (2016) gradient vs. SONGspecific PCA comparison with a more direct Margulies et al.’s (2016) gradient vs. SONG-specific gradient comparison as described on page 21-23 of this response letter.

More broadly, we elected to use HMM because of recent work showing correspondence between low-dimensional HMM states and behavior (Cornblath et al., 2020; Taghia et al., 2018; van der Meer et al., 2020; Yamashita et al., 2021). We also found the model’s assumption—a mixture Gaussian emission probability and first-order Markovian transition probability—to be the most suited to analyzing the fMRI time series data. We do not intend to claim that other data-reduction techniques would not also capture low-dimensional, behaviorally relevant changes in brain activity. Instead, our primary focus was identifying a set of latent states that generalize (i.e., recur) across multiple contexts and understanding how those states reflect cognitive and attentional states.

Although a comparison of possible data-reduction algorithms is out of the scope of the current work, an exhaustive comparison of different models can be found in Bolt et al. (2022). The authors compared dozens of latent brain state algorithms spanning zero-lag analysis (e.g., principal component analysis, principal component analysis with Varimax rotation, Laplacian eigenmaps, spatial independent component analysis, temporal independent component analysis, hidden Markov model, seed-based correlation analysis, and co-activation patterns) to time-lag analysis (e.g., quasi-periodic pattern and lag projections). Bolt et al. (2022) writes “a range of empirical phenomena, including functional connectivity gradients, the task-positive/task-negative anticorrelation pattern, the global signal, time-lag propagation patterns, the quasiperiodic pattern and the functional connectome network structure, are manifestations of the three spatiotemporal patterns.” That is, many previous findings that used different methods essentially describe the same recurring latent states. A similar argument was made in previous papers (Brown et al., 2021; Karapanagiotidis et al., 2020; Turnbull et al., 2020).

We agree that the HMM is a simplified idea of brain dynamics. We do not argue that the four number of states can fully explain the complexity and flexibility of cognition. Instead, we hoped to show that there are different dimensionalities to which the brain systems can operate, and they may have different consequences to cognition. We “simplified” neural dynamics to a discrete sequence of a small number of states. However, what is fascinating is that these overly “simplified” brain state dynamics can explain certain cognitive and attentional dynamics, such as event segmentation and sustained attention fluctuations. We highlight this point in the Discussion.

[Manuscript, page 16] “Our study adopted the assumption of low dimensionality of large-scale neural systems, which led us to intentionally identify only a small number of states underlying whole-brain dynamics. Importantly, however, we do not claim that the four states will be the optimal set of states in every dataset and participant population. Instead, latent states and patterns of state occurrence may vary as a function of individuals and tasks (Figure 1—figure supplement 2). Likewise, while the lowest dimensions of the manifold (i.e., the first two gradients) were largely shared across datasets tested here, we do not argue that it will always be identical. If individuals and tasks deviate significantly from what was tested here, the manifold may also differ along with changes in latent states (Samara et al., 2023). Brain systems operate at different dimensionalities and spatiotemporal scales (Greene et al., 2023), which may have different consequences for cognition. Asking how brain states and manifolds—probed at different dimensionalities and scales—flexibly reconfigure (or not) with changes in contexts and mental states is an important research question for understanding complex human cognition.”

• For the 25- ROI replication it seems like they again do not try multiple K values for the number of states to validate that 4 states are in fact the correct number.

In the manuscript, we do not argue that the four will be the optimal number of states in any dataset. (We actually predict that this may differ depending on the amount of data, participant population, tasks, etc.) Instead, we claim that the four identified in the SONG dataset are not specific (i.e., overfit) to that sample, but rather recur in independent datasets as well. More broadly we argue that the complexity and flexibility of human cognition stem from the fact that computation occurs at multiple dimensions and that the low-dimensional states observed here are robustly related to cognitive and attentional states. To prevent misunderstanding of our results, we emphasized in the Discussion that we are not arguing for a fixed number of states. A paragraph included in our response to the previous comment (page 16 in the manuscript) illustrates this point.

• Fig 2B: Colorbar goes from -0.05 to 0.05 but values are up to 0.87

We apologize for the confusion. The current version of the figure is correct. The figure legend states, “The values indicate transition probabilities, such that values in each row sums to 1. The colors indicate differences from the mean of the null distribution where the HMMs were conducted on the circular-shifted time series.”

We recognize that this complicates the interpretation of the figure. However, after much consideration, we decided that it was valuable to show both the actual transition probabilities (values) and their difference from the mean of null HMMs (colors). The values demonstrate the Markovian property of latent state dynamics, with a high probability of remaining in the same state at consecutive moments and a low probability of transitioning to a different state. The colors indicate that the base state is a transitional hub state by illustrating that the DMN, DAN, and SM states are more likely to transition to the base state than would be expected by chance.

• P 16 L4 near-critical, authors need to be more specific in their terminology here especially since they talk about dynamic systems, where near-criticality has a specific definition. It is unclear which definition they are looking for here.

We agree that our explanation was vague. Because we do not have evidence for this speculative proposal, we removed the mention of near-criticality. Instead, we focus on our observation as the base state being the transitional hub state within a metastable system.

[Manuscript, page 17-18] “However, the functional relevance of the base state to human cognition had not been explored previously. We propose that the base state, a transitional hub (Figure 2B) positioned at the center of the gradient subspace (Figure 1D), functions as a state of natural equilibrium. Transitioning to the DMN, DAN, or SM states reflects incursion away from natural equilibrium (Deco et al., 2017; Gu et al., 2015), as the brain enters a functionally modular state. Notably, the base state indicated high attentional engagement (Figure 5E and F) and exhibited the highest occurrence proportion (Figure 3B) as well as the longest dwell times (Figure 3—figure supplement 1) during naturalistic movie watching, whereas its functional involvement was comparatively minor during controlled tasks. This significant relevance to behavior verifies that the base state cannot simply be a byproduct of the model. We speculate that susceptibility to both external and internal information is maximized in the base state—allowing for roughly equal weighting of both sides so that they can be integrated to form a coherent representation of the world—at the expense of the stability of a certain functional network (Cocchi et al., 2017; Fagerholm et al., 2015). When processing rich narratives, particularly when a person is fully immersed without having to exert cognitive effort, a less modular state with high degrees of freedom to reach other states may be more likely to be involved. The role of the base state should be further investigated in future studies.”

• P16 L13-L17 unnecessary

We prefer to have the last paragraph as a summary of the implications of this paper. However, if the length of this paper becomes a problem as we work towards publication with the editors, we are happy to remove these lines.

• I think this paper is solid, but my main issue is with using an HMM, never explaining why, not showing inference results on test data, not reporting an R^2 score for it, and not comparing it to other models. Secondly, they use the Calinski-Harabasz score to determine the number of states, but not the log-likelihood of the fit. This clearly creates a bias in what types of states you will find, namely states that are far away from each other, which likely also leads to the functional gradient and PCA results they have. Where they specifically talk about how their states are far away from each other in the functional gradient space and correlated to (orthogonal) components. It is completely unclear to me why they used this measure because it also seems to be one of many scores you could use with respect to clustering (with potentially different results), and even odd in the presence of a loglikelihood fit to the data and with the model they use (which does not perform clustering).

(1) Showing inference results on test data

We address this concern on page 19-21 of this response letter.

(2) Not reporting 𝑹𝟐 score

We address this concern on page 21-23 of this response letter.

(3) Not comparing the HMM model to other models

We address this concern on page 27-28 of this response letter.

(4) The use of the Calinski-Harabasz score to determine the number of states rather than the log-likelihood of the model fit

To our knowledge, the log-likelihood of the model fit is not used in the HMM literature. It is because the log-likelihood tends to increase monotonically as the number of states increases. Baker et al. (2014) illustrates this problem, writing:

“In theory, it should be possible to pick the optimal number of states by selecting the model with the greatest (negative) free energy. In practice however, we observe that the free energy increases monotonically up to K = 15 states, suggesting that the Bayes-optimal model may require an even higher number of states.”

Similarly, the following figure is the log-likelihood estimated from the SONG dataset. Similar to the findings of Baker et al. (2014), the log-likelihood monotonically increased as the number of states increased (Author response image 8, right). The measures like AIC or BIC, which account for the number of parameters, also have the same issue of monotonic increase.

Author response image 8.

Because there is “no straightforward data-driven approach to model order selection” (Baker et al., 2014), past work has used different approaches to decide on the number of states. For example, Vidaurre et al. (2018) iterated over a range of the number of states to repeat the same HMM training and inference procedures 5 times using the same hyperparameters. They selected the number of states that showed the highest consistency across iterations. Gao et al. (2021) tested the clustering performance of the model output using the Calinski-Harabasz score. The number of states that showed the highest within-cluster cohesion compared to the across-cluster separation was selected as the number of states. Chang et al. (2021) applied HMM to voxels of the ventromedial prefrontal cortex using a similar clustering algorithm, writing: “To determine the number of states for the HMM estimation procedure, we identified the number of states that maximized the average within-state spatial similarity relative to the average between-state similarity”. In our previous paper (Song et al., 2021b), we reported both the reliability and clustering performance measures to decide on the number of states.

In the current manuscript, the model consistency criterion from Vidaurre et al. (2018) was ineffective because the HMM inference was extremely robust (i.e., always inferring the exact same sequence) due to a large number of data points. Thus, we used the Calinski-Harabasz score as our criterion for the number of states selected.

We agree with the reviewer that the selection of the number of states is critical to any study that implements HMM. However, the field lacks a consensus on how to decide on the number of states in the HMM, and the Calinski-Harabasz score has been validated in previous studies. Most importantly, the latent states’ relationships with behavioral and cognitive measures give strong evidence that the latent states are indeed meaningful states. Again, we are not arguing that the optimal set of states in any dataset will be four nor are we arguing that these four states will always be the optimal states. Instead, the manuscript proposes that a small number of latent states explains meaningful variance in cognitive dynamics.

• Grammatical error: P24 L29 rendering seems to have gone wrong

Our intention was correct here. To avoid confusion, we changed “(number of participantsC2 iterations)” to “(#𝐶!iterations, where N=number of participants)” (page 26 in the manuscript).

Questions:

• Comment on subject differences, it seems like they potentially found group dynamics based on stimuli, but interesting to see individual differences in large-scale dynamics, and do they believe the states they find mostly explain global linear dynamics?

We agree with the reviewer that whether low-dimensional latent state dynamics explain individual differences—above and beyond what could be explained by the high-dimensional, temporally static neural signatures of individuals (e.g., Finn et al., 2015)—is an important research question. However, because the SONG dataset was collected in a single lab, with a focus on covering diverse contexts (rest, task, and movie watching) over 2 sessions, we were only able to collect 27 participants. Due to this small sample size, we focused on investigating group-level, shared temporal dynamics and across-condition differences, rather than on investigating individual differences.

Past work has studied individual differences (e.g., behavioral traits like well-being, intelligence, and personality) using the HMM (Vidaurre et al., 2017). In the lab, we are working on a project that investigates latent state dynamics in relation to individual differences in clinical symptoms using the Healthy Brain Network dataset (Ji et al., 2022, presented at SfN; Alexander et al., 2017).

Finally, the reviewer raises an interesting question about whether the latent state sequence that was derived here mostly explains global linear dynamics as opposed to nonlinear dynamics. We have two responses: one methodological and one theoretical. First, methodologically, we defined the emission probabilities as a linear mixture of Gaussian distributions for each input dimension with the state-specific mean (mean fMRI activity patterns of the networks) and variance (functional covariance across networks). Therefore, states are modeled with an assumption of linearity of feature combinations. Theoretically, recent work supports in favor of nonlinearity of large-scale neural dynamics, especially as tasks get richer and more complex (Cunningham and Yu, 2014; Gao et al., 2021). However, whether low-dimensional latent states should be modeled nonlinearly—that is, whether linear algorithms are insufficient at capturing latent states compared to nonlinear algorithms—is still unknown. We agree with the reviewer that the assumption of linearity is an interesting topic in systems neuroscience. However, together with prior work which showed how numerous algorithms—either linear or nonlinear—recapitulated a common set of latent states, we argue that the HMM provides a strong low-dimensional model of large-scale neural activity and interaction.

• P19 L40 why did the authors interpolate incorrect or no-responses for the gradCPT runs? It seems more logical to correct their results for these responses or to throw them out since interpolation can induce huge biases in these cases because the data is likely not missing at completely random.

Interpolating the RTs of the trials without responses (omission errors and incorrect trials) is a standardized protocol for analyzing gradCPT data (Esterman et al., 2013; Fortenbaugh et al., 2018, 2015; Jayakumar et al., 2023; Rosenberg et al., 2013; Terashima et al., 2021; Yamashita et al., 2021). The choice of this analysis is due to an assumption that sustained attention is a continuous attentional state; the RT, a proxy for the attentional state in the gradCPT literature, is a noisy measure of a smoothed, continuous attentional state. Thus, the RTs of the trials without responses are interpolated and the RT time courses are smoothed by convolving with a gaussian kernel.

References

Abbas A, Belloy M, Kashyap A, Billings J, Nezafati M, Schumacher EH, Keilholz S. 2019. Quasiperiodic patterns contribute to functional connectivity in the brain. Neuroimage 191:193–204.

Alexander LM, Escalera J, Ai L, Andreotti C, Febre K, Mangone A, Vega-Potler N, Langer N, Alexander A, Kovacs M, Litke S, O’Hagan B, Andersen J, Bronstein B, Bui A, Bushey M, Butler H, Castagna V, Camacho N, Chan E, Citera D, Clucas J, Cohen S, Dufek S, Eaves M, Fradera B, Gardner J, Grant-Villegas N, Green G, Gregory C, Hart E, Harris S, Horton M, Kahn D, Kabotyanski K, Karmel B, Kelly SP, Kleinman K, Koo B, Kramer E, Lennon E, Lord C, Mantello G, Margolis A, Merikangas KR, Milham J, Minniti G, Neuhaus R, Levine A, Osman Y, Parra LC, Pugh KR, Racanello A, Restrepo A, Saltzman T, Septimus B, Tobe R, Waltz R, Williams A, Yeo A, Castellanos FX, Klein A, Paus T, Leventhal BL, Craddock RC, Koplewicz HS, Milham MP. 2017. Data Descriptor: An open resource for transdiagnostic research in pediatric mental health and learning disorders. Sci Data 4:1–26.

Allen EA, Damaraju E, Plis SM, Erhardt EB, Eichele T, Calhoun VD. 2014. Tracking whole-brain connectivity dynamics in the resting state. Cereb Cortex 24:663–676.

Baker AP, Brookes MJ, Rezek IA, Smith SM, Behrens T, Probert Smith PJ, Woolrich M. 2014. Fast transient networks in spontaneous human brain activity. Elife 3:e01867.

Bolt T, Nomi JS, Bzdok D, Salas JA, Chang C, Yeo BTT, Uddin LQ, Keilholz SD. 2022. A Parsimonious Description of Global Functional Brain Organization in Three Spatiotemporal Patterns. Nat Neurosci 25:1093–1103.

Brown JA, Lee AJ, Pasquini L, Seeley WW. 2021. A dynamic gradient architecture generates brain activity states. Neuroimage 261:119526.

Chang C, Leopold DA, Schölvinck ML, Mandelkow H, Picchioni D, Liu X, Ye FQ, Turchi JN, Duyn JH. 2016. Tracking brain arousal fluctuations with fMRI. Proc Natl Acad Sci U S A 113:4518–4523.

Chang CHC, Lazaridi C, Yeshurun Y, Norman KA, Hasson U. 2021. Relating the past with the present: Information integration and segregation during ongoing narrative processing. J Cogn Neurosci 33:1–23.

Chang LJ, Jolly E, Cheong JH, Rapuano K, Greenstein N, Chen P-HA, Manning JR. 2021. Endogenous variation in ventromedial prefrontal cortex state dynamics during naturalistic viewing reflects affective experience. Sci Adv 7:eabf7129.

Chen J, Leong YC, Honey CJ, Yong CH, Norman KA, Hasson U. 2017. Shared memories reveal shared structure in neural activity across individuals. Nat Neurosci 20:115–125.

Chen S, Langley J, Chen X, Hu X. 2016. Spatiotemporal Modeling of Brain Dynamics Using RestingState Functional Magnetic Resonance Imaging with Gaussian Hidden Markov Model. Brain Connect 6:326–334.

Cocchi L, Gollo LL, Zalesky A, Breakspear M. 2017. Criticality in the brain: A synthesis of neurobiology, models and cognition. Prog Neurobiol 158:132–152.

Cornblath EJ, Ashourvan A, Kim JZ, Betzel RF, Ciric R, Adebimpe A, Baum GL, He X, Ruparel K, Moore TM, Gur RC, Gur RE, Shinohara RT, Roalf DR, Satterthwaite TD, Bassett DS. 2020. Temporal sequences of brain activity at rest are constrained by white matter structure and modulated by cognitive demands. Commun Biol 3:261.

Cunningham JP, Yu BM. 2014. Dimensionality reduction for large-scale neural recordings. Nat Neurosci 17:1500–1509.

Deco G, Kringelbach ML, Jirsa VK, Ritter P. 2017. The dynamics of resting fluctuations in the brain: Metastability and its dynamical cortical core. Sci Rep 7:3095.

Esterman M, Noonan SK, Rosenberg M, Degutis J. 2013. In the zone or zoning out? Tracking behavioral and neural fluctuations during sustained attention. Cereb Cortex 23:2712–2723.

Esterman M, Rothlein D. 2019. Models of sustained attention. Curr Opin Psychol 29:174–180.

Fagerholm ED, Lorenz R, Scott G, Dinov M, Hellyer PJ, Mirzaei N, Leeson C, Carmichael DW, Sharp DJ, Shew WL, Leech R. 2015. Cascades and cognitive state: Focused attention incurs subcritical dynamics. J Neurosci 35:4626–4634.

Falahpour M, Chang C, Wong CW, Liu TT. 2018. Template-based prediction of vigilance fluctuations in resting-state fMRI. Neuroimage 174:317–327.

Finn ES, Shen X, Scheinost D, Rosenberg MD, Huang J, Chun MM, Papademetris X, Constable RT. 2015. Functional connectome fingerprinting: Identifying individuals using patterns of brain connectivity. Nat Neurosci 18:1664–1671.

Fortenbaugh FC, Degutis J, Germine L, Wilmer JB, Grosso M, Russo K, Esterman M. 2015. Sustained attention across the life span in a sample of 10,000: Dissociating ability and strategy. Psychol Sci 26:1497–1510.

Fortenbaugh FC, Rothlein D, McGlinchey R, DeGutis J, Esterman M. 2018. Tracking behavioral and neural fluctuations during sustained attention: A robust replication and extension. Neuroimage 171:148–164.

Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME. 2005. The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proc Natl Acad Sci U S A 102:9673–9678.

Gao S, Mishne G, Scheinost D. 2021. Nonlinear manifold learning in functional magnetic resonance imaging uncovers a low-dimensional space of brain dynamics. Hum Brain Mapp 42:4510–4524.

Goodale SE, Ahmed N, Zhao C, de Zwart JA, Özbay PS, Picchioni D, Duyn J, Englot DJ, Morgan VL, Chang C. 2021. Fmri-based detection of alertness predicts behavioral response variability. Elife 10:1–20.

Greene AS, Horien C, Barson D, Scheinost D, Constable RT. 2023. Why is everyone talking about brain state? Trends Neurosci.

Greene DJ, Marek S, Gordon EM, Siegel JS, Gratton C, Laumann TO, Gilmore AW, Berg JJ, Nguyen AL, Dierker D, Van AN, Ortega M, Newbold DJ, Hampton JM, Nielsen AN, McDermott KB, Roland JL, Norris SA, Nelson SM, Snyder AZ, Schlaggar BL, Petersen SE, Dosenbach NUF. 2020. Integrative and Network-Specific Connectivity of the Basal Ganglia and Thalamus Defined in Individuals. Neuron 105:742-758.e6.

Gu S, Pasqualetti F, Cieslak M, Telesford QK, Yu AB, Kahn AE, Medaglia JD, Vettel JM, Miller MB, Grafton ST, Bassett DS. 2015. Controllability of structural brain networks. Nat Commun 6:8414.

Jayakumar M, Balusu C, Aly M. 2023. Attentional fluctuations and the temporal organization of memory. Cognition 235:105408.

Ji E, Lee JE, Hong SJ, Shim W (2022). Idiosyncrasy of latent neural state dynamic in ASD during movie watching. Poster presented at the Society for Neuroscience 2022 Annual Meeting.

Karapanagiotidis T, Vidaurre D, Quinn AJ, Vatansever D, Poerio GL, Turnbull A, Ho NSP, Leech R, Bernhardt BC, Jefferies E, Margulies DS, Nichols TE, Woolrich MW, Smallwood J. 2020. The psychological correlates of distinct neural states occurring during wakeful rest. Sci Rep 10:1–11.

Liu X, Duyn JH. 2013. Time-varying functional network information extracted from brief instances of spontaneous brain activity. Proc Natl Acad Sci U S A 110:4392–4397.

Liu X, Zhang N, Chang C, Duyn JH. 2018. Co-activation patterns in resting-state fMRI signals. Neuroimage 180:485–494.

Lynn CW, Cornblath EJ, Papadopoulos L, Bertolero MA, Bassett DS. 2021. Broken detailed balance and entropy production in the human brain. Proc Natl Acad Sci 118:e2109889118.

Margulies DS, Ghosh SS, Goulas A, Falkiewicz M, Huntenburg JM, Langs G, Bezgin G, Eickhoff SB, Castellanos FX, Petrides M, Jefferies E, Smallwood J. 2016. Situating the default-mode network along a principal gradient of macroscale cortical organization. Proc Natl Acad Sci U S A 113:12574–12579.

Mesulam MM. 1998. From sensation to cognition. Brain 121:1013–1052.

Munn BR, Müller EJ, Wainstein G, Shine JM. 2021. The ascending arousal system shapes neural dynamics to mediate awareness of cognitive states. Nat Commun 12:1–9.

Raut R V., Snyder AZ, Mitra A, Yellin D, Fujii N, Malach R, Raichle ME. 2021. Global waves synchronize the brain’s functional systems with fluctuating arousal. Sci Adv 7.

Rosenberg M, Noonan S, DeGutis J, Esterman M. 2013. Sustaining visual attention in the face of distraction: A novel gradual-onset continuous performance task. Attention, Perception, Psychophys 75:426–439.

Rosenberg MD, Finn ES, Scheinost D, Papademetris X, Shen X, Constable RT, Chun MM. 2016. A neuromarker of sustained attention from whole-brain functional connectivity. Nat Neurosci 19:165–171.

Rosenberg MD, Scheinost D, Greene AS, Avery EW, Kwon YH, Finn ES, Ramani R, Qiu M, Todd Constable R, Chun MM. 2020. Functional connectivity predicts changes in attention observed across minutes, days, and months. Proc Natl Acad Sci U S A 117:3797–3807.

Saggar M, Shine JM, Liégeois R, Dosenbach NUF, Fair D. 2022. Precision dynamical mapping using topological data analysis reveals a hub-like transition state at rest. Nat Commun 13.

Schaefer A, Kong R, Gordon EM, Laumann TO, Zuo X-N, Holmes AJ, Eickhoff SB, Yeo BTT. 2018. Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI. Cereb Cortex 28:3095–3114.

Shine JM. 2019. Neuromodulatory Influences on Integration and Segregation in the Brain. Trends Cogn Sci 23:572–583.

Shine JM, Bissett PG, Bell PT, Koyejo O, Balsters JH, Gorgolewski KJ, Moodie CA, Poldrack RA. 2016. The Dynamics of Functional Brain Networks: Integrated Network States during Cognitive Task Performance. Neuron 92:544–554.

Shine JM, Breakspear M, Bell PT, Ehgoetz Martens K, Shine R, Koyejo O, Sporns O, Poldrack RA. 2019. Human cognition involves the dynamic integration of neural activity and neuromodulatory systems. Nat Neurosci 22:289–296.

Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, Filippini N, Watkins KE, Toro R, Laird AR, Beckmann CF. 2009. Correspondence of the brain’s functional architecture during activation and rest. Proc Natl Acad Sci 106:13040–13045.

Song H, Emily FS, Rosenberg MD. 2021a. Neural signatures of attentional engagement during narratives and its consequences for event memory. Proc Natl Acad Sci 118:e2021905118.

Song H, Park B-Y, Park H, Shim WM. 2021b. Cognitive and Neural State Dynamics of Narrative Comprehension. J Neurosci 41:8972–8990.

Taghia J, Cai W, Ryali S, Kochalka J, Nicholas J, Chen T, Menon V. 2018. Uncovering hidden brain state dynamics that regulate performance and decision-making during cognition. Nat Commun 9:2505.

Terashima H, Kihara K, Kawahara JI, Kondo HM. 2021. Common principles underlie the fluctuation of auditory and visual sustained attention. Q J Exp Psychol 74:705–715.

Tian Y, Margulies DS, Breakspear M, Zalesky A. 2020. Topographic organization of the human subcortex unveiled with functional connectivity gradients. Nat Neurosci 23:1421–1432.

Turnbull A, Karapanagiotidis T, Wang HT, Bernhardt BC, Leech R, Margulies D, Schooler J, Jefferies E, Smallwood J. 2020. Reductions in task positive neural systems occur with the passage of time and are associated with changes in ongoing thought. Sci Rep 10:1–10.

Unsworth N, Robison MK. 2018. Tracking arousal state and mind wandering with pupillometry. Cogn Affect Behav Neurosci 18:638–664.

Unsworth N, Robison MK. 2016. Pupillary correlates of lapses of sustained attention. Cogn Affect Behav Neurosci 16:601–615.

van der Meer JN, Breakspear M, Chang LJ, Sonkusare S, Cocchi L. 2020. Movie viewing elicits rich and reliable brain state dynamics. Nat Commun 11:1–14.

Van Essen DC, Smith SM, Barch DM, Behrens TEJ, Yacoub E, Ugurbil K. 2013. The WU-Minn Human Connectome Project: An overview. Neuroimage 80:62–79.

Vidaurre D, Abeysuriya R, Becker R, Quinn AJ, Alfaro-Almagro F, Smith SM, Woolrich MW. 2018. Discovering dynamic brain networks from big data in rest and task. Neuroimage, Brain Connectivity Dynamics 180:646–656.

Vidaurre D, Smith SM, Woolrich MW. 2017. Brain network dynamics are hierarchically organized in time. Proc Natl Acad Sci U S A 114:12827–12832.

Yamashita A, Rothlein D, Kucyi A, Valera EM, Esterman M. 2021. Brain state-based detection of attentional fluctuations and their modulation. Neuroimage 236:118072.

Yeo BTT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, Roffman JL, Smoller JW, Zöllei L, Polimeni JR, Fisch B, Liu H, Buckner RL. 2011. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 106:1125–1165.

Yousefi B, Keilholz S. 2021. Propagating patterns of intrinsic activity along macroscale gradients coordinate functional connections across the whole brain. Neuroimage 231:117827.

Zhang S, Goodale SE, Gold BP, Morgan VL, Englot DJ, Chang C. 2023. Vigilance associates with the low-dimensional structure of fMRI data. Neuroimage 267.

Author Response:

Reviewer #1:

Insulin-secreting beta-cells are electrically excitable, and action potential firing in these cells leads to an increase in the cytoplasmic calcium concentration that in turn stimulates insulin release. Beta-cells are electrically coupled to their neighbours and electrical activity and calcium waves are synchronised across the pancreatic islets. How these oscillations are initiated are not known. In this study, the authors identify a subset of 'first responders' beta-cells that are the first to respond to glucose and that initiate a propagating Ca2+ wave across the islet. These cells may be particularly responsive because of their intrinsic electrophysiological properties. Somewhat unexpectedly, the electrical coupling of first responder cells appears weaker than that in the other islet cells but this paradox is well explained by the authors. Finally, the authors provide evidence of a hierarchy of beta-cells within the islets and that if the first responder cells are destroyed, other islet cells are ready to take over.

The strengths of the paper are the advanced calcium imaging, the photoablation experiments and the longitudinal measurements (up to 48h).

Whilst I find the evidence for the existence of first responders and hierarchy convincing, the link between the first responders in isolated individual islets and first phase insulin secretion seen in vivo (which becomes impaired in type-2 diabetes) seems somewhat overstated. It is is difficult to see how first responders in an islet can synchronise secretion from 1000s (rodents) to millions of islets (man) and it might be wise to down-tone this particular aspect.

We thank the reviewer for highlighting this point. We acknowledge that we did not measure insulin from individual islets post first responder cell ablation, where we observed diminished first phase Ca2+. We do note that studies have linked the first phase Ca2+ response to first phase insulin release [Henquin et al, Diabetes (2006) and Head et al, Diabetes (2012)], albeit with additional amplification signals for higher glucose elevations. Thus a diminished first phase Ca2+ would imply a diminished first phase insulin (although given the amplifying signals the converse would not necessarily be the case).

Nevertheless there are also important caveats to our experiment. Within islets we ablated a single first responder cell. In small islets this ablation diminished Ca2+ in the plane that we imaged. In larger islets this ablation did not, pointing to the presence of multiple first responder cells. Furthermore we only observed the plane of the islet containing the ablated first responder. It is possible elsewhere in the islet that [Ca2+] was not significantly disrupted. Thus even within a small islet it is possible for redundancy, where multiple first responder cells are present and that together drive first phase [Ca2+] across the islet. Loss of a single first responder cell only disrupts Ca2+ locally. That we see a relationship between the timing of the [Ca2+] response and distance from the first responder would support this notion. Results from the islet model also support this notion, where >10% of cells were required to be ablate to significantly disrupt first-phase Ca2+.

While we already discuss the issue of redundancy in large islets and in 3D, we now briefly mention the importance of measuring insulin release.

Reviewer #2:

Kravets et al. further explored the functional heterogeneity in insulin-secreting beta cells in isolated mouse islets. They used slow cytosolic calcium [Ca2+] oscillations with a cycle period of 2 to several minutes in both phases of glucose-dependent beta cell activity that got triggered by a switch from unphysiologically low (2 mM) to unphysiologically high (11 mM) glucose concentration. Based on the presented evidence, they described a distinct population of beta cells responsible for driving the first phase [Ca2+] elevation and characterised it to be different from some other previously described functional subpopulations.

Strengths:

The study uses advanced experimental approaches to address a specific role a subpopulation of beta cells plays during the first phase of an islet response to 11 mM glucose or strong secretagogues like glibenclamide. It finds elements of a broadscale complex network on the events of the slow time scale [Ca2+] oscillations. For this, they appropriately discuss the presence of most connected cells (network hubs) also in slower [Ca2+] oscillations.

Weakness:

The critical weakness of the paper is the evaluation of linear regressions that should support the impact of relative proximity (Fig. 1E), of the response consistency (Fig. 2C), and of increased excitability of the first responder cells (Fig. 3B). None of the datasets provided in the submission satisfies the criterion of normality of the distribution of regression residuals. In addition, the interpretation that the majority of first responder cells retain their early response time could as well be interpreted that the majority does not.

We thank the reviewers for their input, as it really opened multiple opportunities for us to improve our analysis and strengthen our arguments of the existence and consistency of the first responder cells. We present more detailed analysis for these respective figures below and describe how these are included in the manuscript.

As it is described below, we performed additional in-depth analysis and statistical evaluation of the data presented in figures 1E, 2C, and 3B. We now report that two of the datasets (Fig.1 E, Fig.2 C) satisfy the criterion of normality of the distribution of regression residuals. The third dataset (Fig.3 B) does not satisfy this criterion, and we update our interpretation of this data in the text.

Figure 1E Statistics, Scatter: We now show the slope and p-value indicating deviation of the slope from 0, and r^2 values in Fig.1 E. While the scatter is large (r^2=0.1549 in Fig.1E) for cells located at all distances from the first responder cell, we found that scatter substantially diminishes when we consider cells located closer to the first responder (r^2=0.3219 in Fig.S1 F): the response time for cells at distances up to 60 μm from the first responder cells now is shown in Fig.S1 F. The choice of 60 μm comes from it being the maximum first-to-last responder distance in our data set (see red box in Fig.1D).

Additionally, we noticed that within larger islets there may be multiple domains with their own first responder in the center (now in Fig.S1 E) and below. Linear distance/time dependence is preserved withing each domain.

Figure 1E Normality of residuals: We appreciate reviewer’s suggestion and now see that the original “distance vs time” dependence in Fig.1 E did not meet normality of residuals test. When plotted as distance (μm)/response time (percentile), the cumulative distribution still did not meet the Shapiro-Wilk test for normality of residuals (see QQ plot “All distances” below). However, for cells located in the 60 μm proximity of the first responder, the residuals pass the Shapiro- Wilk normality test. The QQ-plots for “up to 60 μm distances” are included in Fig.S1 G.

Figure 2C Statistic and Scatter: After consulting a biostatistician (Dr. Laura Pyle), we realized that since the Response time during initial vs repeated glucose elevation was measured in the same islet, these were repeated measurements on the same statistical units (i.e. a longitudinal study). Therefore, it required a mixed model analysis, as opposed to simple linear regression which we used initially. We now have applied linear mixed effects model (LMEM) to LN- transformed (original data + 0.0001). The 0.0001 value was added to avoid issues of LN(0).

We now show LMEM-derived slope and p-value indicating deviation of the slope from 0 in Fig.2 C. Further, we performed sorting of the data presented in Fig.2 C by distance to each of the first responders (now added to Fig.2D). An example of the sorted vs non-sorted time of response in the large islet with multiple first responders is added to the Source Data – Figure 1. We found a substantial improvement of the scatter in the distance- sorted data, compared to the non-sorted, which indicates that consistency of the glucose response of a cell correlates with it’s proximity to the first responder. We also discuss this in the first sub-section of the Discussion.

Figure 2C Normality of residuals: The residuals pass Shapiro-Wilk normality test for LMEM of the LN-transformed data. We added very small number (0.0001) to all 0 values in our data set, presented in Fig.2C, D, and Fig.S4 A, to perform natural-log transformation. Details on the LMEM and it’s output are added to the Source data – Statistical analysis file.

Figure 3B Statistic and Scatter: We now show LMEM-derived slope and p-value, indicating deviation of the slope from 0, values in Fig.3 B (below). The LMEM-derived slope has p-value of 0.1925, indicating that the slope is not significantly different from 0. This result changes our original interpretation, and we now edit the associated results and discussion.

Figure 3B Normality of residuals: This data set does not pass Shapiro-Wilk test.

A major issue of the work is also that it is unnecessarily complicated. In the Results section, the authors introduce a number of beta cell subpopulations: first responder cell, last responder cell, wave origin cell, wave end cell, hub-like phase 1, hub-like phase 2, and random cells, which are all defined in exclusively relative terms, regarding the time within which the cells responded, phase lags of their oscillations, or mutual distances within the islet. These cell types also partially overlap.

To address this comment, we added Table 1 to describe the properties of these different populations.

Their choice to use the diameter percentile as a metrics for distances between the cells is not well substantiated since they do not demonstrate in what way would the islet size variability influence the conclusion. All presented islets are of rather a comparable size within the diffusion limits.

We replaced normalized distances in Fig.1 D with absolute distance from first responder in μm.

The functional hierarchy of cells defining the first response should be reflected in the consistency of their relative response time. The authors claim that the spatial organisation is consistent over a time of up to 24 hours. In the first place, it is not clear why would this prolonged consistency be of an advantage in comparison to the absence of such consistency. The linear regression analysis between the initial and repeated relative activation times does suggest a significant correlation, but the distribution of regression residuals of the provided data is again not normal and non-conclusive, despite the low p-value. 50% of the cells defined a first responder in the initial stimulation were part of that subpopulation also during the second stimulation, which is rather random.

We began to describe our analysis of the response time to initial and repeated glucose stimulation earlier in this reply. Further evidence of the distance-dependence of the consistency of the response time is now presented in Fig.S4 A: a response time consistency for cells at 60 μm, 50μm, and 40 μm proximity to the first responder. The closer a cell is located to the first responder, the higher is the consistency of its response time (the lower the scatter), below.

If we analyze this data with a linear regression model, where the r^2 allows us to quantitatively demonstrate decrease of the scatter, we observe r^2 of 0.3013, 0.3228, 0.3674 respectively for cells at 60 μm, 50μm, and 40 μm proximity to the first responder (below). This data is not included in the manuscript because residuals do not pass Shapiro-Wilk Normality test for this model (while they do for the LMEM).

One of the most surprising features of this study is the total lack of fast [Ca2+] oscillations, which are in mouse islets, stimulated with 11 mM glucose typically several seconds long and should be easily detected with the measurement speed used.

Our data used in this manuscript contains Ca2+ dynamics from islets with a) slow oscillations only, b) fast oscillations superimposed on the slow oscillations, c) no obvious oscillations (likely continual spiking). Representative curves are below. Because we focused our study on the slow oscillations, we used dynamics of type (a) in our figures, which formed an impression that no fast oscillations were present. In our analysis of dynamics of type (b) we used Fourier transformation to separate slow oscillations from the fast (described in Methods). Dynamics of type (c) were excluded from the analysis of the oscillatory phase, and instead only used for the first-phase analysis. We indicate this exclusion in the methods.

And lastly, we should also not perpetuate imprecise information about the disease if we know better. The first sentence of the Introduction section, stating that "Diabetes is a disease characterised by high blood glucose, …" is not precise. Diabetes only describes polyuria. Regarding the role of high glucose, a quote from a textbook by K. Frayn, R Evans: Human metabolism - a regulatory perspective, 4rd. 2019 „The changes in glucose metabolism are usually regarded as the "hallmark" of diabetes mellitus, and treatment is always monitored by the level of glucose in the blood. However, it has been said that if it were as easy to measure fatty acids in the blood as it is to measure glucose, we would think of diabetes mellitus mainly as a disorder of fat metabolism."

We acknowledge that Diabetes alone refers to polyurea, and instead state Diabetes Mellitus to be more precise to the disease we refer to. We stated “Diabetes is a disease characterized by high blood glucose, ... “ as this is in line with internationally accepted diagnoses and classification criteria, such as position statements from the American Diabetes Association [‘Diagnosis and Classification of Diabetes Mellitus” AMERICAN DIABETES ASSOCIATION, DIABETES CARE, 36, (2013)]. We certainly acknowledge the glucose-centric approach to characterizing and diagnosing Diabetes Mellitus is largely born of the ease of which glucose can be measured. Thus if blood lipids could be easily measured we may be characterizing diabetes as a disease of hyperlipidemia (depending how lipidemia links with complications of diabetes).

Author Response:

Reviewer #1 (Public Review):

In this article, Bollmann and colleagues demonstrated both theoretically and experimentally that blood vessels could be targeted at the mesoscopic scale with time-of-flight magnetic resonance imaging (TOF-MRI). With a mathematical model that includes partial voluming effects explicitly, they outline how small voxels reduce the dependency of blood dwell time, a key parameter of the TOF sequence, on blood velocity. Through several experiments on three human subjects, they show that increasing resolution improves contrast and evaluate additional issues such as vessel displacement artifacts and the separation of veins and arteries.

The overall presentation of the main finding, that small voxels are beneficial for mesoscopic pial vessels, is clear and well discussed, although difficult to grasp fully without a good prior understanding of the underlying TOF-MRI sequence principles. Results are convincing, and some of the data both raw and processed have been provided publicly. Visual inspection and comparisons of different scans are provided, although no quantification or statistical comparison of the results are included.

Potential applications of the study are varied, from modeling more precisely functional MRI signals to assessing the health of small vessels. Overall, this article reopens a window on studying the vasculature of the human brain in great detail, for which studies have been surprisingly limited until recently.

In summary, this article provides a clear demonstration that small pial vessels can indeed be imaged successfully with extremely high voxel resolution. There are however several concerns with the current manuscript, hopefully addressable within the study.

Thank you very much for this encouraging review. While smaller voxel sizes theoretically benefit all blood vessels, we are specifically targeting the (small) pial arteries here, as the inflow-effect in veins is unreliable and susceptibility-based contrasts are much more suited for this part of the vasculature. (We have clarified this in the revised manuscript by substituting ‘vessel’ with ‘artery’ wherever appropriate.) Using a partial-volume model and a relative contrast formulation, we find that the blood delivery time is not the limiting factor when imaging pial arteries, but the voxel size is. Taking into account the comparatively fast blood velocities even in pial arteries with diameters ≤ 200 µm (using t_delivery=l_voxel/v_blood), we find that blood dwell times are sufficiently long for the small voxel sizes considered here to employ the simpler formulation of the flow-related enhancement effect. In other words, small voxels eliminate blood dwell time as a consideration for the blood velocities expected for pial arteries.

We have extended the description of the TOF-MRA sequence in the revised manuscript, and all data and simulations/analyses presented in this manuscript are now publicly available at https://osf.io/nr6gc/ and https://gitlab.com/SaskiaB/pialvesseltof.git, respectively. This includes additional quantifications of the FRE effect for large vessels (adding to the assessment for small vessels already included), and the effect of voxel size on vessel segmentations.

Main points:

1) The manuscript needs clarifying through some additional background information for a readership wider than expert MR physicists. The TOF-MRA sequence and its underlying principles should be introduced first thing, even before discussing vascular anatomy, as it is the key to understanding what aspects of blood physiology and MRI parameters matter here. MR physics shorthand terms should be avoided or defined, as 'spins' or 'relaxation' are not obvious to everybody. The relationship between delivery time and slab thickness should be made clear as well.

Thank you for this valuable comment that the Theory section is perhaps not accessible for all readers. We have adapted the manuscript in several locations to provide more background information and details on time-of-flight contrast. We found, however, that there is no concise way to first present the MR physics part and then introduce the pial arterial vasculature, as the optimization presented therein is targeted towards this structure. To address this comment, we have therefore opted to provide a brief introduction to TOF-MRA first in the Introduction, and then a more in-depth description in the Theory section.

Introduction section:

"Recent studies have shown the potential of time-of-flight (TOF) based magnetic resonance angiography (MRA) at 7 Tesla (T) in subcortical areas (Bouvy et al., 2016, 2014; Ladd, 2007; Mattern et al., 2018; Schulz et al., 2016; von Morze et al., 2007). In brief, TOF-MRA uses the high signal intensity caused by inflowing water protons in the blood to generate contrast, rather than an exogenous contrast agent. By adjusting the imaging parameters of a gradient-recalled echo (GRE) sequence, namely the repetition time (T_R) and flip angle, the signal from static tissue in the background can be suppressed, and high image intensities are only present in blood vessels freshly filled with non-saturated inflowing blood. As the blood flows through the vasculature within the imaging volume, its signal intensity slowly decreases. (For a comprehensive introduction to the principles of MRA, see for example Carr and Carroll (2012)). At ultra-high field, the increased signal-to-noise ratio (SNR), the longer T_1 relaxation times of blood and grey matter, and the potential for higher resolution are key benefits (von Morze et al., 2007)."

Theory section:

"Flow-related enhancement

Before discussing the effects of vessel size, we briefly revisit the fundamental theory of the flow-related enhancement effect used in TOF-MRA. Taking into account the specific properties of pial arteries, we will then extend the classical description to this new regime. In general, TOF-MRA creates high signal intensities in arteries using inflowing blood as an endogenous contrast agent. The object magnetization—created through the interaction between the quantum mechanical spins of water protons and the magnetic field—provides the signal source (or magnetization) accessed via excitation with radiofrequency (RF) waves (called RF pulses) and the reception of ‘echo’ signals emitted by the sample around the same frequency. The T1-contrast in TOF-MRA is based on the difference in the steady-state magnetization of static tissue, which is continuously saturated by RF pulses during the imaging, and the increased or enhanced longitudinal magnetization of inflowing blood water spins, which have experienced no or few RF pulses. In other words, in TOF-MRA we see enhancement for blood that flows into the imaging volume."

"Since the coverage or slab thickness in TOF-MRA is usually kept small to minimize blood delivery time by shortening the path-length of the vessel contained within the slab (Parker et al., 1991), and because we are focused here on the pial vasculature, we have limited our considerations to a maximum blood delivery time of 1000 ms, with values of few hundreds of milliseconds being more likely."

2) The main discussion of higher resolution leading to improvements rather than loss presented here seems a bit one-sided: for a more objective understanding of the differences it would be worth to explicitly derive the 'classical' treatment and show how it leads to different conclusions than the present one. In particular, the link made in the discussion between using relative magnetization and modeling partial voluming seems unclear, as both are unrelated. One could also argue that in theory higher resolution imaging is always better, but of course there are practical considerations in play: SNR, dynamics of the measured effect vs speed of acquisition, motion, etc. These issues are not really integrated into the model, even though they provide strong constraints on what can be done. It would be good to at least discuss the constraints that 140 or 160 microns resolution imposes on what is achievable at present.

Thank you for this excellent suggestion. We found it instructive to illustrate the different effects separately, i.e. relative vs. absolute FRE, and then partial volume vs. no-partial volume effects. In response to comment R2.8 of Reviewer 2, we also clarified the derivation of the relative FRE vs the ‘classical’ absolute FRE (please see R2.8). Accordingly, the manuscript now includes the theoretical derivation in the Theory section and an explicit demonstration of how the classical treatment leads to different conclusions in the Supplementary Material. The important insight gained in our work is that only when considering relative FRE and partial-volume effects together, can we conclude that smaller voxels are advantageous. We have added the following section in the Supplementary Material:

"Effect of FRE Definition and Interaction with Partial-Volume Model

For the definition of the FRE effect employed in this study, we used a measure of relative FRE (Al-Kwifi et al., 2002) in combination with a partial-volume model (Eq. 6). To illustrate the implications of these two effects, as well as their interaction, we have estimated the relative and absolute FRE for an artery with a diameter of 200 µm or 2 000 µm (i.e. no partial-volume effects at the centre of the vessel). The absolute FRE expression explicitly takes the voxel volume into account, and so instead of Eq. (6) for the relative FRE we used"

Eq. (1)

"Note that the division by M_zS^tissue⋅l_voxel^3 to obtain the relative FRE from this expression removes the contribution of the total voxel volume (l_voxel^3). Supplementary Figure 2 shows that, when partial volume effects are present, the highest relative FRE arises in voxels with the same size as or smaller than the vessel diameter (Supplementary Figure 2A), whereas the absolute FRE increases with voxel size (Supplementary Figure 2C). If no partial-volume effects are present, the relative FRE becomes independent of voxel size (Supplementary Figure 2B), whereas the absolute FRE increases with voxel size (Supplementary Figure 2D). While the partial-volume effects for the relative FRE are substantial, they are much more subtle when using the absolute FRE and do not alter the overall characteristics."

Supplementary Figure 2: Effect of voxel size and blood delivery time on the relative flow-related enhancement (FRE) using either a relative (A,B) (Eq. (3)) or an absolute (C,D) (Eq. (12)) FRE definition assuming a pial artery diameter of 200 μm (A,C) or 2 000 µm, i.e. no partial-volume effects at the central voxel of this artery considered here.

In addition, we have also clarified the contribution of the two definitions and their interaction in the Discussion section. Following the suggestion of Reviewer 2, we have extended our interpretation of relative FRE. In brief, absolute FRE is closely related to the physical origin of the contrast, whereas relative FRE is much more concerned with the “segmentability” of a vessel (please see R2.8 for more details):

"Extending classical FRE treatments to the pial vasculature

There are several major modifications in our approach to this topic that might explain why, in contrast to predictions from classical FRE treatments, it is indeed possible to image pial arteries. For instance, the definition of vessel contrast or flow-related enhancement is often stated as an absolute difference between blood and tissue signal (Brown et al., 2014a; Carr and Carroll, 2012; Du et al., 1993, 1996; Haacke et al., 1990; Venkatesan and Haacke, 1997). Here, however, we follow the approach of Al-Kwifi et al. (2002) and consider relative contrast. While this distinction may seem to be semantic, the effect of voxel volume on FRE for these two definitions is exactly opposite: Du et al. (1996) concluded that larger voxel size increases the (absolute) vessel-background contrast, whereas here we predict an increase in relative FRE for small arteries with decreasing voxel size. Therefore, predictions of the depiction of small arteries with decreasing voxel size differ depending on whether one is considering absolute contrast, i.e. difference in longitudinal magnetization, or relative contrast, i.e. contrast differences independent of total voxel size. Importantly, this prediction changes for large arteries where the voxel contains only vessel lumen, in which case the relative FRE remains constant across voxel sizes, but the absolute FRE increases with voxel size (Supplementary Figure 2). Overall, the interpretations of relative and absolute FRE differ, and one measure may be more appropriate for certain applications than the other. Absolute FRE describes the difference in magnetization and is thus tightly linked to the underlying physical mechanism. Relative FRE, however, describes the image contrast and segmentability. If blood and tissue magnetization are equal, both contrast measures would equal zero and indicate that no contrast difference is present. However, when there is signal in the vessel and as the tissue magnetization approaches zero, the absolute FRE approaches the blood magnetization (assuming no partial-volume effects), whereas the relative FRE approaches infinity. While this infinite relative FRE does not directly relate to the underlying physical process of ‘infinite’ signal enhancement through inflowing blood, it instead characterizes the segmentability of the image in that an image with zero intensity in the background and non-zero values in the structures of interest can be segmented perfectly and trivially. Accordingly, numerous empirical observations (Al-Kwifi et al., 2002; Bouvy et al., 2014; Haacke et al., 1990; Ladd, 2007; Mattern et al., 2018; von Morze et al., 2007) and the data provided here (Figure 5, 6 and 7) have shown the benefit of smaller voxel sizes if the aim is to visualize and segment small arteries."

Note that our formulation of the FRE—even without considering SNR—does not suggest that higher resolution is always better, but instead should be matched to the size of the target arteries:

"Importantly, note that our treatment of the FRE does not suggest that an arbitrarily small voxel size is needed, but instead that voxel sizes appropriate for the arterial diameter of interest are beneficial (in line with the classic “matched-filter” rationale (North, 1963)). Voxels smaller than the arterial diameter would not yield substantial benefits (Figure 5) and may result in SNR reductions that would hinder segmentation performance."

Further, we have also extended the concluding paragraph of the Imaging limitation section to also include a practical perspective:

"In summary, numerous theoretical and practical considerations remain for optimal imaging of pial arteries using time-of-flight contrast. Depending on the application, advanced displacement artefact compensation strategies may be required, and zero-filling could provide better vessel depiction. Further, an optimal trade-off between SNR, voxel size and acquisition time needs to be found. Currently, the partial-volume FRE model only considers voxel size, and—as we reduced the voxel size in the experiments—we (partially) compensated the reduction in SNR through longer scan times. This, ultimately, also required the use of prospective motion correction to enable the very long acquisition times necessary for 140 µm isotropic voxel size. Often, anisotropic voxels are used to reduce acquisition time and increase SNR while maintaining in-plane resolution. This may indeed prove advantageous when the (also highly anisotropic) arteries align with the anisotropic acquisition, e.g. when imaging the large supplying arteries oriented mostly in the head-foot direction. In the case of pial arteries, however, there is not preferred orientation because of the convoluted nature of the pial arterial vasculature encapsulating the complex folding of the cortex (see section Anatomical architecture of the pial arterial vasculature). A further reduction in voxel size may be possible in dedicated research settings utilizing even longer acquisition times and/or larger acquisition volumes to maintain SNR. However, if acquisition time is limited, voxel size and SNR need to be carefully balanced against each other."

3) The article seems to imply that TOF-MRA is the only adequate technique to image brain vasculature, while T2 mapping, UHF T1 mapping (see e.g. Choi et al., https://doi.org/10.1016/j.neuroimage.2020.117259) phase (e.g. Fan et al., doi:10.1038/jcbfm.2014.187), QSM (see e.g. Huck et al., https://doi.org/10.1007/s00429-019-01919-4), or a combination (Bernier et al., https://doi.org/10.1002/hbm.24337​, Ward et al., https://doi.org/10.1016/j.neuroimage.2017.10.049) all depict some level of vascular detail. It would be worth quickly reviewing the different effects of blood on MRI contrast and how those have been used in different approaches to measure vasculature. This would in particular help clarify the experiment combining TOF with T2 mapping used to separate arteries from veins (more on this question below).

We apologize if we inadvertently created the impression that TOF-MRA is a suitable technique to image the complete brain vasculature, and we agree that susceptibility-based methods are much more suitable for venous structures. As outlined above, we have revised the manuscript in various sections to indicate that it is the pial arterial vasculature we are targeting. We have added a statement on imaging the venous vasculature in the Discussion section. Please see our response below regarding the use of T2* to separate arteries and veins.

"The advantages of imaging the pial arterial vasculature using TOF-MRA without an exogenous contrast agent lie in its non-invasiveness and the potential to combine these data with various other structural and functional image contrasts provided by MRI. One common application is to acquire a velocity-encoded contrast such as phase-contrast MRA (Arts et al., 2021; Bouvy et al., 2016). Another interesting approach utilises the inherent time-of-flight contrast in magnetization-prepared two rapid acquisition gradient echo (MP2RAGE) images acquired at ultra-high field that simultaneously acquires vasculature and structural data, albeit at lower achievable resolution and lower FRE compared to the TOF-MRA data in our study (Choi et al., 2020). In summary, we expect high-resolution TOF-MRA to be applicable also for group studies to address numerous questions regarding the relationship of arterial topology and morphometry to the anatomical and functional organization of the brain, and the influence of arterial topology and morphometry on brain hemodynamics in humans. In addition, imaging of the pial venous vasculature—using susceptibility-based contrasts such as T2-weighted magnitude (Gulban et al., 2021) or phase imaging (Fan et al., 2015), susceptibility-weighted imaging (SWI) (Eckstein et al., 2021; Reichenbach et al., 1997) or quantitative susceptibility mapping (QSM) (Bernier et al., 2018; Huck et al., 2019; Mattern et al., 2019; Ward et al., 2018)—would enable a comprehensive assessment of the complete cortical vasculature and how both arteries and veins shape brain hemodynamics.*"

4) The results, while very impressive, are mostly qualitative. This seems a missed opportunity to strengthen the points of the paper: given the segmentations already made, the amount/density of detected vessels could be compared across scans for the data of Fig. 5 and 7. The minimum distance between vessels could be measured in Fig. 8 to show a 2D distribution and/or a spatial map of the displacement. The number of vessels labeled as veins instead of arteries in Fig. 9 could be given.

We fully agree that estimating these quantitative measures would be very interesting; however, this would require the development of a comprehensive analysis framework, which would considerably shift the focus of this paper from data acquisition and flow-related enhancement to data analysis. As noted in the discussion section Challenges for vessel segmentation algorithms, ‘The vessel segmentations presented here were performed to illustrate the sensitivity of the image acquisition to small pial arteries’, because the smallest arteries tend to be concealed in the maximum intensity projections. Further, the interpretation of these measures is not straightforward. For example, the number of detected vessels for the artery depicted in Figure 5 does not change across resolutions, but their length does. We have therefore estimated the relative increase in skeleton length across resolutions for Figures 5 and 7. However, these estimates are not only a function of the voxel size but also of the underlying vasculature, i.e. the number of arteries with a certain diameter present, and may thus not generalise well to enable quantitative predictions of the improvement expected from increased resolutions. We have added an illustration of these analyses in the Supplementary Material, and the following additions in the Methods, Results and Discussion sections.

"For vessel segmentation, a semi-automatic segmentation pipeline was implemented in Matlab R2020a (The MathWorks, Natick, MA) using the UniQC toolbox (Frässle et al., 2021): First, a brain mask was created through thresholding which was then manually corrected in ITK-SNAP (http://www.itksnap.org/) (Yushkevich et al., 2006) such that pial vessels were included. For the high-resolution TOF data (Figures 6 and 7, Supplementary Figure 4), denoising to remove high frequency noise was performed using the implementation of an adaptive non-local means denoising algorithm (Manjón et al., 2010) provided in DenoiseImage within the ANTs toolbox, with the search radius for the denoising set to 5 voxels and noise type set to Rician. Next, the brain mask was applied to the bias corrected and denoised data (if applicable). Then, a vessel mask was created based on a manually defined threshold, and clusters with less than 10 or 5 voxels for the high- and low-resolution acquisitions, respectively, were removed from the vessel mask. Finally, an iterative region-growing procedure starting at each voxel of the initial vessel mask was applied that successively included additional voxels into the vessel mask if they were connected to a voxel which was already included and above a manually defined threshold (which was slightly lower than the previous threshold). Both thresholds were applied globally but manually adjusted for each slab. No correction for motion between slabs was applied. The Matlab code describing the segmentation algorithm as well as the analysis of the two-echo TOF acquisition outlined in the following paragraph are also included in our github repository (https://gitlab.com/SaskiaB/pialvesseltof.git). To assess the data quality, maximum intensity projections (MIPs) were created and the outline of the segmentation MIPs were added as an overlay. To estimate the increased detection of vessels with higher resolutions, we computed the relative increase in the length of the segmented vessels for the data presented in Figure 5 (0.8 mm, 0.5 mm, 0.4 mm and 0.3 mm isotropic voxel size) and Figure 7 (0.16 mm and 0.14 mm isotropic voxel size) by computing the skeleton using the bwskel Matlab function and then calculating the skeleton length as the number of voxels in the skeleton multiplied by the voxel size."

"To investigate the effect of voxel size on vessel FRE, we acquired data at four different voxel sizes ranging from 0.8 mm to 0.3 mm isotropic resolution, adjusting only the encoding matrix, with imaging parameters being otherwise identical (FOV, TR, TE, flip angle, R, slab thickness, see section Data acquisition). The total acquisition time increases from less than 2 minutes for the lowest resolution scan to over 6 minutes for the highest resolution scan as a result. Figure 5 shows thin maximum intensity projections of a small vessel. While the vessel is not detectable at the largest voxel size, it slowly emerges as the voxel size decreases and approaches the vessel size. Presumably, this is driven by the considerable increase in FRE as seen in the single slice view (Figure 5, small inserts). Accordingly, the FRE computed from the vessel mask for the smallest part of the vessel (Figure 5, red mask) increases substantially with decreasing voxel size. More precisely, reducing the voxel size from 0.8 mm, 0.5 mm or 0.4 mm to 0.3 mm increases the FRE by 2900 %, 165 % and 85 %, respectively. Assuming a vessel diameter of 300 μm, the partial-volume FRE model (section Introducing a partial-volume model) would predict similar ratios of 611%, 178% and 78%. However, as long as the vessel is larger than the voxel (Figure 5, blue mask), the relative FRE does not change with resolution (see also Effect of FRE Definition and Interaction with Partial-Volume Model in the Supplementary Material). To illustrate the gain in sensitivity to detect smaller arteries, we have estimated the relative increase of the total length of the segmented vasculature (Supplementary Figure 9): reducing the voxel size from 0.8 mm to 0.5 mm isotropic increases the skeleton length by 44 %, reducing the voxel size from 0.5 mm to 0.4 mm isotropic increases the skeleton length by 28 %, and reducing the voxel size from 0.4 mm to 0.3 mm isotropic increases the skeleton length by 31 %. In summary, when imaging small pial arteries, these data support the hypothesis that it is primarily the voxel size, not the blood delivery time, which determines whether vessels can be resolved."

"Indeed, the reduction in voxel volume by 33 % revealed additional small branches connected to larger arteries (see also Supplementary Figure 8). For this example, we found an overall increase in skeleton length of 14 % (see also Supplementary Figure 9)."

"We therefore expect this strategy to enable an efficient image acquisition without the need for additional venous suppression RF pulses. Once these challenges for vessel segmentation algorithms are addressed, a thorough quantification of the arterial vasculature can be performed. For example, the skeletonization procedure used to estimate the increase of the total length of the segmented vasculature (Supplementary Figure 9) exhibits errors particularly in the unwanted sinuses and large veins. While they are consistently present across voxel sizes, and thus may have less impact on relative change in skeleton length, they need to be addressed when estimating the absolute length of the vasculature, or other higher-order features such as number of new branches. (Note that we have also performed the skeletonization procedure on the maximum intensity projections to reduce the number of artefacts and obtained comparable results: reducing the voxel size from 0.8 mm to 0.5 mm isotropic increases the skeleton length by 44 % (3D) vs 37 % (2D), reducing the voxel size from 0.5 mm to 0.4 mm isotropic increases the skeleton length by 28 % (3D) vs 26 % (2D), reducing the voxel size from 0.4 mm to 0.3 mm isotropic increases the skeleton length by 31 % (3D) vs 16 % (2D), and reducing the voxel size from 0.16 mm to 0.14 mm isotropic increases the skeleton length by 14 % (3D) vs 24 % (2D).)"

Supplementary Figure 9: Increase of vessel skeleton length with voxel size reduction. Axial maximum intensity projections for data acquired with different voxel sizes ranging from 0.8 mm to 0.3 mm (TOP) (corresponding to Figure 5) and 0.16 mm to 0.14 mm isotropic (corresponding to Figure 7) are shown. Vessel skeletons derived from segmentations performed for each resolution are overlaid in red. A reduction in voxel size is accompanied by a corresponding increase in vessel skeleton length.

Regarding further quantification of the vessel displacement presented in Figure 8, we have estimated the displacement using the Horn-Schunck optical flow estimator (Horn and Schunck, 1981; Mustafa, 2016) (https://github.com/Mustafa3946/Horn-Schunck-3D-Optical-Flow). However, the results are dominated by the larger arteries, whereas we are mostly interested in the displacement of the smallest arteries, therefore this quantification may not be helpful.

Because the theoretical relationship between vessel displacement and blood velocity is well known (Eq. 7), and we have also outlined the expected blood velocity as a function of arterial diameter in Figure 2, which provided estimates of displacements that matched what was found in our data (as reported in our original submission), we believe that the new quantification in this form does not add value to the manuscript. What would be interesting would be to explore the use of this displacement artefact as a measure of blood velocities. This, however, would require more substantial analyses in particular for estimation of the arterial diameter and additional validation data (e.g. phase-contrast MRA). We have outlined this avenue in the Discussion section. What is relevant to the main aim of this study, namely imaging of small pial arteries, is the insight that blood velocities are indeed sufficiently fast to cause displacement artefacts even in smaller arteries. We have clarified this in the Results section:

"Note that correction techniques exist to remove displaced vessels from the image (Gulban et al., 2021), but they cannot revert the vessels to their original location. Alternatively, this artefact could also potentially be utilised as a rough measure of blood velocity."

"At a delay time of 10 ms between phase encoding and echo time, the observed displacement of approximately 2 mm in some of the larger vessels would correspond to a blood velocity of 200 mm/s, which is well within the expected range (Figure 2). For the smallest arteries, a displacement of one voxel (0.4 mm) can be observed, indicative of blood velocities of 40 mm/s. Note that the vessel displacement can be observed in all vessels visible at this resolution, indicating high blood velocities throughout much of the pial arterial vasculature. Thus, assuming a blood velocity of 40 mm/s (Figure 2) and a delay time of 5 ms for the high-resolution acquisitions (Figure 6), vessel displacements of 0.2 mm are possible, representing a shift of 1–2 voxels."

Regarding the number of vessels labelled as veins, please see our response below to R1.5.

In the main quantification given, the estimation of FRE increase with resolution, it would make more sense to perform the segmentation independently for each scan and estimate the corresponding FRE: using the mask from the highest resolution scan only biases the results. It is unclear also if the background tissue measurement one voxel outside took partial voluming into account (by leaving a one voxel free interface between vessel and background). In this analysis, it would also be interesting to estimate SNR, so you can compare SNR and FRE across resolutions, also helpful for the discussion on SNR.

The FRE serves as an indicator of the potential performance of any segmentation algorithm (including manual segmentation) (also see our discussion on the interpretation of FRE in our response to R1.2). If we were to segment each scan individually, we would, in the ideal case, always obtain the same FRE estimate, as FRE influences the performance of the segmentation algorithm. In practice, this simply means that it is not possible to segment the vessel in the low-resolution image to its full extent that is visible in the high-resolution image, because the FRE is too low for small vessels. However, we agree with the core point that the reviewer is making, and so to help address this, a valuable addition would be to compare the FRE for the section of a vessel that is visible at all resolutions, where we found—within the accuracy of the transformations and resampling across such vastly different resolutions—that the FRE does not increase any further with higher resolution if the vessel is larger than the voxel size (page 18 and Figure 5). As stated in the Methods section, and as noted by the reviewer, we used the voxels immediately next to the vessel mask to define the background tissue signal level. Any resulting potential partial-volume effects in these background voxels would affect all voxel sizes, introducing a consistent bias that would not impact our comparison. However, inspection of the image data in Figure 5 showed partial-volume effects predominantly within those voxels intersecting the vessel, rather than voxels surrounding the vessel, in agreement with our model of FRE.

"All imaging data were slab-wise bias-field corrected using the N4BiasFieldCorrection (Tustison et al., 2010) tool in ANTs (Avants et al., 2009) with the default parameters. To compare the empirical FRE across the four different resolutions (Figure 5), manual masks were first created for the smallest part of the vessel in the image with the highest resolution and for the largest part of the vessel in the image with the lowest resolution. Then, rigid-body transformation parameters from the low-resolution to the high-resolution (and the high-resolution to the low-resolution) images were estimated using coregister in SPM (https://www.fil.ion.ucl.ac.uk/spm/), and their inverse was applied to the vessel mask using SPM’s reslice. To calculate the empirical FRE (Eq. (3)), the mean of the intensity values within the vessel mask was used to approximate the blood magnetization, and the mean of the intensity values one voxel outside of the vessel mask was used as the tissue magnetization."

"To investigate the effect of voxel size on vessel FRE, we acquired data at four different voxel sizes ranging from 0.8 mm to 0.3 mm isotropic resolution, adjusting only the encoding matrix, with imaging parameters being otherwise identical (FOV, TR, TE, flip angle, R, slab thickness, see section Data acquisition). The total acquisition time increases from less than 2 minutes for the lowest resolution scan to over 6 minutes for the highest resolution scan as a result. Figure 5 shows thin maximum intensity projections of a small vessel. While the vessel is not detectable at the largest voxel size, it slowly emerges as the voxel size decreases and approaches the vessel size. Presumably, this is driven by the considerable increase in FRE as seen in the single slice view (Figure 5, small inserts). Accordingly, the FRE computed from the vessel mask for the smallest part of the vessel (Figure 5, red mask) increases substantially with decreasing voxel size. More precisely, reducing the voxel size from 0.8 mm, 0.5 mm or 0.4 mm to 0.3 mm increases the FRE by 2900 %, 165 % and 85 %, respectively. Assuming a vessel diameter of 300 μm, the partial-volume FRE model (section Introducing a partial-volume model) would predict similar ratios of 611%, 178% and 78%. However, if the vessel is larger than the voxel (Figure 5, blue mask), the relative FRE remains constant across resolutions (see also Effect of FRE Definition and Interaction with Partial-Volume Model in the Supplementary Material). To illustrate the gain in sensitivity to smaller arteries, we have estimated the relative increase of the total length of the segmented vasculature (Supplementary Figure 9): reducing the voxel size from 0.8 mm to 0.5 mm isotropic increases the skeleton length by 44 %, reducing the voxel size from 0.5 mm to 0.4 mm isotropic increases the skeleton length by 28 %, and reducing the voxel size from 0.4 mm to 0.3 mm isotropic increases the skeleton length by 31 %. In summary, when imaging small pial arteries, these data support the hypothesis that it is primarily the voxel size, not blood delivery time, which determines whether vessels can be resolved."

Figure 5: Effect of voxel size on flow-related vessel enhancement. Thin axial maximum intensity projections containing a small artery acquired with different voxel sizes ranging from 0.8 mm to 0.3 mm isotropic are shown. The FRE is estimated using the mean intensity value within the vessel masks depicted on the left, and the mean intensity values of the surrounding tissue. The small insert shows a section of the artery as it lies within a single slice. A reduction in voxel size is accompanied by a corresponding increase in FRE (red mask), whereas no further increase is obtained once the voxel size is equal or smaller than the vessel size (blue mask).

After many internal discussions, we had to conclude that deducing a meaningful SNR analysis that would benefit the reader was not possible given the available data due to the complex relationship between voxel size and other imaging parameters in practice. In detail, we have reduced the voxel size but at the same time increased the acquisition time by increasing the number of encoding steps—which we have now also highlighted in the manuscript. We have, however, added additional considerations about balancing SNR and segmentation performance. Note that these considerations are not specific to imaging the pial arteries but apply to all MRA acquisitions, and have thus been discussed previously in the literature. Here, we wanted to focus on the novel insights gained in our study. Importantly, while we previously noted that reducing voxel size improves contrast in vessels whose diameters are smaller than the voxel size, we now explicitly acknowledge that, for vessels whose diameters are larger than the voxel size reducing the voxel size is not helpful---since it only reduces SNR without any gain in contrast---and may hinder segmentation performance, and thus become counterproductive.

"In general, we have not considered SNR, but only FRE, i.e. the (relative) image contrast, assuming that segmentation algorithms would benefit from higher contrast for smaller arteries. Importantly, the acquisition parameters available to maximize FRE are limited, namely repetition time, flip angle and voxel size. SNR, however, can be improved via numerous avenues independent of these parameters (Brown et al., 2014b; Du et al., 1996; Heverhagen et al., 2008; Parker et al., 1991; Triantafyllou et al., 2011; Venkatesan and Haacke, 1997), the simplest being longer acquisition times. If the aim is to optimize a segmentation outcome for a given acquisition time, the trade-off between contrast and SNR for the specific segmentation algorithm needs to be determined (Klepaczko et al., 2016; Lesage et al., 2009; Moccia et al., 2018; Phellan and Forkert, 2017). Our own—albeit limited—experience has shown that segmentation algorithms (including manual segmentation) can accommodate a perhaps surprising amount of noise using prior knowledge and neighborhood information, making these high-resolution acquisitions possible. Importantly, note that our treatment of the FRE does not suggest that an arbitrarily small voxel size is needed, but instead that voxel sizes appropriate for the arterial diameter of interest are beneficial (in line with the classic “matched-filter” rationale (North, 1963)). Voxels smaller than the arterial diameter would not yield substantial benefits (Figure 5) and may result in SNR reductions that would hinder segmentation performance."

5) The separation of arterial and venous components is a bit puzzling, partly because the methodology used is not fully explained, but also partly because the reasons invoked (flow artefact in large pial veins) do not match the results (many small vessels are included as veins). This question of separating both types of vessels is quite important for applications, so the whole procedure should be explained in detail. The use of short T2 seemed also sub-optimal, as both arteries and veins result in shorter T2 compared to most brain tissues: wouldn't a susceptibility-based measure (SWI or better QSM) provide a better separation? Finally, since the T2* map and the regular TOF map are at different resolutions, masking out the vessels labeled as veins will likely result in the smaller veins being left out.

We agree that while the technical details of this approach were provided in the Data analysis section, the rationale behind it was only briefly mentioned. We have therefore included an additional section Inflow-artefacts in sinuses and pial veins in the Theory section of the manuscript. We have also extended the discussion of the advantages and disadvantages of the different susceptibility-based contrasts, namely T2, SWI and QSM. While in theory both T2 and QSM should allow the reliable differentiation of arterial and venous blood, we found T2* to perform more robustly, as QSM can fail in many places, e.g., due to the strong susceptibility sources within superior sagittal and transversal sinuses and pial veins and their proximity to the brain surface, dedicated processing is required (Stewart et al., 2022). Further, we have also elaborated in the Discussion section why the interpretation of Figure 9 regarding the absence or presence of small veins is challenging. Namely, the intensity-based segmentation used here provides only an incomplete segmentation even of the larger sinuses, because the overall lower intensity found in veins combined with the heterogeneity of the intensities in veins violates the assumptions made by most vascular segmentation approaches of homogenous, high image intensities within vessels, which are satisfied in arteries (page 29f) (see also the illustration below). Accordingly, quantifying the number of vessels labelled as veins (R1.4a) would provide misleading results, as often only small subsets of the same sinus or vein are segmented.

"Inflow-artefacts in sinuses and pial veins

Inflow in large pial veins and the sagittal and transverse sinuses can cause flow-related enhancement in these non-arterial vessels. One common strategy to remove this unwanted signal enhancement is to apply venous suppression pulses during the data acquisition, which saturate bloods spins outside the imaging slab. Disadvantages of this technique are the technical challenges of applying these pulses at ultra-high field due to constraints of the specific absorption rate (SAR) and the necessary increase in acquisition time (Conolly et al., 1988; Heverhagen et al., 2008; Johst et al., 2012; Maderwald et al., 2008; Schmitter et al., 2012; Zhang et al., 2015). In addition, optimal positioning of the saturation slab in the case of pial arteries requires further investigation, and in particular supressing signal from the superior sagittal sinus without interfering in the imaging of the pial arteries vasculature at the top of the cortex might prove challenging. Furthermore, this venous saturation strategy is based on the assumption that arterial blood is traveling head-wards while venous blood is drained foot-wards. For the complex and convoluted trajectory of pial vessels this directionality-based saturation might be oversimplified, particularly when considering the higher-order branches of the pial arteries and veins on the cortical surface. Inspired by techniques to simultaneously acquire a TOF image for angiography and a susceptibility-weighted image for venography (Bae et al., 2010; Deistung et al., 2009; Du et al., 1994; Du and Jin, 2008), we set out to explore the possibility of removing unwanted venous structures from the segmentation of the pial arterial vasculature during data postprocessing. Because arteries filled with oxygenated blood have T2-values similar to tissue, while veins have much shorter T2-values due to the presence of deoxygenated blood (Pauling and Coryell, 1936; Peters et al., 2007; Uludağ et al., 2009; Zhao et al., 2007), we used this criterion to remove vessels with short T2* values from the segmentation (see Data Analysis for details). In addition, we also explored whether unwanted venous structures in the high-resolution TOF images—where a two-echo acquisition is not feasible due to the longer readout—can be removed based on detecting them in a lower-resolution image."

"Removal of pial veins

Inflow in large pial veins and the superior sagittal and transverse sinuses can cause a flow-related enhancement in these non-arterial vessels (Figure 9, left). The higher concentration of deoxygenated haemoglobin in these vessels leads to shorter T2 values (Pauling and Coryell, 1936), which can be estimated using a two-echo TOF acquisition (see also Inflow-artefacts in sinuses and pial veins). These vessels can be identified in the segmentation based on their T2 values (Figure 9, left), and removed from the angiogram (Figure 9, right) (Bae et al., 2010; Deistung et al., 2009; Du et al., 1994; Du and Jin, 2008). In particular, the superior and inferior sagittal and the transversal sinuses and large veins which exhibited an inhomogeneous intensity profile and a steep loss of intensity at the slab boundary were identified as non-arterial (Figure 9, left). Further, we also explored the option of removing unwanted venous vessels from the high-resolution TOF image (Figure 7) using a low-resolution two-echo TOF (not shown). This indeed allowed us to remove the strong signal enhancement in the sagittal sinuses and numerous larger veins, although some small veins, which are characterised by inhomogeneous intensity profiles and can be detected visually by experienced raters, remain."

Figure 9: Removal of non-arterial vessels in time-of-flight imaging. LEFT: Segmentation of arteries (red) and veins (blue) using T_2^ estimates. RIGHT: Time-of-flight angiogram after vein removal.*

Our approach also assumes that the unwanted veins are large enough that they are also resolved in the low-resolution image. If we consider the source of the FRE effect, it might indeed be exclusively large veins that are present in TOF-MRA data, which would suggest that our assumption is valid. Fundamentally, the FRE depends on the inflow of un-saturated spins into the imaging slab. However, small veins drain capillary beds in the local tissue, i.e. the tissue within the slab. (Note that due to the slice oversampling implemented in our acquisition, spins just above or below the slab will also be excited.) Thus, small veins only contain blood water spins that have experienced a large number of RF pulses due to the long transit time through the pial arterial vasculature, the capillaries and the intracortical venules. Hence, their longitudinal magnetization would be similar to that of stationary tissue. To generate an FRE effect in veins, “pass-through” venous blood from outside the imaging slab is required. This is only available in veins that are passing through the imaging slab, which have much larger diameters. These theoretical considerations are corroborated by the findings in Figure 9, where large disconnected vessels with varying intensity profiles were identified as non-arterial. Due to the heterogenous intensity profiles in large veins and the sagittal and transversal sinuses, the intensity-based segmentation applied here may only label a subset of the vessel lumen, creating the impression of many small veins. This is particularly the case for the straight and inferior sagittal sinus in the bottom slab of Figure 9. Nevertheless, future studies potentially combing anatomical prior knowledge, advanced segmentation algorithms and susceptibility measures would be capable of removing these unwanted veins in post-processing to enable an efficient TOF-MRA image acquisition dedicated to optimally detecting small arteries without the need for additional venous suppression RF pulses.

6) A more general question also is why this imaging method is limited to pial vessels: at 140 microns, the larger intra-cortical vessels should be appearing (group 6 in Duvernoy, 1981: diameters between 50 and 240 microns). Are there other reasons these vessels are not detected? Similarly, it seems there is no arterial vasculature detected in the white matter here: it is due to the rather superior location of the imaging slab, or a limitation of the method? Likewise, all three results focus on a rather homogeneous region of cerebral cortex, in terms of vascularisation. It would be interesting for applications to demonstrate the capabilities of the method in more complex regions, e.g. the densely vascularised cerebellum, or more heterogeneous regions like the midbrain. Finally, it is notable that all three subjects appear to have rather different densities of vessels, from sparse (participant II) to dense (participant I), with some inhomogeneities in density (frontal region in participant III) and inconsistencies in detection (sinuses absent in participant II). All these points should be discussed.

While we are aware that the diameter of intracortical arteries has been suggested to be up to 240 µm (Duvernoy et al., 1981), it remains unclear how prevalent intracortical arteries of this size are. For example, note that in a different context in the Duvernoy study (in teh revised manuscript), the following values are mentioned (which we followed in Figure 1):

“Central arteries of the Iobule always have a large diameter of 260 µ to 280 µ, at their origin. Peripheral arteries have an average diameter of 150 µ to 180 µ. At the cortex surface, all arterioles of 50 µ or less, penetrate the cortex or form anastomoses. The diameter of most of these penetrating arteries is approximately 40 µ.”

Further, the examinations by Hirsch et al. (2012) (albeit in the macaque brain), showed one (exemplary) intracortical artery belonging to group 6 (Figure 1B), whose diameter appears to be below 100 µm. Given these discrepancies and the fact that intracortical arteries in group 5 only reach 75 µm, we suspect that intracortical arteries with diameters > 140 µm are a very rare occurrence, which we might not have encountered in this data set.

Similarly, arteries in white matter (Nonaka et al., 2003) and the cerebellum (Duvernoy et al., 1983) are beyond our resolution at the moment. The midbrain is an interesting suggesting, although we believe that the cortical areas chosen here with their gradual reduction in diameter along the vascular tree, provide a better illustration of the effect of voxel size than the rather abrupt reduction in vascular diameter found in the midbrain. We have added the even higher resolution requirements in the discussion section:

"In summary, we expect high-resolution TOF-MRA to be applicable also for group studies, to address numerous questions regarding the relationship of arterial topology and morphometry to the anatomical and functional organization of the brain, and the influence of arterial topology and morphometry on brain hemodynamics in humans. Notably, we have focused on imaging pial arteries of the human cerebrum; however, other brain structures such as the cerebellum, subcortex and white matter are of course also of interest. While the same theoretical considerations apply, imaging the arterial vasculature in these structures will require even smaller voxel sizes due to their smaller arterial diameters (Duvernoy et al., 1983, 1981; Nonaka et al., 2003)."

Regarding the apparent sparsity of results from participant II, this is mostly driven by the much smaller coverage in this subject (19.6 mm in Participant II vs. 50 mm and 58 mm in Participant I and III, respectively). The reduction in density in the frontal regions might indeed constitute difference in anatomy or might be driven by the presence or more false-positive veins in Participant I than Participant III in these areas. Following the depiction in Duvernoy et al. (1981), one would not expect large arteries in frontal areas, but large veins are common. Thus, the additional vessels in Participant I in the frontal areas might well be false-positive veins, and their removal would result in similar densities for both participants. Indeed, as pointed out in section Future directions, we would expect a lower arterial density in frontal and posterior areas than in middle areas. The sinuses (and other large false-positive veins) in Participant II have been removed as outlined and discussed in sections Removal of pial veins and Challenges for vessel segmentation algorithms, respectively.

7) One of the main practical limitations of the proposed method is the use of a very small imaging slab. It is mentioned in the discussion that thicker slabs are not only possible, but beneficial both in terms of SNR and acceleration possibilities. What are the limitations that prevented their use in the present study? With the current approach, what would be the estimated time needed to acquire the vascular map of an entire brain? It would also be good to indicate whether specific processing was needed to stitch together the multiple slab images in Fig. 6-9, S2.

Time-of-flight acquisitions are commonly performed with thin acquisition slabs, following initial investigations by Parker et al. (1991) to maximise vessel sensitivity and minimize noise. We therefore followed this practice for our initial investigations but wanted to point out in the discussion that thicker slabs might provide several advantages that need to be evaluated in future studies. This would include theoretical and empirical evaluations balancing SNR gains from larger excitation volumes and SNR losses due to more acceleration. For this study, we have chosen the slab thickness such as to keep the acquisition time at a reasonable amount to minimize motion artefacts (as outlined in the Discussion). In addition, due to the extreme matrix sizes in particular for the 0.14 mm acquisition, we were also limited in the number of data points per image that can be indexed. This would require even more substantial changes to the sequence than what we have already performed. With 16 slabs, assuming optimal FOV orientation, full-brain coverage including the cerebellum of 95 % of the population (Mennes et al., 2014) could be achieved with an acquisition time of (16  11 min 42 s = 3 h 7 min 12 s) at 0.16 mm isotropic voxel size. No stitching of the individual slabs was performed, as subject motion was minimal. We have added a corresponding comment in the Data Analysis.

"Both thresholds were applied globally but manually adjusted for each slab. No correction for motion between slabs was applied as subject motion was minimal. The Matlab code describing the segmentation algorithm as well es the analysis of the two-echo TOF acquisition outlined in the following paragraph are also included in the github repository (https://gitlab.com/SaskiaB/pialvesseltof.git)."

8) Some researchers and clinicians will argue that you can attain best results with anisotropic voxels, combining higher SNR and higher resolution. It would be good to briefly mention why isotropic voxels are preferred here, and whether anisotropic voxels would make sense at all in this context.

Anisotropic voxels can be advantageous if the underlying object is anisotropic, e.g. an artery running straight through the slab, which would have a certain diameter (imaged using the high-resolution plane) and an ‘infinite’ elongation (in the low-resolution direction). However, the vessels targeted here can have any orientation and curvature; an anisotropic acquisition could therefore introduce a bias favouring vessels with a particular orientation relative to the voxel grid. Note that the same argument applies when answering the question why a further reduction slab thickness would eventually result in less increase in FRE (section Introducing a partial-volume model). We have added a corresponding comment in our discussion on practical imaging considerations:

"In summary, numerous theoretical and practical considerations remain for optimal imaging of pial arteries using time-of-flight contrast. Depending on the application, advanced displacement artefact compensation strategies may be required, and zero-filling could provide better vessel depiction. Further, an optimal trade-off between SNR, voxel size and acquisition time needs to be found. Currently, the partial-volume FRE model only considers voxel size, and—as we reduced the voxel size in the experiments—we (partially) compensated the reduction in SNR through longer scan times. This, ultimately, also required the use of prospective motion correction to enable the very long acquisition times necessary for 140 µm isotropic voxel size. Often, anisotropic voxels are used to reduce acquisition time and increase SNR while maintaining in-plane resolution. This may indeed prove advantageous when the (also highly anisotropic) arteries align with the anisotropic acquisition, e.g. when imaging the large supplying arteries oriented mostly in the head-foot direction. In the case of pial arteries, however, there is not preferred orientation because of the convoluted nature of the pial arterial vasculature encapsulating the complex folding of the cortex (see section Anatomical architecture of the pial arterial vasculature). A further reduction in voxel size may be possible in dedicated research settings utilizing even longer acquisition times and a larger field-of-view to maintain SNR. However, if acquisition time is limited, voxel size and SNR need to be carefully balanced against each other."

Reviewer #2 (Public Review):

Overview

This paper explores the use of inflow contrast MRI for imaging the pial arteries. The paper begins by providing a thorough background description of pial arteries, including past studies investigating the velocity and diameter. Following this, the authors consider this information to optimize the contrast between pial arteries and background tissue. This analysis reveals spatial resolution to be a strong factor influencing the contrast of the pial arteries. Finally, experiments are performed on a 7T MRI to investigate: the effect of spatial resolution by acquiring images at multiple resolutions, demonstrate the feasibility of acquiring ultrahigh resolution 3D TOF, the effect of displacement artifacts, and the prospect of using T2* to remove venous voxels.

Impression

There is certainly interest in tools to improve our understanding of the architecture of the small vessels of the brain and this work does address this. The background description of the pial arteries is very complete and the manuscript is very well prepared. The images are also extremely impressive, likely benefiting from motion correction, 7T, and a very long scan time. The authors also commit to open science and provide the data in an open platform. Given this, I do feel the manuscript to be of value to the community; however, there are concerns with the methods for optimization, the qualitative nature of the experiments, and conclusions drawn from some of the experiments.

Specific Comments :

1) Figure 3 and Theory surrounding. The optimization shown in Figure 3 is based fixing the flip angle or the TR. As is well described in the literature, there is a strong interdependency of flip angle and TR. This is all well described in literature dating back to the early 90s. While I think it reasonable to consider these effects in optimization, the language needs to include this interdependency or simply reference past work and specify how the flip angle was chosen. The human experiments do not include any investigation of flip angle or TR optimization.

We thank the reviewer for raising this valuable point, and we fully agree that there is an interdependency between these two parameters. To simplify our optimization, we did fix one parameter value at a time, but in the revised manuscript we clarified that both parameters can be optimized simultaneously. Importantly, a large range of parameter values will result in a similar FRE in the small artery regime, which is illustrated in the optimization provided in the main text. We have therefore chosen the repetition time based on encoding efficiency and then set a corresponding excitation flip angle. In addition, we have also provided additional simulations in the supplementary material outlining the interdependency for the case of pial arteries.

"Optimization of repetition time and excitation flip angle

As the main goal of the optimisation here was to start within an already established parameter range for TOF imaging at ultra-high field (Kang et al., 2010; Stamm et al., 2013; von Morze et al., 2007), we only needed to then further tailor these for small arteries by considering a third parameter, namely the blood delivery time. From a practical perspective, a TR of 20 ms as a reference point was favourable, as it offered a time-efficient readout minimizing wait times between excitations but allowing low encoding bandwidths to maximize SNR. Due to the interdependency of flip angle and repetition time, for any one blood delivery time any FRE could (in theory) be achieved. For example, a similar FRE curve at 18 ° flip angle and 5 ms TR can also be achieved at 28 ° flip angle and 20 ms TR; or the FRE curve at 18 ° flip angle and 30 ms TR is comparable to the FRE curve at 8 ° flip angle and 5 ms TR (Supplementary Figure 3 TOP). In addition, the difference between optimal parameter settings diminishes for long blood delivery times, such that at a blood delivery time of 500 ms (Supplementary Figure 3 BOTTOM), the optimal flip angle at a TR of 15 ms, 20 ms or 25 ms would be 14 °, 16 ° and 18 °, respectively. This is in contrast to a blood delivery time of 100 ms, where the optimal flip angles would be 32 °, 37 ° and 41 °. In conclusion, in the regime of small arteries, long TR values in combination with low flip angles ensure flow-related enhancement at blood delivery times of 200 ms and above, and within this regime there are marginal gains by further optimizing parameter values and the optimal values are all similar."

Supplementary Figure 3: Optimal imaging parameters for small arteries. This assessment follows the simulations presented in Figure 3, but in addition shows the interdependency for the corresponding third parameter (either flip angle or repetition time). TOP: Flip angles close to the Ernst angle show only a marginal flow-related enhancement; however, the influence of the blood delivery time decreases further (LEFT). As the flip angle increases well above the values used in this study, the flow-related enhancement in the small artery regime remains low even for the longer repetition times considered here (RIGHT). BOTTOM: The optimal excitation flip angle shows reduced variability across repetition times in the small artery regime compared to shorter blood delivery times.

"Based on these equations, optimal T_R and excitation flip angle values (θ) can be calculated for the blood delivery times under consideration (Figure 3). To better illustrate the regime of small arteries, we have illustrated the effect of either flip angle or T_R while keeping the other parameter values fixed to the value that was ultimately used in the experiments; although both parameters can also be optimized simultaneously (Haacke et al., 1990). Supplementary Figure 3 further delineates the interdependency between flip angle and T_R within a parameter range commonly used for TOF imaging at ultra-high field (Kang et al., 2010; Stamm et al., 2013; von Morze et al., 2007). Note how longer T_R values still provide an FRE effect even at very long blood delivery times, whereas using shorter T_R values can suppress the FRE effect (Figure 3, left). Similarly, at lower flip angles the FRE effect is still present for long blood delivery times, but it is not available anymore at larger flip angles, which, however, would give maximum FRE for shorter blood delivery times (Figure 3, right). Due to the non-linear relationships of both blood delivery time and flip angle with FRE, the optimal imaging parameters deviate considerably when comparing blood delivery times of 100 ms and 300 ms, but the differences between 300 ms and 1000 ms are less pronounced. In the following simulations and measurements, we have thus used a T_R value of 20 ms, i.e. a value only slightly longer than the readout of the high-resolution TOF acquisitions, which allowed time-efficient data acquisition, and a nominal excitation flip angle of 18°. From a practical standpoint, these values are also favorable as the low flip angle reduces the specific absorption rate (Fiedler et al., 2018) and the long T_R value decreases the potential for peripheral nerve stimulation (Mansfield and Harvey, 1993)."

2) Figure 4 and Theory surrounding. A major limitation of this analysis is the lack of inclusion of noise in the analysis. I believe the results to be obvious that the FRE will be modulated by partial volume effects, here described quadratically by assuming the vessel to pass through the voxel. This would substantially modify the analysis, with a shift towards higher voxel volumes (scan time being equal). The authors suggest the FRE to be the dominant factor effecting segmentation; however, segmentation is limited by noise as much as contrast.

We of course agree with the reviewer that contrast-to-noise ratio is a key factor that determines the detection of vessels and the quality of the segmentation, however there are subtleties regarding the exact inter-relationship between CNR, resolution, and segmentation performance.

The main purpose of Figure 4 is not to provide a trade-off between flow-related enhancement and signal-to-noise ratio—in particular as SNR is modulated by many more factors than voxel size alone, e.g. acquisition time, coil geometry and instrumentation—but to decide whether the limiting factor for imaging pial arteries is the reduction in flow-related enhancement due to long blood delivery times (which is the explanation often found in the literature (Chen et al., 2018; Haacke et al., 1990; Masaryk et al., 1989; Mut et al., 2014; Park et al., 2020; Parker et al., 1991; Wilms et al., 2001; Wright et al., 2013)) or due to partial volume effects. Furthermore, when reducing voxel size one will also likely increase the number of encoding steps to maintain the imaging coverage (i.e., the field-of-view) and so the relationship between voxel size and SNR in practice is not straightforward. Therefore, we had to conclude that deducing a meaningful SNR analysis that would benefit the reader was not possible given the available data due to the complex relationship between voxel size and other imaging parameters. Note that these considerations are not specific to imaging the pial arteries but apply to all MRA acquisitions, and have thus been discussed previously in the literature. Here, we wanted to focus on the novel insights gained in our study, namely that it provides an expression for how relative FRE contrast changes with voxel size with some assumptions that apply for imaging pial arteries.

Further, depending on the definition of FRE and whether partial-volume effects are included (see also our response to R2.8), larger voxel volumes have been found to be theoretically advantageous even when only considering contrast (Du et al., 1996; Venkatesan and Haacke, 1997), which is not in line with empirical observations (Al-Kwifi et al., 2002; Bouvy et al., 2014; Haacke et al., 1990; Ladd, 2007; Mattern et al., 2018; von Morze et al., 2007).

The notion that vessel segmentation algorithms perform well on noisy data but poorly on low-contrast data was mainly driven by our own experiences. However, we still believe that the assumption that (all) segmentation algorithms are linearly dependent on contrast and noise (which the formulation of a contrast-to-noise ratio presumes) is similarly not warranted. Indeed, the necessary trade-off between FRE and SNR might be specific to the particular segmentation algorithm being used than a general property of the acquisition. Please also note that our analysis of the FRE does not suggest that an arbitrarily high resolution is needed. Importantly, while we previously noted that reducing voxel size improves contrast in vessels whose diameters are smaller than the voxel size, we now explicitly acknowledge that, for vessels whose diameters are larger than the voxel size reducing the voxel size is not helpful---since it only reduces SNR without any gain in contrast---and may hinder segmentation performance, and thus become counterproductive. But we take the reviewer’s point and also acknowledge that these intricacies need to be mentioned, and therefore we have rephrased the statement in the discussion in the following way:

"In general, we have not considered SNR, but only FRE, i.e. the (relative) image contrast, assuming that segmentation algorithms would benefit from higher contrast for smaller arteries. Importantly, the acquisition parameters available to maximize FRE are limited, namely repetition time, flip angle and voxel size. SNR, however, can be improved via numerous avenues independent of these parameters (Brown et al., 2014b; Du et al., 1996; Heverhagen et al., 2008; Parker et al., 1991; Triantafyllou et al., 2011; Venkatesan and Haacke, 1997), the simplest being longer acquisition times. If the aim is to optimize a segmentation outcome for a given acquisition time, the trade-off between contrast and SNR for the specific segmentation algorithm needs to be determined (Klepaczko et al., 2016; Lesage et al., 2009; Moccia et al., 2018; Phellan and Forkert, 2017). Our own—albeit limited—experience has shown that segmentation algorithms (including manual segmentation) can accommodate a perhaps surprising amount of noise using prior knowledge and neighborhood information, making these high-resolution acquisitions possible. Importantly, note that our treatment of the FRE does not suggest that an arbitrarily small voxel size is needed, but instead that voxel sizes appropriate for the arterial diameter of interest are beneficial (in line with the classic “matched-filter” rationale (North, 1963)). Voxels smaller than the arterial diameter would not yield substantial benefits (Figure 5) and may result in SNR reductions that would hinder segmentation performance."

3) Page 11, Line 225. "only a fraction of the blood is replaced" I think the language should be reworded. There are certainly water molecules in blood which have experience more excitation B1 pulses due to the parabolic flow upstream and the temporal variation in flow. There is magnetization diffusion which reduces the discrepancy; however, it seems pertinent to just say the authors assume the signal is represented by the average arrival time. This analysis is never verified and is only approximate anyways. The "blood dwell time" is also an average since voxels near the wall will travel more slowly. Overall, I recommend reducing the conjecture in this section.

We fully agree that our treatment of the blood dwell time does not account for the much more complex flow patterns found in cortical arteries. However, our aim was not do comment on these complex patterns, but to help establish if, in the simplest scenario assuming plug flow, the often-mentioned slow blood flow requires multiple velocity compartments to describe the FRE (as is commonly done for 2D MRA (Brown et al., 2014a; Carr and Carroll, 2012)). We did not intend to comment on the effects of laminar flow or even more complex flow patterns, which would require a more in-depth treatment. However, as the small arteries targeted here are often just one voxel thick, all signals are indeed integrated within that voxel (i.e. there is no voxel near the wall that travels more slowly), which may average out more complex effects. We have clarified the purpose and scope of this section in the following way:

"In classical descriptions of the FRE effect (Brown et al., 2014a; Carr and Carroll, 2012), significant emphasis is placed on the effect of multiple “velocity segments” within a slice in the 2D imaging case. Using the simplified plug-flow model, where the cross-sectional profile of blood velocity within the vessel is constant and effects such as drag along the vessel wall are not considered, these segments can be described as ‘disks’ of blood that do not completely traverse through the full slice within one T_R, and, thus, only a fraction of the blood in the slice is replaced. Consequently, estimation of the FRE effect would then need to accommodate contribution from multiple ‘disks’ that have experienced 1 to k RF pulses. In the case of 3D imaging as employed here, multiple velocity segments within one voxel are generally not considered, as the voxel sizes in 3D are often smaller than the slice thickness in 2D imaging and it is assumed that the blood completely traverses through a voxel each T_R. However, the question arises whether this assumption holds for pial arteries, where blood velocity is considerably lower than in intracranial vessels (Figure 2). To answer this question, we have computed the blood dwell time , i.e. the average time it takes the blood to traverse a voxel, as a function of blood velocity and voxel size (Figure 2). For reference, the blood velocity estimates from the three studies mentioned above (Bouvy et al., 2016; Kobari et al., 1984; Nagaoka and Yoshida, 2006) have been added in this plot as horizontal white lines. For the voxel sizes of interest here, i.e. 50–300 μm, blood dwell times are, for all but the slowest flows, well below commonly used repetition times (Brown et al., 2014a; Carr and Carroll, 2012; Ladd, 2007; von Morze et al., 2007). Thus, in a first approximation using the plug-flow model, it is not necessary to include several velocity segments for the voxel sizes of interest when considering pial arteries, as one might expect from classical treatments, and the FRE effect can be described by equations (1) – (3), simplifying our characterization of FRE for these vessels. When considering the effect of more complex flow patterns, it is important to bear in mind that the arteries targeted here are only one-voxel thick, and signals are integrated across the whole artery."

4) Page 13, Line 260. "two-compartment modelling" I think this section is better labeled "Extension to consider partial volume effects" The compartments are not interacting in any sense in this work.

Thank you for this suggestion. We have replaced the heading with Introducing a partial-volume model (page 14) and replaced all instances of ‘two-compartment model’ with ‘partial-volume model’.

5) Page 14, Line 284. "In practice, a reduction in slab …." "reducing the voxel size is a much more promising avenue" There is a fair amount on conjecture here which is not supported by experiments. While this may be true, the authors also use a classical approach with quite thin slabs.

The slab thickness used in our experiments was mainly limited by the acquisition time and the participants ability to lie still. We indeed performed one measurement with a very experienced participant with a thicker slab, but found that with over 20 minutes acquisition time, motion artefacts were unavoidable. The data presented in Figure 5 were acquired with similar slab thickness, supporting the statement that reducing the voxel size is a promising avenue for imaging small pial arteries. However, we indeed have not provided an empirical comparison of the effect of slab thickness. Nevertheless, we believe it remains useful to make the theoretical argument that due to the convoluted nature of the pial arterial vascular geometry, a reduction in slab thickness may not reduce the acquisition time if no reduction in intra-slab vessel length can be achieved, i.e. if the majority of the artery is still contained in the smaller slab. We have clarified the statement and removed the direct comparison (‘much more’ promising) in the following way:

"In theory, a reduction in blood delivery time increases the FRE in both regimes, and—if the vessel is smaller than the voxel—so would a reduction in voxel size. In practice, a reduction in slab thickness―which is the default strategy in classical TOF-MRA to reduce blood delivery time―might not provide substantial FRE increases for pial arteries. This is due to their convoluted geometry (see section Anatomical architecture of the pial arterial vasculature), where a reduction in slab thickness may not necessarily reduce the vessel segment length if the majority of the artery is still contained within the smaller slab. Thus, given the small arterial diameter, reducing the voxel size is a promising avenue when imaging the pial arterial vasculature."

6) Figure 5. These image differences are highly exaggerated by the lack of zero filling (or any interpolation) and the fact that the wildly different. The interpolation should be addressed, and the scan time discrepancy listed as a limitation.

We have extended the discussion around zero-filling by including additional considerations based on the imaging parameters in Figure 5 and highlighted the substantial differences in voxel volume. Our choice not to perform zero-filling was driven by the open question of what an ‘optimal’ zero-filling factor would be. We have also highlighted the substantial differences in acquisition time when describing the results.

Changes made to the results section:

"To investigate the effect of voxel size on vessel FRE, we acquired data at four different voxel sizes ranging from 0.8 mm to 0.3 mm isotropic resolution, adjusting only the encoding matrix, with imaging parameters being otherwise identical (FOV, TR, TE, flip angle, R, slab thickness, see section Data acquisition). The total acquisition time increases from less than 2 minutes for the lowest resolution scan to over 6 minutes for the highest resolution scan as a result."

Changes made to the discussion section:

"Nevertheless, slight qualitative improvements in image appearance have been reported for higher zero-filling factors (Du et al., 1994), presumably owing to a smoother representation of the vessels (Bartholdi and Ernst, 1973). In contrast, Mattern et al. (2018) reported no improvement in vessel contrast for their high-resolution data. Ultimately, for each application, e.g. visual evaluation vs. automatic segmentation, the optimal zero-filling factor needs to be determined, balancing image appearance (Du et al., 1994; Zhu et al., 2013) with loss in statistical independence of the image noise across voxels. For example, in Figure 5, when comparing across different voxel sizes, the visual impression might improve with zero-filling. However, it remains unclear whether the same zero-filling factor should be applied for each voxel size, which means that the overall difference in resolution remains, namely a nearly 20-fold reduction in voxel volume when moving from 0.8-mm isotropic to 0.3-mm isotropic voxel size. Alternatively, the same ’zero-filled’ voxel sizes could be used for evaluation, although then nearly 94 % of the samples used to reconstruct the image with 0.8-mm voxel size would be zero-valued for a 0.3-mm isotropic resolution. Consequently, all data presented in this study were reconstructed without zero-filling."

7) Figure 7. Given the limited nature of experiment may it not also be possible the subject moved more, had differing brain blood flow, etc. Were these lengthy scans acquired in the same session? Many of these differences could be attributed to other differences than the small difference in spatial resolution.

The scans were acquired in the same session using the same prospective motion correction procedure. Note that the acquisition time of the images with 0.16 mm isotropic voxel size was comparatively short, taking just under 12 minutes. Although the difference in spatial resolution may seem small, it still amounts to a 33% reduction in voxel volume. For comparison, reducing the voxel size from 0.4 mm to 0.3 mm also ‘only’ reduces the voxel volume by 58 %—not even twice as much. Overall, we fully agree that additional validation and optimisation of the imaging parameters for pial arteries are beneficial and have added a corresponding statement to the Discussion section.

Changes made to the results section (also in response to Reviewer 1 (R1.22))

"We have also acquired one single slab with an isotropic voxel size of 0.16 mm with prospective motion correction for this participant in the same session to compare to the acquisition with 0.14 mm isotropic voxel size and to test whether any gains in FRE are still possible at this level of the vascular tree."

Changes made to the discussion section:

"Acquiring these data at even higher field strengths would boost SNR (Edelstein et al., 1986; Pohmann et al., 2016) to partially compensate for SNR losses due to acceleration and may enable faster imaging and/or smaller voxel sizes. This could facilitate the identification of the ultimate limit of the flow-related enhancement effect and identify at which stage of the vascular tree does the blood delivery time become the limiting factor. While Figure 7 indicates the potential for voxel sizes below 0.16 mm, the singular nature of this comparison warrants further investigations."

8) Page 22, Line 395. Would the analysis be any different with an absolute difference? The FRE (Eq 6) divides by a constant value. Clearly there is value in the difference as other subtractive inflow imaging would have infinite FRE (not considering noise as the authors do).

Absolutely; using an absolute FRE would result in the highest FRE for the largest voxel size, whereas in our data small vessels are more easily detected with the smallest voxel size. We also note that relative FRE would indeed become infinite if the value in the denominator representing the tissue signal was zero, but this special case highlights how relative FRE can help characterize “segmentability”: a vessel with any intensity surrounded by tissue with an intensity of zero is trivially/infinitely segmentatble. We have added this point to the revised manuscript as indicated below.

Following the suggestion of Reviewer 1 (R1.2), we have included additional simulations to clarify the effects of relative FRE definition and partial-volume model, in which we show that only when considering both together are smaller voxel sizes advantageous (Supplementary Material).

"Effect of FRE Definition and Interaction with Partial-Volume Model

For the definition of the FRE effect in this study, we used a measure of relative FRE (Al-Kwifi et al., 2002) in combination with a partial-volume model (Eq. 6). To illustrate the effect of these two definitions, as well as their interaction, we have estimated the relative and absolute FRE for an artery with a diameter of 200 µm and 2 000 µm (i.e. no partial-volume effects). The absolute FRE explicitly takes the voxel volume into account, i.e. instead of Eq. (6) for the relative FRE we used"

Eq. (1)

Note that the division by

to obtain the relative FRE removes the contribution of the total voxel volume

"Supplementary Figure 2 shows that, when partial volume effects are present, the highest relative FRE arises in voxels with the same size as or smaller than the vessel diameter (Supplementary Figure 2A), whereas the absolute FRE increases with voxel size (Supplementary Figure 2C). If no partial-volume effects are present, the relative FRE becomes independent of voxel size (Supplementary Figure 2B), whereas the absolute FRE increases with voxel size (Supplementary Figure 2D). While the partial-volume effects for the relative FRE are substantial, they are much more subtle when using the absolute FRE and do not alter the overall characteristics."

Supplementary Figure 2: Effect of voxel size and blood delivery time on the relative flow-related enhancement (FRE) using either a relative (A,B) (Eq. (3)) or an absolute (C,D) (Eq. (12)) FRE definition assuming a pial artery diameter of 200 μm (A,C) or 2 000 µm, i.e. no partial-volume effects at the central voxel of this artery considered here.

Following the established literature (Brown et al., 2014a; Carr and Carroll, 2012; Haacke et al., 1990) and because we would ultimately derive a relative measure, we have omitted the effect of voxel volume on the longitudinal magnetization in our derivations, which make it appear as if we are dividing by a constant in Eq. 6, as the effect of total voxel volume cancels out for the relative FRE. We have now made this more explicit in our derivation of the partial volume model.

"Introducing a partial-volume model

To account for the effect of voxel volume on the FRE, the total longitudinal magnetization M_z needs to also consider the number of spins contained within in a voxel (Du et al., 1996; Venkatesan and Haacke, 1997). A simple approximation can be obtained by scaling the longitudinal magnetization with the voxel volume (Venkatesan and Haacke, 1997) . To then include partial volume effects, the total longitudinal magnetization in a voxel M_z^total becomes the sum of the contributions from the stationary tissue M_zS^tissue and the inflowing blood M_z^blood, weighted by their respective volume fractions V_rel:"

A simple approximation can be obtained by scaling the longitudinal magnetization with the voxel volume (Venkatesan and Haacke, 1997) . To then include partial volume effects, the total longitudinal magnetization in a voxel M_z^total becomes the sum of the contributions from the stationary tissue M_zS^tissue and the inflowing blood M_z^blood, weighted by their respective volume fractions V_rel:

Eq. (4)

For simplicity, we assume a single vessel is located at the center of the voxel and approximate it to be a cylinder with diameter d_vessel and length l_voxel of an assumed isotropic voxel along one side. The relative volume fraction of blood V_rel^blood is the ratio of vessel volume within the voxel to total voxel volume (see section Estimation of vessel-volume fraction in the Supplementary Material), and the tissue volume fraction V_rel^tissue is the remainder that is not filled with blood, or

Eq. (5)

We can now replace the blood magnetization in equation Eq. (3) with the total longitudinal magnetization of the voxel to compute the FRE as a function of vessel-volume fraction:

Eq. (6)

Based on your suggestion, we have also extended our interpretation of relative and absolute FRE. Indeed, a subtractive flow technique where no signal in the background remains and only intensities in the object are present would have infinite relative FRE, as this basically constitutes a perfect segmentation (bar a simple thresholding step).

"Extending classical FRE treatments to the pial vasculature

There are several major modifications in our approach to this topic that might explain why, in contrast to predictions from classical FRE treatments, it is indeed possible to image pial arteries. For instance, the definition of vessel contrast or flow-related enhancement is often stated as an absolute difference between blood and tissue signal (Brown et al., 2014a; Carr and Carroll, 2012; Du et al., 1993, 1996; Haacke et al., 1990; Venkatesan and Haacke, 1997). Here, however, we follow the approach of Al-Kwifi et al. (2002) and consider relative contrast. While this distinction may seem to be semantic, the effect of voxel volume on FRE for these two definitions is exactly opposite: Du et al. (1996) concluded that larger voxel size increases the (absolute) vessel-background contrast, whereas here we predict an increase in relative FRE for small arteries with decreasing voxel size. Therefore, predictions of the depiction of small arteries with decreasing voxel size differ depending on whether one is considering absolute contrast, i.e. difference in longitudinal magnetization, or relative contrast, i.e. contrast differences independent of total voxel size. Importantly, this prediction changes for large arteries where the voxel contains only vessel lumen, in which case the relative FRE remains constant across voxel sizes, but the absolute FRE increases with voxel size (Supplementary Figure 9). Overall, the interpretations of relative and absolute FRE differ, and one measure may be more appropriate for certain applications than the other. Absolute FRE describes the difference in magnetization and is thus tightly linked to the underlying physical mechanism. Relative FRE, however, describes the image contrast and segmentability. If blood and tissue magnetization are equal, both contrast measures would equal zero and indicate that no contrast difference is present. However, when there is signal in the vessel and as the tissue magnetization approaches zero, the absolute FRE approaches the blood magnetization (assuming no partial-volume effects), whereas the relative FRE approaches infinity. While this infinite relative FRE does not directly relate to the underlying physical process of ‘infinite’ signal enhancement through inflowing blood, it instead characterizes the segmentability of the image in that an image with zero intensity in the background and non-zero values in the structures of interest can be segmented perfectly and trivially. Accordingly, numerous empirical observations (Al-Kwifi et al., 2002; Bouvy et al., 2014; Haacke et al., 1990; Ladd, 2007; Mattern et al., 2018; von Morze et al., 2007) and the data provided here (Figure 5, 6 and 7) have shown the benefit of smaller voxel sizes if the aim is to visualize and segment small arteries."

9) Page 22, Line 400. "The appropriateness of " This also ignores noise. The absolute enhancement is the inherent magnetization available. The results in Figure 5, 6, 7 don't readily support a ratio over and absolute difference accounting for partial volume effects.

We hope that with the additional explanations on the effects of relative FRE definition in combination with a partial-volume model and the interpretation of relative FRE provided in the previous response (R2.8) and that Figures 5, 6 and 7 show smaller arteries for smaller voxels, we were able to clarify our argument why only relative FRE in combination with a partial volume model can explain why smaller voxel sizes are advantageous for depicting small arteries.

While we appreciate that there exists a fundamental relationship between SNR and voxel volume in MR (Brown et al., 2014b), this relationship is also modulated by many more factors (as we have argued in our responses to R2.2 and R1.4b).

We hope that the additional derivations and simulations provided in the previous response have clarified why a relative FRE model in combination with a partial-volume model helps to explain the enhanced detectability of small vessels with small voxels.

10) Page 24, Line 453. "strategies, such as radial and spiral acquisitions, experience no vessel displacement artefact" These do observe flow related distortions as well, just not typically called displacement.

Yes, this is a helpful point, as these methods will also experience a degradation of spatial accuracy due to flow effects, which will propagate into errors in the segmentation.

As the reviewer suggests, flow-related artefacts in radial and spiral acquisitions usually manifest as a slight blur, and less as the prominent displacement found in Cartesian sampling schemes. We have added a corresponding clarification to the Discussion section:

"Other encoding strategies, such as radial and spiral acquisitions, experience no vessel displacement artefact because phase and frequency encoding take place in the same instant; although a slight blur might be observed instead (Nishimura et al., 1995, 1991). However, both trajectories pose engineering challenges and much higher demands on hardware and reconstruction algorithms than the Cartesian readouts employed here (Kasper et al., 2018; Shu et al., 2016); particularly to achieve 3D acquisitions with 160 µm isotropic resolution."

11) Page 24, Line 272. "although even with this nearly ideal subject behaviour approximately 1 in 4 scans still had to be discarded and repeated" This is certainly a potential source of bias in the comparisons.

We apologize if this section was written in a misleading way. For the comparison presented in Figure 7, we acquired one additional slab in the same session at 0.16 mm voxel size using the same prospective motion correction procedure as for the 0.14 mm data. For the images shown in Figure 6 and Supplementary Figure 4 at 0.16 mm voxel size, we did not use a motion correction system and, thus, had to discard a portion of the data. We have clarified that for the comparison of the high-resolution data, prospective motion correction was used for both resolutions. We have clarified this in the Discussion section:

"This allowed for the successful correction of head motion of approximately 1 mm over the 60-minute scan session, showing the utility of prospective motion correction at these very high resolutions. Note that for the comparison in Figure 7, one slab with 0.16 mm voxel size was acquired in the same session also using the prospective motion correction system. However, for the data shown in Figure 6 and Supplementary Figure 4, no prospective motion correction was used, and we instead relied on the experienced participants who contributed to this study. We found that the acquisition of TOF data with 0.16 mm isotropic voxel size in under 12 minutes acquisition time per slab is possible without discernible motion artifacts, although even with this nearly ideal subject behaviour approximately 1 in 4 scans still had to be discarded and repeated."

12) Page 25, Line 489. "then need to include the effects of various analog and digital filters" While the analysis may benefit from some of this, most is not at all required for analysis based on optimization of the imaging parameters.

We have included all four correction factors for completeness, given the unique acquisition parameter and contrast space our time-of-flight acquisition occupies, e.g. very low bandwidth of only 100 Hz, very large matrix sizes > 1024 samples, ideally zero SNR in the background (fully supressed tissue signal). However, we agree that probably the most important factor is the non-central chi distribution of the noise in magnitude images from multiple-channel coil arrays, and have added this qualification in the text:

"Accordingly, SNR predictions then need to include the effects of various analog and digital filters, the number of acquired samples, the noise covariance correction factor, and—most importantly—the non-central chi distribution of the noise statistics of the final magnitude image (Triantafyllou et al., 2011)."

Al-Kwifi, O., Emery, D.J., Wilman, A.H., 2002. Vessel contrast at three Tesla in time-of-flight magnetic resonance angiography of the intracranial and carotid arteries. Magnetic Resonance Imaging 20, 181–187. https://doi.org/10.1016/S0730-725X(02)00486-1

Arts, T., Meijs, T.A., Grotenhuis, H., Voskuil, M., Siero, J., Biessels, G.J., Zwanenburg, J., 2021. Velocity and Pulsatility Measures in the Perforating Arteries of the Basal Ganglia at 3T MRI in Reference to 7T MRI. Frontiers in Neuroscience 15. Avants, B.B., Tustison, N., Song, G., 2009. Advanced normalization tools (ANTS). Insight j 2, 1–35. Bae, K.T., Park, S.-H., Moon, C.-H., Kim, J.-H., Kaya, D., Zhao, T., 2010. Dual-echo arteriovenography imaging with 7T MRI: CODEA with 7T. J. Magn. Reson. Imaging 31, 255–261. https://doi.org/10.1002/jmri.22019

Bartholdi, E., Ernst, R.R., 1973. Fourier spectroscopy and the causality principle. Journal of Magnetic Resonance (1969) 11, 9–19. https://doi.org/10.1016/0022-2364(73)90076-0

Bernier, M., Cunnane, S.C., Whittingstall, K., 2018. The morphology of the human cerebrovascular system. Human Brain Mapping 39, 4962–4975. https://doi.org/10.1002/hbm.24337

Bouvy, W.H., Biessels, G.J., Kuijf, H.J., Kappelle, L.J., Luijten, P.R., Zwanenburg, J.J.M., 2014. Visualization of Perivascular Spaces and Perforating Arteries With 7 T Magnetic Resonance Imaging: Investigative Radiology 49, 307–313. https://doi.org/10.1097/RLI.0000000000000027

Bouvy, W.H., Geurts, L.J., Kuijf, H.J., Luijten, P.R., Kappelle, L.J., Biessels, G.J., Zwanenburg, J.J.M., 2016. Assessment of blood flow velocity and pulsatility in cerebral perforating arteries with 7-T quantitative flow MRI: Blood Flow Velocity And Pulsatility In Cerebral Perforating Arteries. NMR Biomed. 29, 1295–1304. https://doi.org/10.1002/nbm.3306

Brown, R.W., Cheng, Y.-C.N., Haacke, E.M., Thompson, M.R., Venkatesan, R., 2014a. Chapter 24 - MR Angiography and Flow Quantification, in: Magnetic Resonance Imaging. John Wiley & Sons, Ltd, pp. 701–737. https://doi.org/10.1002/9781118633953.ch24

Brown, R.W., Cheng, Y.-C.N., Haacke, E.M., Thompson, M.R., Venkatesan, R., 2014b. Chapter 15 - Signal, Contrast, and Noise, in: Magnetic Resonance Imaging. John Wiley & Sons, Ltd, pp. 325–373. https://doi.org/10.1002/9781118633953.ch15

Carr, J.C., Carroll, T.J., 2012. Magnetic resonance angiography: principles and applications. Springer, New York. Cassot, F., Lauwers, F., Fouard, C., Prohaska, S., Lauwers-Cances, V., 2006. A Novel Three-Dimensional Computer-Assisted Method for a Quantitative Study of Microvascular Networks of the Human Cerebral Cortex. Microcirculation 13, 1–18. https://doi.org/10.1080/10739680500383407

Chen, L., Mossa-Basha, M., Balu, N., Canton, G., Sun, J., Pimentel, K., Hatsukami, T.S., Hwang, J.-N., Yuan, C., 2018. Development of a quantitative intracranial vascular features extraction tool on 3DMRA using semiautomated open-curve active contour vessel tracing: Comprehensive Artery Features Extraction From 3D MRA. Magn. Reson. Med 79, 3229–3238. https://doi.org/10.1002/mrm.26961

Choi, U.-S., Kawaguchi, H., Kida, I., 2020. Cerebral artery segmentation based on magnetization-prepared two rapid acquisition gradient echo multi-contrast images in 7 Tesla magnetic resonance imaging. NeuroImage 222, 117259. https://doi.org/10.1016/j.neuroimage.2020.117259

Conolly, S., Nishimura, D., Macovski, A., Glover, G., 1988. Variable-rate selective excitation. Journal of Magnetic Resonance (1969) 78, 440–458. https://doi.org/10.1016/0022-2364(88)90131-X

Deistung, A., Dittrich, E., Sedlacik, J., Rauscher, A., Reichenbach, J.R., 2009. ToF-SWI: Simultaneous time of flight and fully flow compensated susceptibility weighted imaging. J. Magn. Reson. Imaging 29, 1478–1484. https://doi.org/10.1002/jmri.21673

Detre, J.A., Leigh, J.S., Williams, D.S., Koretsky, A.P., 1992. Perfusion imaging. Magnetic Resonance in Medicine 23, 37–45. https://doi.org/10.1002/mrm.1910230106

Du, Y., Parker, D.L., Davis, W.L., Blatter, D.D., 1993. Contrast-to-Noise-Ratio Measurements in Three-Dimensional Magnetic Resonance Angiography. Investigative Radiology 28, 1004–1009. Du, Y.P., Jin, Z., 2008. Simultaneous acquisition of MR angiography and venography (MRAV). Magn. Reson. Med. 59, 954–958. https://doi.org/10.1002/mrm.21581

Du, Y.P., Parker, D.L., Davis, W.L., Cao, G., 1994. Reduction of partial-volume artifacts with zero-filled interpolation in three-dimensional MR angiography. J. Magn. Reson. Imaging 4, 733–741. https://doi.org/10.1002/jmri.1880040517

Du, Y.P., Parker, D.L., Davis, W.L., Cao, G., Buswell, H.R., Goodrich, K.C., 1996. Experimental and theoretical studies of vessel contrast-to-noise ratio in intracranial time-of-flight MR angiography. Journal of Magnetic Resonance Imaging 6, 99–108. https://doi.org/10.1002/jmri.1880060120

Duvernoy, H., Delon, S., Vannson, J.L., 1983. The Vascularization of The Human Cerebellar Cortex. Brain Research Bulletin 11, 419–480. Duvernoy, H.M., Delon, S., Vannson, J.L., 1981. Cortical blood vessels of the human brain. Brain Research Bulletin 7, 519–579. https://doi.org/10.1016/0361-9230(81)90007-1

Eckstein, K., Bachrata, B., Hangel, G., Widhalm, G., Enzinger, C., Barth, M., Trattnig, S., Robinson, S.D., 2021. Improved susceptibility weighted imaging at ultra-high field using bipolar multi-echo acquisition and optimized image processing: CLEAR-SWI. NeuroImage 237, 118175. https://doi.org/10.1016/j.neuroimage.2021.118175

Edelstein, W.A., Glover, G.H., Hardy, C.J., Redington, R.W., 1986. The intrinsic signal-to-noise ratio in NMR imaging. Magn. Reson. Med. 3, 604–618. https://doi.org/10.1002/mrm.1910030413

Fan, A.P., Govindarajan, S.T., Kinkel, R.P., Madigan, N.K., Nielsen, A.S., Benner, T., Tinelli, E., Rosen, B.R., Adalsteinsson, E., Mainero, C., 2015. Quantitative oxygen extraction fraction from 7-Tesla MRI phase: reproducibility and application in multiple sclerosis. J Cereb Blood Flow Metab 35, 131–139. https://doi.org/10.1038/jcbfm.2014.187

Fiedler, T.M., Ladd, M.E., Bitz, A.K., 2018. SAR Simulations & Safety. NeuroImage 168, 33–58. https://doi.org/10.1016/j.neuroimage.2017.03.035

Frässle, S., Aponte, E.A., Bollmann, S., Brodersen, K.H., Do, C.T., Harrison, O.K., Harrison, S.J., Heinzle, J., Iglesias, S., Kasper, L., Lomakina, E.I., Mathys, C., Müller-Schrader, M., Pereira, I., Petzschner, F.H., Raman, S., Schöbi, D., Toussaint, B., Weber, L.A., Yao, Y., Stephan, K.E., 2021. TAPAS: An Open-Source Software Package for Translational Neuromodeling and Computational Psychiatry. Front. Psychiatry 12. https://doi.org/10.3389/fpsyt.2021.680811

Gulban, O.F., Bollmann, S., Huber, R., Wagstyl, K., Goebel, R., Poser, B.A., Kay, K., Ivanov, D., 2021. Mesoscopic Quantification of Cortical Architecture in the Living Human Brain. https://doi.org/10.1101/2021.11.25.470023

Haacke, E.M., Masaryk, T.J., Wielopolski, P.A., Zypman, F.R., Tkach, J.A., Amartur, S., Mitchell, J., Clampitt, M., Paschal, C., 1990. Optimizing blood vessel contrast in fast three-dimensional MRI. Magn. Reson. Med. 14, 202–221. https://doi.org/10.1002/mrm.1910140207

Helthuis, J.H.G., van Doormaal, T.P.C., Hillen, B., Bleys, R.L.A.W., Harteveld, A.A., Hendrikse, J., van der Toorn, A., Brozici, M., Zwanenburg, J.J.M., van der Zwan, A., 2019. Branching Pattern of the Cerebral Arterial Tree. Anat Rec 302, 1434–1446. https://doi.org/10.1002/ar.23994

Heverhagen, J.T., Bourekas, E., Sammet, S., Knopp, M.V., Schmalbrock, P., 2008. Time-of-Flight Magnetic Resonance Angiography at 7 Tesla. Investigative Radiology 43, 568–573. https://doi.org/10.1097/RLI.0b013e31817e9b2c

Hirsch, S., Reichold, J., Schneider, M., Székely, G., Weber, B., 2012. Topology and Hemodynamics of the Cortical Cerebrovascular System. J Cereb Blood Flow Metab 32, 952–967. https://doi.org/10.1038/jcbfm.2012.39

Horn, B.K.P., Schunck, B.G., 1981. Determining optical flow. Artificial Intelligence 17, 185–203. https://doi.org/10.1016/0004-3702(81)90024-2

Huck, J., Wanner, Y., Fan, A.P., Jäger, A.-T., Grahl, S., Schneider, U., Villringer, A., Steele, C.J., Tardif, C.L., Bazin, P.-L., Gauthier, C.J., 2019. High resolution atlas of the venous brain vasculature from 7 T quantitative susceptibility maps. Brain Struct Funct 224, 2467–2485. https://doi.org/10.1007/s00429-019-01919-4

Johst, S., Wrede, K.H., Ladd, M.E., Maderwald, S., 2012. Time-of-Flight Magnetic Resonance Angiography at 7 T Using Venous Saturation Pulses With Reduced Flip Angles. Investigative Radiology 47, 445–450. https://doi.org/10.1097/RLI.0b013e31824ef21f

Kang, C.-K., Park, C.-A., Kim, K.-N., Hong, S.-M., Park, C.-W., Kim, Y.-B., Cho, Z.-H., 2010. Non-invasive visualization of basilar artery perforators with 7T MR angiography. Journal of Magnetic Resonance Imaging 32, 544–550. https://doi.org/10.1002/jmri.22250

Kasper, L., Engel, M., Barmet, C., Haeberlin, M., Wilm, B.J., Dietrich, B.E., Schmid, T., Gross, S., Brunner, D.O., Stephan, K.E., Pruessmann, K.P., 2018. Rapid anatomical brain imaging using spiral acquisition and an expanded signal model. NeuroImage 168, 88–100. https://doi.org/10.1016/j.neuroimage.2017.07.062

Klepaczko, A., Szczypiński, P., Deistung, A., Reichenbach, J.R., Materka, A., 2016. Simulation of MR angiography imaging for validation of cerebral arteries segmentation algorithms. Computer Methods and Programs in Biomedicine 137, 293–309. https://doi.org/10.1016/j.cmpb.2016.09.020

Kobari, M., Gotoh, F., Fukuuchi, Y., Tanaka, K., Suzuki, N., Uematsu, D., 1984. Blood Flow Velocity in the Pial Arteries of Cats, with Particular Reference to the Vessel Diameter. J Cereb Blood Flow Metab 4, 110–114. https://doi.org/10.1038/jcbfm.1984.15

Ladd, M.E., 2007. High-Field-Strength Magnetic Resonance: Potential and Limits. Top Magn Reson Imaging 18, 139–152. Lesage, D., Angelini, E.D., Bloch, I., Funka-Lea, G., 2009. A review of 3D vessel lumen segmentation techniques: Models, features and extraction schemes. Medical Image Analysis 13, 819–845. https://doi.org/10.1016/j.media.2009.07.011

Maderwald, S., Ladd, S.C., Gizewski, E.R., Kraff, O., Theysohn, J.M., Wicklow, K., Moenninghoff, C., Wanke, I., Ladd, M.E., Quick, H.H., 2008. To TOF or not to TOF: strategies for non-contrast-enhanced intracranial MRA at 7 T. Magn Reson Mater Phy 21, 159. https://doi.org/10.1007/s10334-007-0096-9

Manjón, J.V., Coupé, P., Martí‐Bonmatí, L., Collins, D.L., Robles, M., 2010. Adaptive non-local means denoising of MR images with spatially varying noise levels. Journal of Magnetic Resonance Imaging 31, 192–203. https://doi.org/10.1002/jmri.22003

Mansfield, P., Harvey, P.R., 1993. Limits to neural stimulation in echo-planar imaging. Magn. Reson. Med. 29, 746–758. https://doi.org/10.1002/mrm.1910290606

Masaryk, T.J., Modic, M.T., Ross, J.S., Ruggieri, P.M., Laub, G.A., Lenz, G.W., Haacke, E.M., Selman, W.R., Wiznitzer, M., Harik, S.I., 1989. Intracranial circulation: preliminary clinical results with three-dimensional (volume) MR angiography. Radiology 171, 793–799. https://doi.org/10.1148/radiology.171.3.2717754

Mattern, H., Sciarra, A., Godenschweger, F., Stucht, D., Lüsebrink, F., Rose, G., Speck, O., 2018. Prospective motion correction enables highest resolution time-of-flight angiography at 7T: Prospectively Motion-Corrected TOF Angiography at 7T. Magn. Reson. Med 80, 248–258. https://doi.org/10.1002/mrm.27033

Mattern, H., Sciarra, A., Lüsebrink, F., Acosta‐Cabronero, J., Speck, O., 2019. Prospective motion correction improves high‐resolution quantitative susceptibility mapping at 7T. Magn. Reson. Med 81, 1605–1619. https://doi.org/10.1002/mrm.27509

Mennes, M., Jenkinson, M., Valabregue, R., Buitelaar, J.K., Beckmann, C., Smith, S., 2014. Optimizing full-brain coverage in human brain MRI through population distributions of brain size. NeuroImage 98, 513–520. https://doi.org/10.1016/j.neuroimage.2014.04.030 Moccia, S., De Momi, E., El Hadji, S., Mattos, L.S., 2018. Blood vessel segmentation algorithms — Review of methods, datasets and evaluation metrics. Computer Methods and Programs in Biomedicine 158, 71–91. https://doi.org/10.1016/j.cmpb.2018.02.001

Mustafa, M.A.R., 2016. A data-driven learning approach to image registration. Mut, F., Wright, S., Ascoli, G.A., Cebral, J.R., 2014. Morphometric, geographic, and territorial characterization of brain arterial trees. International Journal for Numerical Methods in Biomedical Engineering 30, 755–766. https://doi.org/10.1002/cnm.2627

Nagaoka, T., Yoshida, A., 2006. Noninvasive Evaluation of Wall Shear Stress on Retinal Microcirculation in Humans. Invest. Ophthalmol. Vis. Sci. 47, 1113. https://doi.org/10.1167/iovs.05-0218

Nishimura, D.G., Irarrazabal, P., Meyer, C.H., 1995. A Velocity k-Space Analysis of Flow Effects in Echo-Planar and Spiral Imaging. Magnetic Resonance in Medicine 33, 549–556. https://doi.org/10.1002/mrm.1910330414

Nishimura, D.G., Jackson, J.I., Pauly, J.M., 1991. On the nature and reduction of the displacement artifact in flow images. Magnetic Resonance in Medicine 22, 481–492. https://doi.org/10.1002/mrm.1910220255

Nonaka, H., Akima, M., Hatori, T., Nagayama, T., Zhang, Z., Ihara, F., 2003. Microvasculature of the human cerebral white matter: Arteries of the deep white matter. Neuropathology 23, 111–118. https://doi.org/10.1046/j.1440-1789.2003.00486.x

North, D.O., 1963. An Analysis of the factors which determine signal/noise discrimination in pulsed-carrier systems. Proceedings of the IEEE 51, 1016–1027. https://doi.org/10.1109/PROC.1963.2383

Park, C.S., Hartung, G., Alaraj, A., Du, X., Charbel, F.T., Linninger, A.A., 2020. Quantification of blood flow patterns in the cerebral arterial circulation of individual (human) subjects. Int J Numer Meth Biomed Engng 36. https://doi.org/10.1002/cnm.3288

Parker, D.L., Goodrich, K.C., Roberts, J.A., Chapman, B.E., Jeong, E.-K., Kim, S.-E., Tsuruda, J.S., Katzman, G.L., 2003. The need for phase-encoding flow compensation in high-resolution intracranial magnetic resonance angiography. J. Magn. Reson. Imaging 18, 121–127. https://doi.org/10.1002/jmri.10322

Parker, D.L., Yuan, C., Blatter, D.D., 1991. MR angiography by multiple thin slab 3D acquisition. Magn. Reson. Med. 17, 434–451. https://doi.org/10.1002/mrm.1910170215

Pauling, L., Coryell, C.D., 1936. The magnetic properties and structure of hemoglobin, oxyhemoglobin and carbonmonoxyhemoglobin. Proceedings of the National Academy of Sciences 22, 210–216. https://doi.org/10.1073/pnas.22.4.210

Payne, S.J., 2017. Cerebral Blood Flow And Metabolism: A Quantitative Approach. World Scientific. Peters, A.M., Brookes, M.J., Hoogenraad, F.G., Gowland, P.A., Francis, S.T., Morris, P.G., Bowtell, R., 2007. T2* measurements in human brain at 1.5, 3 and 7 T. Magnetic Resonance Imaging 25, 748–753. https://doi.org/10.1016/j.mri.2007.02.014

Pfeifer, R.A., 1930. Grundlegende Untersuchungen für die Angioarchitektonik des menschlichen Gehirns. Berlin: Julius Springer. Phellan, R., Forkert, N.D., 2017. Comparison of vessel enhancement algorithms applied to time-of-flight MRA images for cerebrovascular segmentation. Medical Physics 44, 5901–5915. https://doi.org/10.1002/mp.12560

Pohmann, R., Speck, O., Scheffler, K., 2016. Signal-to-Noise Ratio and MR Tissue Parameters in Human Brain Imaging at 3, 7, and 9.4 Tesla Using Current Receive Coil Arrays. Magn. Reson. Med. 75, 801–809. https://doi.org/10.1002/mrm.25677

Reichenbach, J.R., Venkatesan, R., Schillinger, D.J., Kido, D.K., Haacke, E.M., 1997. Small vessels in the human brain: MR venography with deoxyhemoglobin as an intrinsic contrast agent. Radiology 204, 272–277. https://doi.org/10.1148/radiology.204.1.9205259 Schmid, F., Barrett, M.J.P., Jenny, P., Weber, B., 2019. Vascular density and distribution in neocortex. NeuroImage 197, 792–805. https://doi.org/10.1016/j.neuroimage.2017.06.046

Schmitter, S., Bock, M., Johst, S., Auerbach, E.J., Uğurbil, K., Moortele, P.-F.V. de, 2012. Contrast enhancement in TOF cerebral angiography at 7 T using saturation and MT pulses under SAR constraints: Impact of VERSE and sparse pulses. Magnetic Resonance in Medicine 68, 188–197. https://doi.org/10.1002/mrm.23226

Schulz, J., Boyacioglu, R., Norris, D.G., 2016. Multiband multislab 3D time-of-flight magnetic resonance angiography for reduced acquisition time and improved sensitivity. Magn Reson Med 75, 1662–8. https://doi.org/10.1002/mrm.25774

Shu, C.Y., Sanganahalli, B.G., Coman, D., Herman, P., Hyder, F., 2016. New horizons in neurometabolic and neurovascular coupling from calibrated fMRI, in: Progress in Brain Research. Elsevier, pp. 99–122. https://doi.org/10.1016/bs.pbr.2016.02.003

Stamm, A.C., Wright, C.L., Knopp, M.V., Schmalbrock, P., Heverhagen, J.T., 2013. Phase contrast and time-of-flight magnetic resonance angiography of the intracerebral arteries at 1.5, 3 and 7 T. Magnetic Resonance Imaging 31, 545–549. https://doi.org/10.1016/j.mri.2012.10.023

Stewart, A.W., Robinson, S.D., O’Brien, K., Jin, J., Widhalm, G., Hangel, G., Walls, A., Goodwin, J., Eckstein, K., Tourell, M., Morgan, C., Narayanan, A., Barth, M., Bollmann, S., 2022. QSMxT: Robust masking and artifact reduction for quantitative susceptibility mapping. Magnetic Resonance in Medicine 87, 1289–1300. https://doi.org/10.1002/mrm.29048

Stucht, D., Danishad, K.A., Schulze, P., Godenschweger, F., Zaitsev, M., Speck, O., 2015. Highest Resolution In Vivo Human Brain MRI Using Prospective Motion Correction. PLoS ONE 10, e0133921. https://doi.org/10.1371/journal.pone.0133921

Szikla, G., Bouvier, G., Hori, T., Petrov, V., 1977. Angiography of the Human Brain Cortex. Springer Berlin Heidelberg, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-81145-6

Triantafyllou, C., Polimeni, J.R., Wald, L.L., 2011. Physiological noise and signal-to-noise ratio in fMRI with multi-channel array coils. NeuroImage 55, 597–606. https://doi.org/10.1016/j.neuroimage.2010.11.084

Tustison, N.J., Avants, B.B., Cook, P.A., Zheng, Y., Egan, A., Yushkevich, P.A., Gee, J.C., 2010. N4ITK: Improved N3 Bias Correction. IEEE Transactions on Medical Imaging 29, 1310–1320. https://doi.org/10.1109/TMI.2010.2046908

Uludağ, K., Müller-Bierl, B., Uğurbil, K., 2009. An integrative model for neuronal activity-induced signal changes for gradient and spin echo functional imaging. NeuroImage 48, 150–165. https://doi.org/10.1016/j.neuroimage.2009.05.051

Venkatesan, R., Haacke, E.M., 1997. Role of high resolution in magnetic resonance (MR) imaging: Applications to MR angiography, intracranial T1-weighted imaging, and image interpolation. International Journal of Imaging Systems and Technology 8, 529–543. https://doi.org/10.1002/(SICI)1098-1098(1997)8:6<529::AID-IMA5>3.0.CO;2-C

von Morze, C., Xu, D., Purcell, D.D., Hess, C.P., Mukherjee, P., Saloner, D., Kelley, D.A.C., Vigneron, D.B., 2007. Intracranial time-of-flight MR angiography at 7T with comparison to 3T. J. Magn. Reson. Imaging 26, 900–904. https://doi.org/10.1002/jmri.21097

Ward, P.G.D., Ferris, N.J., Raniga, P., Dowe, D.L., Ng, A.C.L., Barnes, D.G., Egan, G.F., 2018. Combining images and anatomical knowledge to improve automated vein segmentation in MRI. NeuroImage 165, 294–305. https://doi.org/10.1016/j.neuroimage.2017.10.049

Wilms, G., Bosmans, H., Demaerel, Ph., Marchal, G., 2001. Magnetic resonance angiography of the intracranial vessels. European Journal of Radiology 38, 10–18. https://doi.org/10.1016/S0720-048X(01)00285-6

Wright, S.N., Kochunov, P., Mut, F., Bergamino, M., Brown, K.M., Mazziotta, J.C., Toga, A.W., Cebral, J.R., Ascoli, G.A., 2013. Digital reconstruction and morphometric analysis of human brain arterial vasculature from magnetic resonance angiography. NeuroImage 82, 170–181. https://doi.org/10.1016/j.neuroimage.2013.05.089

Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., Gerig, G., 2006. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. NeuroImage 31, 1116–1128. https://doi.org/10.1016/j.neuroimage.2006.01.015

Zhang, Z., Deng, X., Weng, D., An, J., Zuo, Z., Wang, B., Wei, N., Zhao, J., Xue, R., 2015. Segmented TOF at 7T MRI: Technique and clinical applications. Magnetic Resonance Imaging 33, 1043–1050. https://doi.org/10.1016/j.mri.2015.07.002

Zhao, J.M., Clingman, C.S., Närväinen, M.J., Kauppinen, R.A., van Zijl, P.C.M., 2007. Oxygenation and hematocrit dependence of transverse relaxation rates of blood at 3T. Magn. Reson. Med. 58, 592–597. https://doi.org/10.1002/mrm.21342

Zhu, X., Tomanek, B., Sharp, J., 2013. A pixel is an artifact: On the necessity of zero-filling in fourier imaging. Concepts Magn. Reson. 42A, 32–44. https://doi.org/10.1002/cmr.a.21256

Author Response

Reviewer #1 (Public Review):

The authors present a PyTorch-based simulator for prosthetic vision. The model takes in the anatomical location of a visual cortical prostheses as well as a series of electrical stimuli to be applied to each electrode, and outputs the resulting phosphenes. To demonstrate the usefulness of the simulator, the paper reproduces psychometric curves from the literature and uses the simulator in the loop to learn optimized stimuli.

One of the major strengths of the paper is its modeling work - the authors make good use of existing knowledge about retinotopic maps and psychometric curves that describe phosphene appearance in response to single-electrode stimulation. Using PyTorch as a backbone is another strength, as it allows for GPU integration and seamless integration with common deep learning models. This work is likely to be impactful for the field of sight restoration.

1) However, one of the major weaknesses of the paper is its model validation - while some results seem to be presented for data the model was fit on (as opposed to held-out test data), other results lack quantitative metrics and a comparison to a baseline ("null hypothesis") model. On the one hand, it appears that the data presented in Figs. 3-5 was used to fit some of the open parameters of the model, as mentioned in Subsection G of the Methods. Hence it is misleading to present these as model "predictions", which are typically presented for held-out test data to demonstrate a model's ability to generalize. Instead, this is more of a descriptive model than a predictive one, and its ability to generalize to new patients remains yet to be demonstrated.

We agree that the original presentation of the model fits might give rise to unwanted confusion. In the revision, we have adapted the fit of the thresholding mechanism to include a 3-fold cross validation, where part of the data was excluded during the fitting, and used as test sets to calculate the model’s performance. The results of the cross- validation are now presented in panel D of Figure 3. The fitting of the brightness and temporal dynamics parameters using cross-validation was not feasible due to the limited amount of quantitative data describing temporal dynamics and phosphene size and brightness for intracortical electrodes. To avoid confusion, we have adapted the corresponding text and figure captions to specify that we are using a fit as description of the data.

We note that the goal of the simulator is not to provide a single set of parameters that describes precise phosphene perception for all patients but that it could also be used to capture variability among patients. Indeed, the model can be tailored to new patients based on a small data set. Figure 3-figure supplement 1 exemplifies how our simulator can be tailored to several data sets collected from patients with surface electrodes. Future clinical experiments might be used to verify how well the simulator can be tailored to the data of other patients.

Specifically, we have made the following changes to the manuscript:

• Caption Figure 2: the fitted peak brightness levels reproduced by our model

• Caption Figure 3: The model's probability of phosphene perception is visualized as a function of charge per phase

• Caption Figure 3: Predicted probabilities in panel (d) are the results of a 3-fold cross- validation on held-out test data.

• Line 250: we included biologically inspired methods to model the perceptual effects of different stimulation parameters

• Line 271: Each frame, the simulator maps electrical stimulation parameters (stimulation current, pulse width and frequency) to an estimated phosphene perception

• Lines 335-336: such that 95% of the Gaussian falls within the fitted phosphene size.

• Line 469-470: Figure 4 displays the simulator's fit on the temporal dynamics found in a previous published study by Schmidt et al. (1996).

• Lines 922-925: Notably, the trade-off between model complexity and accurate psychophysical fits or predictions is a recurrent theme in the validation of the components implemented in our simulator.

2) On the other hand, the results presented in Fig. 8 as part of the end-to-end learning process are not accompanied by any sorts of quantitative metrics or comparison to a baseline model.

We now realize that the presentation of the end-to-end results might have given the impression that we present novel image processing strategies. However, the development of a novel image processing strategy is outside the scope of the study. Instead, The study aims to provide an improved simulation which can be used for more realistic assessment of different stimulation protocols. The simulator needs to fit experimental data, and it should run fast (so it can be used in behavioral experiments). Importantly, as demonstrated in our end-to-end experiments, the model can be used in differentiable programming pipelines (so it can be used in computational optimization experiments), which is a valuable contribution in itself because it lends itself to many machine learning approaches which can improve the realism of the simulation.

We have rephrased our study aims in the discussion to improve clarity.

• Lines 275-279: In the sections below, we discuss the different components of the simulator model, followed by a description of some showcase experiments that assess the ability to fit recent clinical data and the practical usability of our simulator in simulation experiments

• Lines 810-814: Computational optimization approaches can also aid in the development of safe stimulation protocols, because they allow a faster exploration of the large parameter space and enable task-driven optimization of image processing strategies (Granley et al., 2022; Fauvel et al., 2022; White et al., 2019; Küçükoglü et al. 2022; de Ruyter van Steveninck et al., 2022; Ghaffari et al., 2021).

• Lines 814-819: Ultimately, the development of task-relevant scene-processing algorithms will likely benefit both from computational optimization experiments as well as exploratory SPV studies with human observers. With the presented simulator we aim to contribute a flexible toolkit for such experiments.

• Lines 842-853: Eventually, the functional quality of the artificial vision will not only depend on the correspondence between the visual environment and the phosphene encoding, but also on the implant recipient's ability to extract that information into a usable percept. The functional quality of end-to-end generated phosphene encodings in daily life tasks will need to be evaluated in future experiments. Regardless of the implementation, it will always be important to include human observers (both sighted experimental subjects and actual prosthetic implant users in the optimization cycle to ensure subjective interpretability for the end user (Fauvel et al., 2022; Beyeler & Sanchez-Garcia, 2022).

3) The results seem to assume that all phosphenes are small Gaussian blobs, and that these phosphenes combine linearly when multiple electrodes are stimulated. Both assumptions are frequently challenged by the field. For all these reasons, it is challenging to assess the potential and practical utility of this approach as well as get a sense of its limitations.

The reviewer raises a valid point and a similar point was raised by a different reviewer (our response is duplicated). As pointed out in the discussion, many aspects about multi- electrode phosphene perception are still unclear. On the one hand, the literature is in agreement that there is some degree of predictability: some papers explicitly state that phosphenes produced by multiple patterns are generally additive (Dobelle & Mladejovsky, 1974), that the locations are predictable (Bosking et al., 2018) and that multi-electrode stimulation can be used to generate complex, interpretable patterns of phosphenes (Chen et al., 2020, Fernandez et al., 2021). On the other hand, however, in some cases, the stimulation of multiple electrodes is reported to lead to brighter phosphenes (Fernandez et al., 2021), fused or displaced phosphenes (Schmidt et al., 1996, Bak et al., 1990) or unpredicted phosphene patterns (Fernández et al., 2021). It is likely that the probability of these interference patterns decreases when the distance between the stimulated electrodes increases. An empirical finding is that the critical distance for intracortical stimulation is approximately 1 mm (Ghose & Maunsell, 2012).

We note that our simulator is not restricted to the simulation of linearly combined Gaussian blobs. Some irregularities, such as elongated phosphene shapes were already supported in the previous version of our software. Furthermore, we added a supplementary figure that displays a possible approach to simulate some of the more complex electrode interactions that are reported in the literature, with only minor adaptations to the code. Our study thereby aims to present a flexible simulation toolkit that can be adapted to the needs of the user.

Adjustments:

• Added Figure 1-figure supplement 3 on irregular phosphene percepts.

• Lines 957-970: Furthermore, in contrast to the assumptions of our model, interactions between simultaneous stimulation of multiple electrodes can have an effect on the phosphene size and sometimes lead to unexpected percepts (Fernandez et al., 2021, Dobelle & Mladejovsky 1974, Bak et al., 1990). Although our software supports basic exploratory experimentation of non-linear interactions (see Figure 1-figure supplement 3), by default, our simulator assumes independence between electrodes. Multi- phosphene percepts are modeled using linear summation of the independent percepts. These assumptions seem to hold for intracortical electrodes separated by more than 1 mm (Ghose & Maunsell, 2012), but may underestimate the complexities observed when electrodes are nearer. Further clinical and theoretical modeling work could help to improve our understanding of these non-linear dynamics.

4) Another weakness of the paper is the term "biologically plausible", which appears throughout the manuscript but is not clearly defined. In its current form, it is not clear what makes this simulator "biologically plausible" - it certainly contains a retinotopic map and is fit on psychophysical data, but it does not seem to contain any other "biological" detail.

We thank the reviewer for the remark. We improved our description of what makes the simulator “biologically plausible” in the introduction (line 78): ‘‘Biological plausibility, in our work's context, points to the simulation's ability to capture essential biological features of the visual system in a manner consistent with empirical findings: our simulator integrates quantitative findings and models from the literature on cortical stimulation in V1 [...]”. In addition, we mention in the discussion (lines 611 - 621): “The aim of this study is to present a biologically plausible phosphene simulator, which takes realistic ranges of stimulation parameters, and generates a phenomenologically accurate representation of phosphene vision using differentiable functions. In order to achieve this, we have modeled and incorporated an extensive body of work regarding the psychophysics of phosphene perception. From the results presented in section H, we observe that our simulator is able to produce phosphene percepts that match the descriptions of phosphene vision that were gathered in basic and clinical visual neuroprosthetics studies over the past decades.”

5) In fact, for the most part the paper seems to ignore the fact that implanting a prosthesis in one cerebral hemisphere will produce phosphenes that are restricted to one half of the visual field. Yet Figures 6 and 8 present phosphenes that seemingly appear in both hemifields. I do not find this very "biologically plausible".

We agree with the reviewer that contemporary experiments with implantable electrodes usually test electrodes in a single hemisphere. However, future clinically useful approaches should use bilaterally implanted electrode arrays. Our simulator can either present phosphene locations in either one or both hemifields.

We have made the following textual changes:

• Fig. 1 caption: Example renderings after initializing the simulator with four 10 × 10 electrode arrays (indicated with roman numerals) placed in the right hemisphere (electrode spacing: 4 mm, in correspondence with the commonly used 'Utah array' (Maynard et al., 1997)).

• Line 518-525: The simulator is initialized with 1000 possible phosphenes in both hemifields, covering a field of view of 16 degrees of visual angle. Note that the simulated electrode density and placement differs from current prototype implants and the simulation can be considered to be an ambitious scenario from a surgical point of view, given the folding of the visual cortex and the part of the retinotopic map in V1 that is buried in the calcarine sulcus. Line 546-547: with the same phosphene coverage as the previously described experiment

Reviewer #2 (Public Review):

Van der Grinten and De Ruyter van Steveninck et al. present a design for simulating cortical- visual-prosthesis phosphenes that emphasizes features important for optimizing the use of such prostheses. The characteristics of simulated individual phosphenes were shown to agree well with data published from the use of cortical visual prostheses in humans. By ensuring that functions used to generate the simulations were differentiable, the authors permitted and demonstrated integration of the simulations into deep-learning algorithms. In concept, such algorithms could thereby identify parameters for translating images or videos into stimulation sequences that would be most effective for artificial vision. There are, however, limitations to the simulation that will limit its applicability to current prostheses.

The verification of how phosphenes are simulated for individual electrodes is very compelling. Visual-prosthesis simulations often do ignore the physiologic foundation underlying the generation of phosphenes. The authors' simulation takes into account how stimulation parameters contribute to phosphene appearance and show how that relationship can fit data from actual implanted volunteers. This provides an excellent foundation for determining optimal stimulation parameters with reasonable confidence in how parameter selections will affect individual-electrode phosphenes.

We thank the reviewer for these supportive comments.

Issues with the applicability and reliability of the simulation are detailed below:

1) The utility of this simulation design, as described, unfortunately breaks down beyond the scope of individual electrodes. To model the simultaneous activation of multiple electrodes, the authors' design linearly adds individual-electrode phosphenes together. This produces relatively clean collections of dots that one could think of as pixels in a crude digital display. Modeling phosphenes in such a way assumes that each electrode and the network it activates operate independently of other electrodes and their neuronal targets. Unfortunately, as the authors acknowledge and as noted in the studies they used to fit and verify individual-electrode phosphene characteristics, simultaneous stimulation of multiple electrodes often obscures features of individual-electrode phosphenes and can produce unexpected phosphene patterns. This simulation does not reflect these nonlinearities in how electrode activations combine. Nonlinearities in electrode combinations can be as subtle the phosphenes becoming brighter while still remaining distinct, or as problematic as generating only a single small phosphene that is indistinguishable from the activation of a subset of the electrodes activated, or that of a single electrode.

If a visual prosthesis happens to generate some phosphenes that can be elicited independently, a simulator of this type could perhaps be used by processing stimulation from independent groups of electrodes and adding their phosphenes together in the visual field.

The reviewer raises a valid point and a similar point was raised by a different reviewer (our response is duplicated). As pointed out in the discussion, many aspects about multi- electrode phosphene perception are still unclear. On the one hand, the literature is in agreement that there is some degree of predictability: some papers explicitly state that phosphenes produced by multiple patterns are generally additive (Dobelle & Mladejovsky, 1974), that the locations are predictable (Bosking et al., 2018) and that multi-electrode stimulation can be used to generate complex, interpretable patterns of phosphenes (Chen et al., 2020, Fernandez et al., 2021). On the other hand, however, in some cases, the stimulation of multiple electrodes is reported to lead to brighter phosphenes (Fernandez et al., 2021), fused or displaced phosphenes (Schmidt et al., 1996, Bak et al., 1990) or unpredicted phosphene patterns (Fernández et al., 2021). It is likely that the probability of these interference patterns decreases when the distance between the stimulated electrodes increases. An empirical finding is that the critical distance for intracortical stimulation is approximately 1 mm (Ghose & Maunsell, 2012).

We note that our simulator is not restricted to the simulation of linearly combined Gaussian blobs. Some irregularities, such as elongated phosphene shapes were already supported in the previous version of our software. Furthermore, we added a supplementary figure that displays a possible approach to simulate some of the more complex electrode interactions that are reported in the literature, with only minor adaptations to the code. Our study thereby aims to present a flexible simulation toolkit that can be adapted to the needs of the user.

Adjustments:

• Lines 957-970: Furthermore, in contrast to the assumptions of our model, interactions between simultaneous stimulation of multiple electrodes can have an effect on the phosphene size and sometimes lead to unexpected percepts (Fernandez et al., 2021, Dobelle & Mladejovsky 1974, Bak et al., 1990). Although our software supports basic exploratory experimentation of non-linear interactions (see Figure 1-figure supplement 3), by default, our simulator assumes independence between electrodes. Multi- phosphene percepts are modeled using linear summation of the independent percepts. These assumptions seem to hold for intracortical electrodes separated by more than 1 mm (Ghose & Maunsell, 2012), but may underestimate the complexities observed when electrodes are nearer. Further clinical and theoretical modeling work could help to improve our understanding of these non-linear dynamics.

• Added Figure 1-figure supplement 3 on irregular phosphene percepts.

2) Verification of how the simulation renders individual phosphenes based on stimulation parameters is an important step in confirming agreement between the simulation and the function of implanted devices. That verification was well demonstrated. The end use a visual-prosthesis simulation, however, would likely not be optimizing just the appearance of phosphenes, but predicting and optimizing functional performance in visual tasks. Investigating whether this simulator can suggest visual-task performance, either with sighted volunteers or a decoder model, that is similar to published task performance from visual-prosthesis implantees would be a necessary step for true validation.

We agree with the reviewer that it will be vital to investigate the utility of the simulator in tasks. However, the literature on the performance of users of a cortical prosthesis in visually-guided tasks is scarce, making it difficult to compare task performance between simulated versus real prosthetic vision.

Secondly, the main objective of the current study is to propose a simulator that emulates the sensory / perceptual experience, i.e. the low-level perceptual correspondence. Once more behavioral data from prosthetic users become available, studies can use the simulator to make these comparisons.

Regarding the comparison to simulated prosthetic vision in sighted volunteers, there are some fundamental limitations. For instance, sighted subjects are exposed for a shorter duration to the (simulated) artificial percept and lack the experience and training that prosthesis users get. Furthermore, sighted subjects may be unfamiliar with compensation strategies that blind individuals have developed. It will therefore be important to conduct clinical experiments.

To convey more clearly that our experiments are performed to verify the practical usability in future behavioral experiments, we have incorporated the following textual adjustments:

• Lines 275-279: In the sections below, we discuss the different components of the simulator model, followed by a description of some showcase experiments that assess the ability to fit recent clinical data and the practical usability of our simulator in simulation experiments.

• Lines 842-853: Eventually, the functional quality of the artificial vision will not only depend on the correspondence between the visual environment and the phosphene encoding, but also on the implant recipient's ability to extract that information into a usable percept. The functional quality of end-to-end generated phosphene encodings in daily life tasks will need to be evaluated in future experiments. Regardless of the implementation, it will always be important to include human observers (both sighted experimental subjects and actual prosthetic implant users in the optimization cycle to ensure subjective interpretability for the end (Fauvel et al., 2022; Beyeler & Sanchez- Garcia, 2022).

3) A feature of this simulation is being able to convert stimulation of V1 to phosphenes in the visual field. If used, this feature would likely only be able to simulate a subset of phosphenes generated by a prosthesis. Much of V1 is buried within the calcarine sulcus, and electrode placement within the calcarine sulcus is not currently feasible. As a result, stimulation of visual cortex typically involves combinations of the limited portions of V1 that lie outside the sulcus and higher visual areas, such as V2.

We agree that some areas (most notably the calcarine sulcus) are difficult to access in a surgical implantation procedure. A realistic simulation of state-of-the-art cortical stimulation should only partially cover the visual field with phosphenes. However, it may be predicted that some of these challenges will be addressed by new technologies. We chose to make the simulator as generally applicable as possible and users of the simulator can decide which phosphene locations are simulated. To demonstrate that our simulator can be flexibly initialized to simulate specific implantation locations using third- party software, we have now added a supplementary figure (Figure 1-figure supplement 1) that displays a demonstration of an electrode grid placement on a 3D brain model, generating the phosphene locations from receptive field maps. However, the simulator is general and can also be used to guide future strategies that aim to e.g. cover the entire field with electrodes, compare performance between upper and lower hemifields etc.

Reviewer #3 (Public Review):

The authors are presenting a new simulation for artificial vision that incorporates many recent advances in our understanding of the neural response to electrical stimulation, specifically within the field of visual prosthetics. The authors succeed in integrating multiple results from other researchers on aspects of V1 response to electrical stimulation to create a system that more accurately models V1 activation in a visual prosthesis than other simulators. The authors then attempt to demonstrate the value of such a system by adding a decoding stage and using machine-learning techniques to optimize the system to various configurations.

1) While there is merit to being able to apply various constraints (such as maximum current levels) and have the system attempt to find a solution that maximizes recoverable information, the interpretability of such encodings to a hypothetical recipient of such a system is not addressed. The authors demonstrate that they are able to recapitulate various standard encodings through this automated mechanism, but the advantages to using it as opposed to mechanisms that directly detect and encode, e.g., edges, are insufficiently justified.

We thank the reviewer for this constructive remark. Our simulator is designed for more realistic assessment of different stimulation protocols in behavioral experiments or in computational optimization experiments. The presented end-to-end experiments are a demonstration of the practical usability of our simulator in computational experiments, building on a previously existing line of research. In fact, our simulator is compatible with any arbitrary encoding strategy.

As our paper is focused on the development of a novel tool for this existing line of research, we do not aim to make claims about the functional quality of end-to-end encoders compared to alternative encoding methods (such as edge detection). That said, we agree with the reviewer that it is useful to discuss the benefits of end-to-end optimization compared to e.g. edge detection will be useful.

We have incorporated several textual changes to give a more nuanced overview and to acknowledge that many benefits remain to be tested. Furthermore, we have restated our study aims more clearly in the discussion to clarify the distinction between the goals of the current paper and the various encoding strategies that remain to be tested.

• Lines 275-279: In the sections below, we discuss the different components of the simulator model, followed by a description of some showcase experiments that assess the ability to fit recent clinical data and the practical usability of our simulator in simulation experiments

• Lines 810-814: Computational optimization approaches can also aid in the development of safe stimulation protocols, because they allow a faster exploration of the large parameter space and enable task-driven optimization of image processing strategies (Granley et al., 2022; Fauvel et al., 2022; White et al., 2019; Küçükoglü et al. 2022; de Ruyter van Steveninck, Güçlü et al., 2022; Ghaffari et al., 2021).

• Lines 842-853: Eventually, the functional quality of the artificial vision will not only depend on the correspondence between the visual environment and the phosphene encoding, but also on the implant recipient's ability to extract that information into a usable percept. The functional quality of end-to-end generated phosphene encodings in daily life tasks will need to be evaluated in future experiments. Regardless of the implementation, it will always be important to include human observers (both sighted experimental subjects and actual prosthetic implant users in the optimization cycle to ensure subjective interpretability for the end user (Fauvel et al., 2022; Beyeler & Sanchez-Garcia, 2022).

2) The authors make a few mistakes in their interpretation of biological mechanisms, and the introduction lacks appropriate depth of review of existing literature, giving the reader the mistaken impression that this is simulator is the only attempt ever made at biologically plausible simulation, rather than merely the most recent refinement that builds on decades of work across the field.

We thank the reviewer for this insight. We have improved the coverage of the previous literature to give credit where credit is due, and to address the long history of simulated phosphene vision.

Textual changes:

• Lines 64-70: Although the aforementioned SPV literature has provided us with major fundamental insights, the perceptual realism of electrically generated phosphenes and some aspects of the biological plausibility of the simulations can be further improved and by integrating existing knowledge of phosphene vision and its underlying physiology.

• Lines 164-190: The aforementioned studies used varying degrees of simplification of phosphene vision in their simulations. For instance, many included equally-sized phosphenes that were uniformly distributed over the visual field (informally referred to as the ‘scoreboard model’). Furthermore, most studies assumed either full control over phosphene brightness or used binary levels of brightness (e.g. 'on' / 'off'), but did not provide a description of the associated electrical stimulation parameters. Several studies have explicitly made steps towards more realistic phosphene simulations, by taking into account cortical magnification or using visuotopic maps (Fehervari et al., 2010;, Li et al., 2013; Srivastava et al., 2009; Paraskevoudi et al., 2021), simulating noise and electrode dropout (Dagnelie et al., 2007), or using varying levels of brightness (Vergnieux et al., 2017; Sanchez-Garcia et al., 2022; Parikh et al., 2013). However, no phosphene simulations have modeled temporal dynamics or provided a description of the parameters used for electrical stimulation. Some recent studies developed descriptive models of the phosphene size or brightness as a function of the stimulation parameters (Winawer et al., 2016; Bosking et al., 2017). Another very recent study has developed a deep-learning based model for predicting a realistic phosphene percept for single stimulating electrodes (Granley et al., 2022). These studies have made important contributions to improve our understanding of the effects of different stimulation parameters. The present work builds on these previous insights to provide a full simulation model that can be used for the functional evaluation of cortical visual prosthetic systems.

• Lines 137-140: Due to the cortical magnification (the foveal information is represented by a relatively large surface area in the visual cortex as a result of variation of retinal RF size) the size of the phosphene increases with its eccentricity (Winawer & Parvizi, 2016, Bosking et al., 2017).

• Lines 883-893: Even after loss of vision, the brain integrates eye movements for the localization of visual stimuli (Reuschel et al., 2012), and in cortical prostheses the position of the artificially induced percept will shift along with eye movements (Brindley & Lewin, 1968, Schmidt et al., 1996). Therefore, in prostheses with a head-mounted camera, misalignment between the camera orientation and the pupillary axes can induce localization problems (Caspi et al., 2018; Paraskevoudi & Pezaris, 2019; Sabbah et al., 2014; Schmidt et al., 1996). Previous SPV studies have demonstrated that eye-tracking can be implemented to simulate the gaze-coupled perception of phosphenes (Cha et al., 1992; Sommerhalder et al., 2004; Dagnelie et al., 2006; McIntosh et al., 2013, Paraskevoudi & Pezaris, 2021; Rassia & Pezaris 2018, Titchener et al., 2018, Srivastava et al., 2009)

3) The authors have importantly not included gaze position compensation which adds more complexity than the authors suggest it would, and also means the simulator lacks a basic, fundamental feature that strongly limits its utility.

We agree with the reviewer that the inclusion of gaze position to simulate gaze-centered phosphene locations is an important requirement for a realistic simulation. We have made several textual adjustments to section M1 to improve the clarity of the explanation and we have added several references to address the simulation literature that took eye movements into account.

In addition, we included a link to some demonstration videos in which we illustrate that the simulator can be used for gaze-centered phosphene simulation. The simulation models the phosphene locations based on the gaze direction, and updates the input with changes in the gaze direction. The stimulation pattern is chosen to encode the visual environment at the location where the gaze is directed. Gaze contingent processing has been implemented in prior simulation studies (for instance: Paraskevoudi et al., 2021; Rassia et al., 2018; Titchener et al., 2018) and even in the clinical setting with users of the Argus II implant (Caspi et al., 2018). From a modeling perspective, it is relatively straightforward to simulate gaze-centered phosphene locations and gaze contingent image processing (our code will be made publicly available). At the same time, however, seen from a clinical and hardware engineering perspective, the implementation of eye-tracking in a prosthetic system for blind individuals might come with additional complexities. This is now acknowledged explicitly in the manuscript.

Textual adjustment:

Lines 883-910: Even after loss of vision, the brain integrates eye movements for the localization of visual stimuli (Reuschel et al., 2012), and in cortical prostheses the position of the artificially induced percept will shift along with eye movements (Brindley & Lewin, 1968, Schmidt et al., 1996). Therefore, in prostheses with a head-mounted camera, misalignment between the camera orientation and the pupillary axes can induce localization problems (Caspi et al., 2018; Paraskevoudi & Pezaris, 2019; Sabbah et al., 2014; Schmidt et al., 1996). Previous SPV studies have demonstrated that eye-tracking can be implemented to simulate the gaze-coupled perception of phosphenes (Cha et al., 1992; Sommerhalder et al., 2004; Dagnelie et al., 2006, McIntosh et al., 2013; Paraskevoudi et al., 2021; Rassia et al., 2018; Titchener et al., 2018; Srivastava et al., 2009). Note that some of the cited studies implemented a simulation condition where not only the simulated phosphene locations, but also the stimulation protocol depended on the gaze direction. More specifically, instead of representing the head-centered camera input, the stimulation pattern was chosen to encode the external environment at the location where the gaze was directed. While further research is required, there is some preliminary evidence that such a gaze-contingent image processing can improve the functional and subjective quality of prosthetic vision (Caspi et al., 2018; Paraskevoudi et al., 2021; Rassia et al., 2018; Titchener et al., 2018). Some example videos of gaze-contingent simulated prosthetic vision can be retrieved from our repository (https://github.com/neuralcodinglab/dynaphos/blob/main/examples/). Note that an eye-tracker will be required to produce gaze-contingent image processing in visual prostheses and there might be unforeseen complexities in the clinical implementation thereof. The study of oculomotor behavior in blind individuals (with or without a visual prosthesis) is still an ongoing line of research (Caspi et al.,2018; Kwon et al., 2013; Sabbah et al., 2014; Hafed et al., 2016).

4) Finally, the computational capacity required to run the described system is substantial and is not one that would plausibly be used as part of an actual device, suggesting that there may be difficulties with converting results from this simulator to an implantable system.

The software runs in real time with affordable, consumer-grade hardware. In Author response image 1 we present the results of performance testing with a 2016 model MSI GeForce GTX 1080 (priced around €600).

Author response image 1.

Note that the GPU is used only for the computation and rendering of the phosphene representations from given electrode stimulation patterns, which will never be part of any prosthetic device. The choice of encoder to generate the stimulation patterns will determine the required processing capacity that needs to be included in the prosthetic system, which is unrelated to the simulator’s requirements.

The following addition was made to the text:

• Lines 488-492: Notably, even on a consumer-grade GPU (e.g. a 2016 model GeForce GTX 1080) the simulator still reaches real-time processing speeds (>100 fps) for simulations with 1000 phosphenes at 256x256 resolution.

5) With all of that said, the results do represent an advance, and one that could have wider impact if the authors were to reduce the computational requirements, and add gaze correction.

We appreciate the kind compliment from the reviewer and sincerely hope that our revised manuscript meets their expectations. Their feedback has been critical to reshape and improve this work.

Author Response

Reviewer #2 (Public Review):

Wang et al. elegantly exploit single-cell RNA-seq datasets to question the putative involvement of lncRNAs in human germ cell development. In the first part of the study, the authors use computational approaches to identify and characterize, from existing data, lncRNAs expressed in the germline. Of note, the scRNA-seq data used were generated from polyA+ RNAs, and thus non-polyadenylated lncRNAs could not be retrieved. Most of the lncRNAs identified in the germ cells and in the somatic cells of the gonads were previously unannotated. While this increases the catalog of lncRNA genes in the human genome, further characterization is needed to determine which fraction of these newly identified lncRNAs represent bona fide transcripts or transcriptional noise.

Differential expression analysis between developmental stages, sexes, or cell types led to several observations: (i) whatever the stage of development, the number of expressed lncRNAs is higher in fetal germ cells compared to gonadal somatic cells; (ii) there is a continuous increase in the number of expressed lncRNA during the development of the germline; of note, a similar, although the more subtle trend is observed for protein-coding genes; (iii) the developmental stage at which there is the highest number of lncRNA expressed differs between male and female germ cells. While convincing, the significance of these observations is difficult to assess. However, the authors remain prudent with their conclusion and are not over-interpreting their findings.

We appreciate Reviewer #2 precise summary of our analysis and highlighting the significances of these datasets for other researchers and future studies.

Interestingly, integrating lncRNA expression to classify cell types led to the identification of a novel population of cells in the female germline that had not been revealed by protein-coding gene only-based classification. The biological relevance of this population, which cluster with mitotic populations, remains to be demonstrated. Finally, by examining lncRNA biotype, the authors could demonstrate an enrichment, in the germ cells, of the antisense head-to-head organization (in relation to the nearby protein-coding gene) compared to other biotypes. Whether this is different from the general distribution of lncRNA should be discussed.

We analyzed the lncRNAs in NONCODEv5 database (human genome), and the result showed that XH type occupied 21.73% of the intragenic lncRNA-mRNA pairs in NONCODEv5 database (human genome), which is lower than 26.58% in fGC and 26.23% in mGC (Response Figure 1).

Response Figure 1. Genomic distribution and biotypes of the lncRNAs in NONCODEv5 database and lncRNAs expressed in human gonad.

In the second part of the manuscript, Wang et al focus on one pair of divergent lncRNA-protein coding genes (LNC1845-LHX8). To document the choice of this particular pair, it would be informative to have its correlation score indicated in Figure 3C. he existence of this transcript was validated using female fetal ovaries, and its function was addressed in late primordial germ cells like cells (PGCLC) derived from human embryonic stem cells (hESCs). The authors have used an admirable set of orthogonal approaches that led them to conclude as to a role for LNC1845 in regulating in cis the nearby gene LHX8. They further went on to identify the underlying mechanisms, which involve modification of the chromatin landscape through direct interaction of LNC1845 with a histone modifier. Among the different strategies used (KO, stop transcription, overexpression), the shRNA-mediated knock-down is the only one to specifically address the function of the transcript itself, as opposed to the active transcription. The result of this experiment led the authors to conclude that the LNC1845 RNA is functional, a conclusion that is reinforced by the demonstration of physical interaction between the LNC1845 RNA and WDR5, a component of MLL methyltransferase complexes. The result of the KD experiment is however puzzling as RNAi has been shown not to be the method of choice for targeting nuclear lncRNAs (Lennox et al. NAR 2016).

We thank the Reviewer #2’s suggestion to add the correlation score of LNC1845-LHX8 pair and the Pearson Correlation of this pair is 0.3268. We have added the number to Figure 4C because which the expression correlation of LNC1845 and LHX8 was first mentioned. We have compared many other similar studies, shRNA knockdown has been widely used to target nuclear lncRNAs (Guttman et al. Nature 2011; Luo et al. Cell Stem Cell 2016; Subhash et al. Nucleic Acids Res. 2018; Li et al. Genome Res 2021), and the knockdown efficiency seemed to be feasible and acceptable to be used. The knockdown results are consistent with the deletion mutation and stop transcription approaches, all three showed that LNC1845 transcriptional expression is required for proper LHX8 expression in late PGCLCs.

Overall, the functional investigation is convincing and strengthened by the inclusion of multiple clones for each approach, and by the convergence in the outcome of each individual approach. The depth of characterization is also remarkable. The analyses of the mechanisms at stake are somehow less solid, as there is less evidence demonstrating the involvement of the LNC1845 RNA and its interaction with WDR5.

We have added more experimental evidence to strengthen the model especially the interaction of LNC1845 and WDR5. Apart from the RIP-qPCR results of WDR5 demonstrating the enrichment of LNC1845 by WDR5 pulldown (Figure S8D), we performed chromatin isolation by RNA purification (ChIRP) assay using antisense oligos along the entire LNC1845 transcript sequence. ChIRP results confirmed that WDR5 protein were enriched when anti-LNC1845 oligo probes were used to isolate the complex but not the controls without the probes or without overexpression of LNC1845 transcript (Response Figure 2). Taken together, the findings of both approaches support the model that LNC1845 directly interacts with WDR5 to modulate the H3K4me3 modification for LHX8 transcriptional activation. (Related to supplementary figure 8D and 8E.)

Response Figure 2. LNC1845 binding for WDR5 was verified by CHIRP-western blot.

Altogether, this study provides a convincing demonstration of the role of a lncRNA on the regulation of a nearby gene in the context of the germline. However, to have a better understanding of the functionality of lncRNA genes in general, it would be interesting to know whether other pairs of lncRNA-PC genes have been functionally investigated in this context, where no function for the lncRNA gene could be demonstrated. Negative results are highly informative and if so, these could be included in the manuscript.

We appreciate Reviewer #2 suggestion to add other lncRNA-PC gene pairs results. In fact, we have analyzed and presented the results of another 2 pairs in figure 7D. LncRNAs LNC3346 and LNC15266 were also transcriptionally regulated by FOXP3, and they may regulate their neighbor genes TMCO1 and MPP5, as figure 7D showed. Our analysis showed that other lncRNA-PC gene pairs may also have the similar transcriptional regulation as LNC1845-LHX8 during germ cell development.

Author Response

Reviewer #1 (Public Review):

The role of the parietal (PPC), the retrospenial (RSP) and the the visual cortex (S1) was assessed in three tasks corresponding a simple visual discrimination task, a working-memory task and a two-armed bandit task all based on the same sensory-motor requirements within a virtual reality framework. A differential involvement of these areas was reported in these tasks based on the effect of optogenetic manipulations. Photoinhibition of PPC and RSP was more detrimental than photoinhibition of S1 and more drastic effects were observed in presumably more complex tasks (i.e. working-memory and bandit task). If mice were trained with these more complex tasks prior to training in the simple discrimination task, then the same manipulations produced large deficits suggesting that switching from one task to the other was more challenging, resulting in the involvement of possibly larger neural circuits, especially at the cortical level. Calcium imaging also supported this view with differential signaling in these cortical areas depending on the task considered and the order to which they were presented to the animals. Overall the study is interesting and the fact that all tasks were assessed relying on the same sensory-motor requirements is a plus, but the theoretical foundations of the study seems a bit loose, opening the way to alternate ways of interpreting the data than "training history".

1) Theoretical framework:

The three tasks used by the authors should be better described at the theoretical level. While the simple task can indeed be considered a visual discrimination task, the other two tasks operationally correspond to a working-memory task (i.e. delay condition which is indeed typically assessed in a Y- or a T-maze in rodent) or a two-armed bandit task (i.e. the switching task), respectively. So these three tasks are qualitatively different, are therefore reliant on at least partially dissociable neural circuits and this should be clearly analyzed to explain the rationale of the focus on the three cortical regions of interest.

We are glad to see that the reviewer finds our study interesting overall and sees value in the experimental design. We agree that in the previous version, we did not provide enough motivation for the specific tasks we employed and the cortical areas studied.

Navigating to reward locations based on sensory cues is a behavior that is crucial for survival and amenable to a head-fixed laboratory setting in virtual reality for mice. In this context of goal-directed navigation based on sensory cues, we chose to center our study on posterior cortical association areas, PPC and RSC, for several reasons. RSC has been shown to be crucial for navigation across species, poised to enable the transformation between egocentric and allocentric reference frames and to support spatial memory across various timescales (Alexander & Nitz, 2015; Fischer et al., 2020; Pothuizen et al., 2009; Powell et al., 2017). It furthermore has been shown to be involved in cognitive processes beyond spatial navigation, such as temporal learning and value coding (Hattori et al., 2019; Todd et al., 2015), and is emerging as a crucial region for the flexible integration of sensory and internal signals (Stacho & ManahanVaughan, 2022). It thus is a prime candidate area in the study of how cognitive experience may affect cortical involvement in goal-directed navigation.

RSC is heavily interconnected with PPC, which is generally thought to convert sensory cues into actions (Freedman & Ibos, 2018) and has been shown to be important for navigation-based decision tasks (Harvey et al., 2012; Pinto et al., 2019). Specific task components involving short-term memory have been suggested to cause PPC to be necessary for a given task (Lyamzin & Benucci, 2019), so we chose such task components in our complex tasks to maximize the likelihood of large PPC involvement to compare the simple task to.

One such task component is a delay period between cue and the ultimate choice report, which is a common design in decision tasks (Goard et al., 2016; Harvey et al., 2012; Katz et al., 2016; Pinto et al., 2019). We agree with the reviewer that traditionally such a task would be referred to as a workingmemory task. However, we refrain from using this terminology because it may cause readers to expect that to solve the task, mice use a working-memory dependent strategy in its strictest and most traditional sense, that is mice show no overt behaviors indicative of the ultimate choice until the end of the delay period. If the ultimate choice is apparent earlier, mice may use what is sometimes referred to as an embodiment-based strategy, which by some readers may be seen as precluding working memory. Indeed, in new choice-decoding analyses from the mice’s running patterns, we show that mice start running towards the side of the ultimate choice during the cue period already (Figure 1—figure supplement 1). Regardless of these seemingly early choices, however, we crucially have found much larger performance decrements from inhibition in mice performing the delay task compared to mice performing the simple task, along with lower overall task performance in the delay task, indicating that the insertion of a delay period increased subjective task difficulty. As traditional working-memory versus embodiment-based strategies are not the focus of our study here and do not seem to inform the performance decrements from inhibition, we chose to label the task descriptively with the crucial task parameter rather than with the supposedly underlying cognitive process.

For the switching task, we appreciate that the reviewer sees similarities to a two-armed bandit task. However, in a two-armed bandit task, rewards are typically delivered probabilistically, whereas in our task, cue and action values are constant within each of the two rule blocks, and only the rule, i.e. the cuechoice association, reverses across blocks. This is a crucial distinction because in our design, blocks of Rule A in the switching task are identical to the simple task, with fixed cue-choice associations and guaranteed reward delivery if the correct choice is made, allowing a fair comparison of cortical involvement across tasks.

We have now heavily revised the introduction, results, and discussion sections of the manuscript to better explain the motivation for the tasks and the investigated brain areas. These revisions cover all the points mentioned in this response.

Furthermore, we agree with the reviewer that the three tasks are qualitatively different and likely depend on at least partially dissociable circuits. We consider the large differences in cortical inhibition effects between the simple and the complex tasks as evidence for this notion. We also want to highlight that in fact, we performed task-specific optogenetic manipulations presented in the Supplementary Material to further understand the involvement of different areas in task-specific processes. In what is now Figure 1—figure supplement 4, we restricted inhibition in the delay task to either the cue period only or delay period only, finding that interestingly, PPC or RSC inhibition during either period caused larger performance drops than observed in the simple task. We also performed epoch-specific inhibition of PPC in the switching task, targeting specifically reward and inter-trial-interval periods following rule switches, in what is now Figure 1—figure supplement 5. With such PPC inhibition during the ITI, we observed no effect on performance recovery after rule switches and thus found PPC activity to be dispensable for rule updates.

For the working-memory task we do not know the duration of the delay but this really is critical information; per definition, performance in such a task is delay-dependent, this is not explored in the paper.

We thank the reviewer for pointing out the lack of information on delay duration and have now added this to the Methods section.

We agree that in classical working memory tasks where the delay duration is purely defined by the experimenter and varied throughout a session, performance is typically dependent on delay duration. However, in our delay task, the delay distance is kept constant, and thus the delay is not varied by the experimenter. Instead, the time spent in the delay period is determined by the mouse, and the only source of variability in the time spent in the delay period is minor differences in the mice’s running speeds across trials or sessions. Notably, the differences in time in the delay period were greatest between mice because some mice ran faster than others. Within a mouse, the time spent in the delay period was generally rather consistent due to relatively constant running speeds. Also, because the mouse had full control over the delay duration, it could very well speed up its running if it started to forget the cue and run more slowly if it was confident in its memory. Thus, because the delay duration was set by the mouse and not the experimenter, it is very challenging or impossible to interpret the meaning and impact of variations in the delay duration. Accordingly, we had no a priori reason to expect a relationship between task performance and delay duration once mice have become experts at the delay task. Indeed, we do not see such a relationship in our data (see plot here, n = 85 sessions across 7 mice). In order to test the effect of delay duration on behavioral performance, we would have to systematically change the length of the delay period in the maze, which we did not do and which would require an entirely new set of experiments.

Also, the authors heavily rely on "decision-making" but I am genuinely wondering if this is at all needed to account for the behavior exhibited by mice in these tasks (it would be more accurate for the bandit task) as with the perspective developed by the authors, any task implies a "decision-making" component, so that alone is not very informative on the nature of the cognitive operations that mice must compute to solve the tasks. I think a more accurate terminology in line with the specific task considered should be employed to clarify this.

We acknowledge that the previous emphasis on decision-making may have created expectations that we demonstrate effects that are specific to the ‘decision-making’ aspect of a decision task. As we do not isolate the decision-making process specifically, we have substantially revised our wording around the tasks and removed the emphasis on decision-making, including in the title. Rather than decision-making, we now highlight the navigational aspect of the tasks employed.

The "switching"/bandit task is particularly interesting. But because the authors only consider trials with highest accuracy, I think they are missing a critical component of this task which is the balance between exploiting current knowledge and the necessity to explore alternate options when the former strategy is no longer effective. So trials with poor performance are thus providing an essential feedback which is a major drive to support exploratory actions and a critical asset of the bandit task. There is an ample literature documenting how these tasks assess the exploration/exploitation trade-off.

We completely agree with the reviewer that the periods following rule switches are an essential part of the switching task and of high interest. Indeed, ongoing work in the lab is carefully quantifying the mice’s strategy in this task and exploring how mice use errors after switches to update their belief about the rule. In this project, however, a detailed quantification of switching task strategy seemed beyond the scope because our focus was on training history and not on the specifics of each task. While we agree with the reviewer about the interesting nature of the switching period, it would be too much for a single paper to investigate the detailed mechanisms of each task on top of what we already report for training history. Instead, we have now added quantifications of performance recovery after rule switches in Figure 1— figure supplement 2, showing that rule switches cause below-chance performance initially, followed by recovery within tens of trials.

2) Training history vs learning sets vs behavioral flexibility:

The authors consider "training history" as the unique angle to interpret the data. Because the experimental setup is the same throughout all experiments, I am wondering if animals are just simply provided with a cognitive challenge assessing behavioral flexibility given that they must identify the new rule while restraining from responding using previously established strategies. According to this view, it may be expected for cortical lesions to be more detrimental because multiple cognitive processes are now at play.

It is also possible that animals form learning sets during successive learning episodes which may interfere with or facilitate subsequent learning. Little information is provided regarding learning dynamics in each task (e.g. trials to criterion depending on the number of tasks already presented) to have a clear view on that.

We thank the reviewer for raising these interesting ideas. We have now evaluated these ideas in the context of our experimental design and results. One of the main points to consider is that for mice transitioned from either of the complex tasks to the simple task, the simple task is not a novel task, but rather a well-known simplification of the previous tasks. Mice that are experts on the delay task have experienced the simple task, i.e. trials without a delay period, during their training procedure before being exposed to delay periods. Switching task expert mice know the simple task as one rule of the switching task and have performed according to this rule in each session prior to the task transition. Accordingly, upon to the transition to the simple task, both delay task expert mice and switching task expert mice perform at very high levels on the very first simple task session. We now quantify and report this in Figure 2—figure supplement 1 (A, B). This is crucial to keep in mind when assessing ‘learning sets’ or ‘behavioral flexibility’ as possible explanations for the persistent cortical involvement after the task transitions. In classical learning sets paradigms, animals are exposed to a series of novel associations, and the learning of previous associations speeds up the learning of subsequent ones (Caglayan et al., 2021; Eichenbaum et al., 1986; Harlow, 1949). This is a distinct paradigm from ours because the simple task does not contain novel associations that are new to the mice already trained on the complex tasks. Relatedly, the simple task is unlikely to present a challenge of behavioral flexibility to these mice given our experimental design and the observation of high simple task performance in the first session after the task transition.

We now clarify these points in the introduction, results, and discussion sections, also acknowledging that it will be of interest for future work to investigate how learning sets may affect cortical task involvement.

3) Calcium imaging data versus interventions:

The value of the calcium imaging data is not entirely clear. Does this approach bring a new point to consider to interpret or conclude on behavioral data or is it to be considered convergent with the optogenetic interventions? Very specific portions of behavioral data are considered for these analyses (e.g. only highly successful trials for the switching/bandit task) and one may wonder if considering larger or different samples would bring similar insights. The whole take on noise correlation is difficult to apprehend because of the same possible interpretation issue, does this really reflect training history, or that a new rule now must be implemented or something else? I don't really get how this correlative approach can help to address this issue.

We thank the reviewer for pointing out that the relationship between the inhibition dataset and calcium imaging dataset is not clear enough. We restricted analyses of inhibition and calcium imaging data in the switching task to the identical cue-choice associations as present in the simple task (i.e. Rule A trials of the switching task). We did this because we sought to make the fairest and most convincing comparison across tasks for both datasets. However, we can now see that not reporting results with trials from the other rule causes concerns that the reported differences across tasks may only hold for a specific subset of trials.

We have now added analyses of optogenetic inhibition effects and calcium imaging results considering Rule B trials. In Figure 1—figure supplement 2, we show that when considering only Rule B trials in the switching task, effects of RSC or PPC inhibition on task performance are still increased relative to the ones observed in mice trained on and performing the simple task. We also show that overall task performance is lower in Rule B trials of the switching task than in the simple task, mirroring the differences across tasks when considering Rule A trials only.

We extended the equivalent comparisons to the calcium imaging dataset, only considering Rule B trials of the switching task in Figure 4—figure supplement 3. With Rule B trials only, we still find larger mean activity and trial-type selectivity levels in RSC and PPC, but not in V1, compared to the simple task, as well as lower noise correlations. We thus find that our conclusions about area necessity and activity differences across tasks hold for Rule B trials and are not due to only considering a subset of the switching task data.

In Figure 4—figure supplement 4, we further leverage the inclusion of Rule B trials and present new analyses of different single-neuron selectivity categories across rules in the switching task, reporting a prevalence of mixed selectivity in our dataset.

Furthermore, to clarify the link between the optogenetic inhibition and the calcium imaging datasets, we have revised the motivation for the imaging dataset, as well as the presentation of its results and discussion. Investigating an area’s neural activity patterns is a crucial first step towards understanding how differential necessity of an area across tasks or experience can be explained mechanistically on a circuit level. We now elaborate on the fact that mechanistically, changes in an area’s necessity may or may not be accompanied by changes in activity within that area, as previous work in related experimental paradigms has reported differences in necessity in the absence of differences in activity (Chowdhury & DeAngelis, 2008; Liu & Pack, 2017). This phenomenon can be explained by differences in the readout of an area’s activity. We now make more explicit that in contrast to the scenario where only the readout changes, we find an intriguing correspondence between increased necessity (as seen in the inhibition experiments) and increased activity and selectivity levels (as seen in the imaging experiments) in cortical association areas depending on the current task and previous experience. Rather than attributing the increase in necessity solely to these observed changes in activity, we highlight that in the simple task condition already, cortical areas contain a high amount of task information, ruling out the idea that insufficient local information would cause the small performance deficits from inhibition. Our results thus suggest that differential necessity across tasks and experience may still require changes at the readout level despite changes in local activity. We view our imaging results as an exciting first step towards a mechanistic understanding of how cognitive experience affects cortical necessity, but we stress that future work will need to test directly the relationship between cortical necessity and various specific features of the neural code.

Reviewer #2 (Public Review):

The authors use a combination of optogenetics and calcium imaging to assess the contribution of cortical areas (posterior parietal cortex, retrosplenial cortex, S1/V1) on a visual-place discrimination task. Headfixed mice were trained on a simple version of the task where they were required to turn left or right depending on the visual cue that was present (e.g. X = go left; Y = go right). In a more complex version of the task the configurations were either switched during training or the stimuli were only presented at the beginning of the trial (delay).

The authors found that inhibiting the posterior parietal cortex and retrosplenial cortex affected performance, particularly on the complex tasks. However, previous training on the complex tasks resulted in more pronounced impairments on the simple task than when behaviourally naïve animals were trained/tested on a simple task. This suggests that the more complex tasks recruit these cortical areas to a greater degree, potentially due to increased attention required during the tasks. When animals then perform the simple version of the task their previous experience of the complex tasks is transferred to the simple task resulting in a different pattern of impairments compared to that found in behaviorally naïve animals.

The calcium imaging data showed a similar pattern of findings to the optogenetic study. There was overall increased activity in the switching tasks compared to the simple tasks consistent with the greater task demands. There was also greater trial-type selectivity in the switching task compared to the simple task. This increased trial-type selectivity in the switching tasks was subsequently carried forward to the simple task so that activity patterns were different when animals performed the simple task after experiencing the complex task compared to when they were trained on the simple task alone

Strengths:

The use of optogenetics and calcium-imaging enables the authors to look at the requirement of these brain structures both in terms of necessity for the task when disrupted as well as their contribution when intact.

The use of the same experimental set up and stimuli can provide a nice comparison across tasks and trials.

The study nicely shows that the contribution of cortical regions varies with task demands and that longerterm changes in neuronal responses c can transfer across tasks.

The study highlights the importance of considering previous experience and exposure when understanding behavioural data and the contribution of different regions.

The authors include a number of important controls that help with the interpretation of the findings.

We thank the reviewer for pointing out these strengths in our work and for finding our main conclusions supported.

Weaknesses:

There are some experimental details that need to be clarified to help with understanding the paper in terms of behavior and the areas under investigation.

The use of the same stimuli throughout is beneficial as it allows direct comparisons with animals experiencing the same visual cues. However, it does limit the extent to which you can extrapolate the findings. It is perhaps unsurprising to find that learning about specific visual cues affects subsequent learning and use of those specific cues. What would be interesting to know is how much of what is being shown is cue specific learning or whether it reflects something more general, for example schema learning which could be generalised to other learning situations. If animals were then trained on a different discrimination with different stimuli would this previous training modify behavior and neural activity in that instance. This would perhaps be more reflective of the types of typical laboratory experiments where you may find an impairment on a more complex task and then go on to rule out more simple discrimination impairments. However, this would typically be done with slightly different stimuli so you don't introduce transfer effects.

We agree with the reviewer that investigating the effects of schema learning on cortical task involvement is an exciting future direction and have now explicitly mentioned this in the Discussion section. As the reviewer points out, however, our study was not designed to test this idea specifically. Because investigating schema learning would require developing and implementing an entirely new set of behavioral task variants, we feel this is beyond the scope of the current work. As to the question of how generalized the effects of cognitive experience are, our data in the run-to-target task suggest that if task settings are sufficiently distinct, cortical involvement can be similarly low regardless of complex task experience (now Figure 3—figure supplement 1). This finding is in line with recent work from (Pinto et al., 2019), where cortical involvement appears to change rapidly depending on major differences in task demands. However, work in MT has shown that previous motion discrimination training using dots can alter MT involvement in motion discrimination of gratings (Liu & Pack, 2017), highlighting that cortical involvement need not be tightly linked to the sensory cue identity.

It is not clear whether length of training has been taken into account for the calcium imaging study given the slow development of neural representations when animals acquire spatial tasks.

We apologize that the training duration and the temporal relationship between task acquisition and calcium imaging was not documented for the calcium imaging dataset. Please see our detailed reply below the ‘recommendations for the authors’ from Reviewer 2 below.

The authors are presenting the study in terms of decision-making, however, it is unclear from the data as presented whether the findings specifically relate to decision making. I'm not sure the authors are demonstrating differential effects at specific decision points.

We understand that the previous emphasis on decision-making may have created expectations that we demonstrate effects that are specific to the ‘decision-making’ aspect of a decision task. As we do not isolate the decision-making process specifically, we have substantially revised our wording around the tasks and removed the emphasis on decision-making, including in the title. Rather than decision-making, we now highlight the navigational aspect of the tasks employed.

While we removed the emphasis on the decision-making process in our tasks, we found the reviewer’s suggestion to measure ‘decision points’ a useful additional behavioral characterization across tasks. So, we quantified how soon a mouse’s ultimate choice can be decoded from its running pattern as it progresses through the maze towards the Y-intersection. We now show these results in Figure 1—figure supplement 1. Interestingly, we found that in the delay task, choice decoding accuracy was already very high during the cue period before the onset of the delay. Nevertheless, we had shown that overall task performance and performance with inhibition were lower in the delay task compared to the simple task. Also, in segment-specific inhibition experiments, we had found that inhibition during only the delay period or only the cue period decreased task performance substantially more than in the simple task, thus finding an interesting absence of differential inhibition effects around decision points. Overall, how early a mouse made its ultimate decision did not appear predictive of the inhibition-induced task decrements, which we also directly quantify in Figure 1—figure supplement 1.

Author Response

Reviewer #1 (Public Review):

In this manuscript, Abdellatef et al. describe the reconstitution of axonemal bending using polymerized microtubules (MTs), purified outer-arm dyneins, and synthesized DNA origami. Specifically, the authors purified axonemal dyneins from Chlamydomonas flagella and combined the purified motors with MTs polymerized from purified brain tubulin. Using electron microscopy, the authors demonstrate that patches of dynein motors of the same orientation at both MT ends (i.e., with their tails bound to the same MT) result in pairs of MTs of parallel alignment, while groups of dynein motors of opposite orientation at both MT ends (i.e., with the tails of the dynein motors of both groups bound to different MTs) result in pairs of MTs with anti-parallel alignment. The authors then show that the dynein motors can slide MTs apart following photolysis of caged ATP, and using optical tweezers, demonstrate active force generation of up to ~30 pN. Finally, the authors show that pairs of anti-parallel MTs exhibit bidirectional motion on the scale of ~50-100 nm when both MTs are cross-linked using DNA origami. The findings should be of interest for the cytoskeletal cell and biophysics communities.

We thank the reviewer for these comments.

We might be misunderstanding this reviewer’s comment, but the complexes with both parallel and anti-parallel MTs had dynein molecules with their tails bound to two different MTs in most cases, as illustrated in Fig.2 – suppl.1. The two groups of dyneins produce opposing forces in a complex with parallel MTs, and majority of our complexes had parallel arrangement of the MTs. To clarify the point, we have modified the Abstract:

“Electron microscopy (EM) showed pairs of parallel MTs crossbridged by patches of regularly arranged dynein molecules bound in two different orientations depending on which of the MTs their tails bind to. The oppositely oriented dyneins are expected to produce opposing forces when the pair of MTs have the same polarity.”

Reviewer #2 (Public Review):

Motile cilia generate rhythmic beating or rotational motion to drive cells or produce extracellular fluid flow. Cilia is made of nine microtubule doublets forming a spoke-like structure and it is known that dynein motor proteins, which connects adjacent microtubule doublet, are the driving force of ciliary motion. However the molecular mechanism to generate motion is still unclear. The authors proved that a pair of microtubules stably linked by DNA-origami and driven by outer dynein arms (ODA) causes beating motion. They employed in vitro motility assay and negative stain TEM to characterize this complex. They demonstrated stable linking of microtubules and ODAs anchored on the both microtubules are essential for oscillatory motion and bending of the microtubules.

Strength

This is an interesting work, addressing an important question in the motile cilia community: what is the minimum system to generate a beating motion? It is an established fact that dynein power stroke on the microtubule doublet is the driving force of the beating motion. It was also known that the radial spoke and the central pair are essential for ciliary motion under the physiological condition, but cilia without radial spokes and the central pair can beat under some special conditions (Yagi and Kamiya, 2000). Therefore in the mechanistic point of view, they are not prerequisite. It is generally thought that fixed connection between adjacent microtubules by nexin converts sliding motion of dyneins to bending, but it was never experimentally investigated. Here the authors successfully enabled a simple system of nexin-like inter-microtubule linkage using DNA origami technique to generate oscillatory and beating motions. This enables an interesting system where ODAs form groups, anchored on two microtubules, orienting oppositely and therefore cause tag-of-war type force generation. The authors demonstrated this system under constraints by DNA origami generates oscillatory and beating motions.

The authors carefully coordinated the experiments to demonstrate oscillations using optical tweezers and sophisticated data analysis (Fourier analysis and a step-finding algorithm). They also proved, using negative stain EM, that this system contains two groups of ODAs forming arrays with opposite polarity on the parallel microtubules. The manuscript is carefully organized with impressive movies. Geometrical and motility analyses of individual ODAs used for statistics are provided in the supplementary source files. They appropriately cited similar past works from Kamiya and Shingyoji groups (they employed systems closer to the physiological axoneme to reproduce beating) and clarify the differences from this study.

We thank the reviewer for these comments.

Weakness

The authors claim this system mimics two pairs of doublets at the opposite sites from 9+2 cilia structure by having two groups of ODAs between two microtubules facing opposite directions within the pair. It is not exactly the case. In the real axoneme, ODA makes continuous array along the entire length of doublets, which means at any point there are ODAs facing opposite directions. In their system, opposite ODAs cannot exist at the same point (therefore the scheme of Dynein-MT complex of Fig.1B is slightly misleading).

Actually, opposite ODAs can exist at the same point in our system as well, and previous work using much higher concentration of dyneins (e.g, Oda et al., J. Cell biol., 2007) showed two continuous arrays of dynein molecules between a pair of microtubules. To observe the structures of individual dynein molecules we used low concentrations of dynein and searched for the areas where dynein could be observed without superposition, but there were some areas where opposite dyneins existed at the same point.

We realize that we did not clearly explain this issue, so we have revised the text accordingly.

In the 1st paragraph of Results: “In the dynein-MT complexes prepared with high concentrations of dynein, a pair of MTs in bundles are crossbridged by two continuous arrays of dynein, so that superposition of two rows of dynein molecules is observed in EM images (Haimo et al., 1979; Oda et al., 2007). On the other hand, when a low concentration of the dynein preparation (6.25–12.5 µg/ml (corresponding to ~3-6 nM outer-arm dynein)) was mixed with 20-25 µg/ml MTs (200-250 nM tubulin dimers), the MTs were only partially decorated with dynein, so that we were able to observe single layers of crossbridges without superposition in many regions.” Legend of Fig. 1(C): “Note that the geometry of dyneins in the dynein-MT complex shown in (B) mimics that of a combination of the dyneins on two opposite sides of the axoneme (cyan boxes), although the dynein arrays in (B) are not continuous.”

If they want to project their result to the ciliary beating model, more insight/explanation would be necessary. For example, arrays of dyneins at certain positions within the long array along one doublet are activated and generate force, while dyneins at different positions are activated on another doublet at the opposite site of the axoneme. This makes the distribution of dyneins and their orientations similar to the system described in this work. Such a localized activation, shown in physiological cilia by Ishikawa and Nicastro groups, may require other regulatory proteins.

We agree that the distributions of activated dyneins in 3D are extremely important in understanding ciliary beating, and that other regulatory proteins would be required to coordinate activation in different places in an axoneme. However, the main goal of this manuscript is to show the minimal components for oscillatory movements, and we feel that discussing the distributions of activated dyneins along the length of the MTs would be too complicated and beyond the scope of this study.

They attempted to reveal conformational change of ODAs induced by power stroke using negative stain EM images, which is less convincing compared to the past cryo-ET works (Ishikawa, Nicastro, Pigino groups) and negative stain EM of sea urchin outer dyneins (Hirose group), where the tail and head parts were clearly defined from the 3D map or 2D averages of two-dynein ODAs. Probably three heavy chains and associated proteins hinder detailed visualization of the tail structure. Because of this, Fig.2C is not clear enough to prove conformational change of ODA. This reviewer imagines refined subaverage (probably with larger datasets) is necessary.

As the reviewer suggests, one of the reasons for less clear averaged images compared to the past images of sea urchin ODA is the three-headed structure of Chlamydomonas ODA. Another and perhaps the bigger reason is the difficulty of obtaining clear images of dynein molecules bound between 2 MTs by negative stain EM: the stain accumulates between MTs that are ~25 nm in diameter and obscures the features of smaller structures. We used cryo-EM with uranyl acetate staining instead of negative staining for the images of sea urchin ODA-MT complexes we previously published (Ueno et al., 2008) in order to visualize dynein stalks. We agree with the reviewer that future work with larger datasets and by cryo-ET is necessary for revealing structural differences.

That having been said, we did not mean to prove structural changes, but rather intended to show that our observation suggests structural changes and thus this system is useful for analyzing structural changes in future. In the revised manuscript, we have extensively modified the parts of the paper discussing structural changes (Please see our response to the next comment).

It is not clear, from the inset of Fig.2 supplement3, how to define the end of the tail for the length measurement, which is the basis for the authors to claim conformational change (Line263-265). The appearance of the tail would be altered, seen from even slightly different view angles. Comparison with 2D projection from apo- and nucleotide-bound 3-headed ODA structures from EM databank will help.

We agree with the reviewer that difference in the viewing angle affects the apparent length of a dynein molecule, although the 2 MTs crossbridged by dyneins lie on the carbon membrane and thus the variation in the viewing angle is expected to be relatively small. To examine how much the apparent length is affected by the view angle, we calculated 2D-projected images of the cryo-ET structures of Chlamydomonas axoneme (emd_1696 and emd_1697; Movassagh et al., 2010) with different view angles, and measured the apparent length of the dynein molecule using the same method we used for our negative-stain images (Author response image 1). As shown in the plot, the effect of view angles on the apparent lengths is smaller than the difference between the two nucleotide states in the range of 40 degrees measured here. Thus, we think that the length difference shown in Fig.2-suppl.4 reflects a real structural difference between no-ATP and ATP states. In addition, it would be reasonable to think that distributions of the view angles in the negative stain images are similar for both absence and presence of ATP, again supporting the conclusion.

Nevertheless, since we agree with the reviewer that we cannot measure the precise length of the molecule using these 2D images, we have revised the corresponding parts of the manuscript, adding description about the effect of view angles on the measured length in the manuscript.

Author response image 1. Effects of viewing angles on apparent length. (A) and (B) 2D-projected images of cryo-electron tomograms of Chlamydomonas outer arm dynein in an axoneme (Movassagh et al., 2010) viewed from different angles. (C) apparent length of the dynein molecule measured in 2D-projected images.

In this manuscript, we discuss two structural changes: 1) a difference in the dynein length between no-nucleotide and +ATP states (Fig.2-suppl.4), and 2) possible structural differences in the arrangement of the dynein heads (Fig.2-suppl.3). Although we realize that extensive analysis using cryo-ET is necessary for revealing the second structural change, we attempted to compare the structures of oppositely oriented dyneins, hoping that it would lead to future research. In the revised manuscript, we have added 2D projection images of emd_1696 and emd_1697 in Fig.2-suppl.3, so that the readers can compare them with our negative stain images. We had an impression that some of our 2D images in the presence of ATP resembled the cryo-ET structure with ADP.Vi, whereas some others appeared to be closer to the no-nucleotide cryo-ET structure. We have also attempted to calculate cross-correlations, but difficulties in removing the effect of MTs sometimes overlapped with a part of dynein, adjusting the magnifications and contrast of different images prevented us from obtaining reliable results.

To address this and the previous comments, we have extensively modified the section titled ‘Structures of dynein in the dynein-MT-DNA-origami complex’.

In Fig.5B (where the oscillation occurs), the microtubule was once driven >150nm unidirectionally and went back to the original position, before oscillation starts. Is it always the case that relatively long unidirectional motion and return precede oscillation? In Fig.7B, where the authors claim no oscillation happened, only one unidirectional motion was shown. Did oscillation not happen after MT returned to the original position?

Long unidirectional movement of ~150 nm was sometimes observed, but not necessarily before the start of oscillation. For example, in Figure 5 – figure supplement 1A, oscillation started soon after the UV flash, and then unidirectional movement occurred.

With the dynein-MT complex in which dyneins are unidirectionally aligned (Fig.7B, Fig.7-suppl.2), the MTs kept moving and escaped from the trap or just stopped moving probably due to depletion of ATP, so we did not see a MT returning to the original position.

Line284-290: More characterization of bending motion will be necessary (and should be possible). How high frequency is it? Do they confirm that other systems (either without DNA-origami or without ODAs arraying oppositely) cannot generate repetitive beating?

The frequencies of the bending motions measured from the movies in Fig.8 and Fig.8-suppl.1 were 0.6 – 1 Hz, and the motions were rather irregular. Even if there were complexes bending at high frequencies, it would not have been possible to detect them due to the low time resolution of these fluorescence microscopy experiments (~0.1 s). Future studies at a higher time resolution will be necessary for further characterization of bending motions.

To observe bending motions, the dynein-MT complex should be fixed to the glass or a bead at one part of the complex while the other end is free in solution. With the dynein-MT-DNA-origami complexes, we looked for such complexes and found some showing bending motions as in Fig. 8. To answer the reviewer’s question asking if we saw repetitive bending in other systems, we checked the movies of the complexes without DNA-origami or without ODAs arraying oppositely but did not notice any repetitive bending motions. However, future studies using the system with a higher temporal resolution and perhaps with an improved method for attaching the complex would be necessary in these cases as well.

Author Response

Reviewer #1 (Public Review):

Building upon the previous evidence of activation of auditory cortex VIP interneurons in response to non-classical stimuli like reward and punishment, Szadai et al., extended the investigation to multiple cortical regions. Use of three-dimensional acousto-optical two-photon microscopy along with the 3D chessboard scanning method allowed high-speed signal acquisition from numerous VIP interneurons in a large brain volume. Additionally, activity of VIP interneurons in deep cortical regions was obtained using fiber photometry. With the help of these two imaging methods authors were able to extract and analyze the VIP cell signal from different cortical regions. Study of VIP interneuron activity during an auditory go-no-go task revealed that more than half of recorded cortical VIP interneurons were responding to both reward and punishment with high reliability. Fiber photometry data revealed similar observations; however, the temporal dynamics of reinforcement stimuli-related response in mPFC was slower than in the auditory cortex. The authors performed detailed analysis of individual cell activity dynamics, which revealed five categories of VIP cells based on their temporal profiles. Further, animals with higher performance on the discrimination task showed stronger VIP responses to 'go trials' possibly suggesting the role of VIP interneurons in discrimination learning. Authors found that reinforcement related response of VIP interneurons in visual cortex was not correlated with their sensory tuning, unveiling an interesting idea that VIP interneurons take part in both local as well as global processing. These observations bring attention to the possible involvement of VIP interneurons in reinforcement stimuli-associated global signaling that would regulate local connectivity and information processing leading to learning.

The state-of-the-art imaging technique allowed authors to succeed in imaging VIP interneurons from several cortical regions. Advanced analyses revealed the nuances, similarities and differences in the VIP activity trend in various regions. The conclusions about reinforcement stimuli related activity of VIP interneurons made by the authors are well supported by the results obtained, however some claims and interpretations require more attention and clarification.

We thank Reviewer #1 for the positive general comments.

Reviewer #2 (Public Review):

In recent years the activity of cortical VIP+ interneurons in relation to learning and sensory processing has raised great interest and has been intensely investigated. The ability of VIP+ interneurons in the auditory cortex to respond to both reward and punishment was already reported a few years ago by some of the authors (Pi et al., 2013, Nature). However, this work importantly adds to their previous study demonstrating a largely similar and synchronous response of a large fraction of these interneurons across the neocortex to salient stimuli of different valence during the performance of an auditory discrimination task.

An additional strength of this study is the analysis and identification of the general pattern of VIP+ interneuron responses associated to specific behaviors in the different layers of the neocortex depth.

Interestingly, the authors also identified using cluster analysis 5 different classes of VIP+ interneurons, based on the dynamic of their responses, that were unequally distributed in distinct cortical areas.

This is a well performed study that took advantage of a cutting-edge imaging approach with high recording speed and good signal-to-noise ratio. Experiments are well performed and the data are properly analyzed and nicely illustrated. However, one shortcoming of this paper, in my opinion, is the "case report" structure of the data. Essentially for each neocortical area the activity of VIP+ interneurons was analyzed only in one animal. This limits the assessment of the stability of the response/recruitment of these interneurons. I appreciate the high number of recorded VIP+ interneurons per area/animal and I do understand that it would be excessively laborious to perform 3D random-access two-photon microscopy in several mice for each cortical area. On the other hand, it would be important to have some knowledge of the general variability of the responses of these neurons among animals.

In conclusion, despite the findings described in this manuscript being generally sound, additional experiments are recommended to further substantiate the conclusions.

Thank you for pointing out this potential misunderstanding. Although we mentioned the number of animals the recordings were obtained from (n=22 total), we repeated this multiple times to alleviate the potential confusion. The data recorded with the 2-photon microscope are from 16 animals, and fiber photometry was performed on a separate 6 animals. Each animal was recorded in one (14 mice) or two areas (8 mice, 2 AOD, 6 photometry). We aimed to acquire data from at least 3 recordings per area (4 in the primary somatosensory cortex, 6 in the primary and secondary motor cortices, 4 in the lateral and medial parietal cortices, 3 in the primary visual cortices, 6 in the auditory and medial prefrontal cortices). In the revised manuscript this information can be found at the beginning of the results section and in the figure legends:

“To probe the behavioral function of VIP interneurons, we trained head-fixed mice (n=22 in total, n=16 for 2-photon microscopy and n=6 for fiber photometry) on a simple auditory discrimination task (Figure 1A).”

“Among the 811 neurons imaged in 18 imaging sessions from 16 mice,”

“Ca2+ responses of individual VIP interneurons recorded separately from 18 different cortical regions from 16 mice using fast 3D AO imaging were averaged for Hit (thick green), FA (thick red), Miss (dark blue), and CR (light blue). Fiber photometry data were recorded simultaneously from mPFC and ACx regions and are shown in gray boxes. Functional map (Kirkcaldie, 2012) used with the permission of the author. Speaker symbols represent the average time of tone onset, and gray triangles mark the reinforcement onset for Hit and FA. Averages of Miss and CR trials were aligned according to the expected reinforcement delivery calculated on the basis of the average reaction time. mPFC: medial prefrontal cortex (n=6 mice), ACx: auditory cortex (n=6), S1Hl/S1Tr/S1Bf/S1Sh: primary somatosensory cortex, hindlimb/trunk/barrel field/shoulder region (n=4), M1/M2: primary/secondary motor cortex (n=6), Mpta/Lpta: medial/lateral parietal cortex (n=4), V1: primary visual cortex (n=3).”

“This approach allowed us to simultaneously measure bulk calcium-dependent signals from VIP interneurons located in the right medial prefrontal (mPFC) and left auditory cortices (ACx) by implanting two 400 µm optical fibers at these locations (n=6 sessions from n=6 mice, Figure 1–figure supplement 1C).”

“Raster plot of the trial-to-trial activation of the responsive VIP neurons in Hit and FA trials during the two-photon imaging sessions (n=18 sessions, n=16 mice, n=746 cells).”

Subregional labels, for example on Figure 2, should be considered as additional information to orient the readers, even if they were very precisely defined on the basis of the coordinates. All analyses considering regional differences were conducted on the level of the main functional areas of the dorsal cortex (motor, somatosensory, parietal, and visual). Despite some location-dependent heterogeneity in the late response phase (Figures 2G and H), even these main dorsal cortical regions were all similar from the perspective of responsiveness to reinforcers and auditory cues.

Reviewer #3 (Public Review):

In this study Szadai et al. show reliable, relatively synchronous activation of VIP neurons across different areas of dorsal cortex in response to reward and punishment of mice performing an auditory discrimination task. The authors use both a relatively fast 2 photon imaging, as well as fiber photometry for some deeper areas. They cluster neurons according to their temporal response profiles and show that these profiles differ across areas and cortical depths. Task performance, running behavior and arousal are all related to VIP response magnitude, as has been previously shown.

Methodologically, this paper is strong: the described imaging technique allows for fairly fast sampling rates, they sample VIP cells from many different areas and the analyses are sophisticated and touch on the most relevant points. The figures are of high quality.

However, as the manuscript is now, the presentation could be clearer, the methods more complete and it is not clear whether their conclusions are entirely supported by the data.

The main issue is that reinforcement and arousal are hard to distinguish in this study. It is well known that VIP activity is correlated with arousal. And it is fairly clear that the reinforcement they use in this study - air puffs to the eye, as well as water rewards - cause arousal. It is possible that the reinforcer responses they observe in VIP neurons throughout all areas merely reflect the increases in arousal caused by these behaviorally salient events. They do discuss this caveat (albeit not fully convincingly) and in their abstract even state that the arousal state was not predictive of reinforcer responses. However their data clearly shows the tight relationship of the VIP reinforcer responses to both arousal (as measured by pupil diameter), as well as running speed of the animal. Both of these variables are well known to be tightly coupled to VIP activity.

Although barely mentioned, the authors do appear to sometimes present uncued reward (Figure S2F). If responses were noticeably different from the same events in the task context (as actual reinforcers) this could at least hint towards the reinforcement signal being distinct from mere arousal. However, this data is only mentioned in one supplementary figure in a different context (comparison with PV cells) and neither directly compared to cued reward, nor is this discussed at all. Were uncued air puffs also presented? How do the responses compare to cued air puffs/punishment?

Our original approach to distinguish between reinforcement- and arousal-related responses aimed:

1) to show that VIP cells with both low and high correlation coefficients with arousal produce large signals upon reinforcement presentation (Figure 3B),

2) the high differences of low and high arousal changes were reflected in a limited way in the VIP activity (Figures 3C and D): as highlighted in Figure R1, where we also added bars to show ∆P/P in high and low pupil change conditions, the difference in ∆P/P is ~5-fold, while it is only ~1.5-fold for ∆F/F. This disproportionality suggests that a large part of the signal below the dashed blue line is independent of arousal. We have added these modifications to the new version of Figure 3 for clarity.

Figure R1 = Figure 3C-D with modification. Comparison of pupil changes and corresponding calcium averages.

We collected further evidence to support our claims. In Figure 3–figure supplement 2 we depicted Hit and FA trials in which the reinforcement didn’t elevate the arousal level any further. Many of these trials were associated with locomotion prior to the reinforcement, but it was also common that the animals remained still during the whole trial. Trials with increased locomotion upon reinforcement presentation were excluded. Reinforcement-related calcium signals were still present under these conditions, indicating that these signals are not simple reflections of arousal. Moreover, we estimate the distinct contributions of arousal, locomotion, and reinforcers in Figure 3–figure supplement 2D in a systematic way with a generalized linear model. This model also confirmed our view about the reinforcement-related coding.

We now say in the results:

“Finally, to assess the motor- and reinforcement-related contributions to VIP interneuronal activity, we built a generalized linear model using the behavior and imaging data of the SS and Mtr recordings (Figure 3–figure supplement 2D, n=3 mice). This model was able to explain 18.8 ± 11.1% of the variance of the VIP population calcium signal, and highlighted that arousal was the best predictor, followed by reward, punishment, locomotion velocity, and auditory cue (weights = 0.055, 0.031, 0.028, 0.020, 0.018 respectively; all predictors, except the auditory cue in the case of one animal, contributed significantly, p<0.001). These observations indicate that running and arousal changes alone cannot fully explain the recruitment of VIP interneurons by reinforcers.”

We apologize for not describing the rational and the result from the uncued reward experiments. Briefly, while recording reinforcement related signals in auditory cortex in our task, we realized that the cue delivery, and the resulting purely sensory response could alter the measurement of the reward-related responses. Hence, in order to disentangle the reward and sensory-related responses, we presented the animals with simple, uncued reward and observed a similar and robust recruitment of VIP interneurons. Based on the same rational, we made similar measurement for PV neurons.

We now say in the results:

“We did not further analyze the FA responses in auditory cortex as those responses also had a sensory component linked to the white noise-like sound created by the air puff delivery. Because the cue delivery could prove as a confound to measure reward-mediated responses from VIP interneurons in auditory cortex (see also methods), we delivered random reward in separate sessions. Water droplets delivery recruited VIP interneurons in both auditory and medial prefrontal cortex in a similar fashion as water delivery during the discrimination task (Figure 2–figure supplement 1G). Like our single cell results, PV-expressing neuronal population in ACx did not show any significant change in activity upon similar random reward delivery (Figure 2–figure supplement 1G).”

Regarding the difference between cued and uncued responses, we definitely agree with the reviewer that it is an important point. The goal of this manuscript is however to study how reward and punishment are being represented by VIP interneurons in cortex.

The imaging method appears well suited for their task, however the improvements listed in table S1 make the method appear far superior to existing methods in many aspects. Published or preprinted papers with 2 photon imaging of VIP populations (eg. from Scanziani lab (Keller et al.), Carandini lab (Dipoppa et al.), deVries lab (Millman et al.), Adesnik lab (Mossing et al.), which use the much more common resonant scanning, seem to be able to image 4-7 layers at 4-8Hz with a good enough SNR and potentially bigger neuronal yield of approximately 100-200 VIP cells, depending on the field of view. While not every single cell in a volume would be captured by these studies, the only main advantage of the here-used technique appears to be the superior temporal resolution.

We thank the reviewer for the positive comment and we agree that interpretation must be improved. We agree that the imaging methods in the papers listed above have good SNR and were proper to address the scientific questions that had arisen. As the reviewer points out, 3D-AOD imaging allows fast 3D measurement that cannot be achieved otherwise. We used these advantages to address the critical question of layer specificity in the response of VIP interneurons to reinforcer presentation (Figure 2–figure supplement 1F, but see also the new Figure 1–figure supplement 1B). Regarding the comparison and quantification of the factual advantages of AOD microscopy over other imaging methods, the reviewer and readers can refer to the methods section (3D AO microscopy), Table S1 and Szalay et al., 2016. We agree with the reviewer that one of the main advantages is the superior temporal resolution. The second main advantage is the improved SNR. This originates from the fact that the entire measurement time is spent on regions of interest; measurement of unnecessary background areas is not required. More specifically, SNR is improved even in the case of 2D imaging by the factor of:

((area of the entire frame )/(area of the recorded VIP cells))^0.5

which is about (100)0.5=10 as VIP interneurons represent about 1% of the brain. We used this second advantage of AO scanning when we determined the activation ratio (e.g., see Figure 2D).

As the resolution of single or a few action potentials is challenging in behaving mice labelled with the GCaMP6 sensor, any improvement in SNR will improve the detection threshold. The higher SNR achieved here improved the detection threshold, which also explains the relatively high activation ratio in our work.

In the case of asynchronous activity patterns, there is negligible contribution of individual small neuropil structures to somatic activities because of the relatively high volume-ratio of a soma and a given small neuropil structure: this minimizes the error during ∆F/F calculation of somatic responses. However, reinforcement, arousal, and running can generate highly synchronous neuronal activities which can synchronize neuropil activity around a given soma and, therefore, effectively and systematically modulating the somatic ∆F/F responses. To avoid this error, we used a high NA objective with proper neuropil resolution and combined it with motion correction. The use of the high NA also decreased the total scanning volume to about 689 µm × 639 µm × 580 µm and, therefore, it limited the maximum number of VIP cells which could be recorded. It is also possible to use a low-NA objective with a much higher FOV and scanning volume and record over 1000 VIP cells, but the extension of the PSF along the z dimension is inversely and quadratically proportional to the NA of the objective, therefore neuropil resolution will be at least partially lost. In summary, using the high-NA Olympus objective we maximized the 2P resolution which, in combination with off-line motion artifact elimination, allowed precise recording of somatic signals without any neuropil contamination: this provided correct activation ratio values.

Even though this is not mentioned at all, it certainly appears possible, that the accousto-optical scanning emits audible noise. In this case it would be good to know the frequency range and level of this background noise, whether there are auditory responses to the scanning itself and if it interferes with the performance of the animals in the auditory task in any way. If this is not the case, this should probably simply be mentioned for non-experts.

While the name of the acousto-optical deflectors seems to refer to “acoustic noise”, these devices are driven in the range of 55-120 MHz, which is 3 orders of magnitude higher frequency than the hearing threshold of animals: mice don’t hear them. Moreover, we developed water-cooled AODs ten years ago which means that ventilators are also not required, therefore AOD-based scanning can be used with zero noise emission. In contrast, galvo, resonant, and piezo scanning work in the kHz frequency range, which is in the middle of the hearing range of mice. Moreover, these technologies can’t be used in a vacuum and the scanner is just a few tens of centimeters away from the mice, which means that acoustic noise can’t be canceled but can only be partially suppressed with white noise. We thank the reviewer for the helpful comment and have added one sentence about the absence of acoustic noise during acousto-optical scanning:

“The deflectors are driven in the 55-120 MHz frequency range, therefore the noise emitted does not interfere with the auditory cues, as mice can’t hear it. This, in combination with the water cooling of the deflectors, makes the AOD-based scanning the quietest technology for in-vivo imaging.”

The authors show a strong correlation between task performance (hit rate) and the response to the auditory cue on hit trials. Was there any other significant correlations of VIP cells' responses to other trial types? Was reinforcer response correlated to behavioral variables at all?

We have not found any remarkable correlations between VIP cell activity and behavioral variables except the one mentioned above.

For example, we tested discrimination rate (hit rate/FA rate) correlation with ∆F/Ftone in Hit trials, but this was not significant (R2=0.03, F=0.49, p=0.69), just like Hit rate vs. ∆F/Ftone in FA trials (R2=0.19, F=3.8, p=0.07), and discrimination rate vs. ∆F/Ftone in FA trials (R2=0.07, F=1.1, p=0.31).

Author Response:

Reviewer #1 (Public Review):

This manuscript describes the role of PMd cck neurons in the invigoration of escape behavior (ie retreat from aversive stimuli located in a circumscribed area of the environment in which testing was conducted). Further, PMd cck neurons are shown to exert their effect on escape via the dorsal PAG. Finally, in an intriguing twist, aversive images are shown to increase the functional coupling between hypothalamus and PAG in the human brain.

The manuscript is broadly interdisciplinary, spanning multiple subfields of neuroscience research from slice physiology to human brain imaging.

We thank the Reviewer for recognizing the interdisciplinarity of our work.

To understand the novelty of the results obtained in the rodent studies, it is important to note that these data are a replication and elaboration of work published recently in Neuron by the primary authors of this manuscript. The current manuscript does not cite the Neuron paper.

We apologize for this omission. At the time of the current submission the Neuron paper had not been accepted and thus we could not cite it. We now discuss this paper in the introduction and highlight how the current manuscripts expand upon the data published in the Neuron paper.

The most novel aspect of the rodent experiments presented in this manuscript is the demonstration of a role for cck PMd neurons in invigorating behavioral withdrawal from cues associated with the kind of artificial stimuli commonly used in laboratory settings (ie a grid floor associated with shock). Unfortunately, these results are made somewhat difficult to interpret by a lack of counterbalancing - all subjects receive an assay of escape from a predator prior to the shock floor assay. Certainly, research on stress and sensitization tells us that prior experience with aversive stimuli can influence the response to aversive stimuli encountered in the future. Because the role of this pMD circuitry in predatory escape has already been demonstrated, this counterbalancing issues does somewhat diminish the impact of the most novel rodent data presented here.

Indeed, as the Reviewer states, prior exposure to aversive stimuli may influence responses to future exposures to threats. We opted to have the rat test before the shock grid test because the rat exposure is a milder experience than the shock grid test, as no actual pain occurs in the rat assay. We thus reasoned that the more intensely aversive assay (the shock assay) was more likely to influence behavior in the rat assay than vice-versa. Nevertheless, we agree with the Reviewer’s point that the lack of counterbalancing between the assays may mask potential influences of the rat assay on the shock grid assay behavior.

To address this issue we ran a cohort of new mice, showing that behavior in the shock grid assay is not affected by prior experience in the rat assay. We now show in Figure R1 and Figure 1, figure supplement 2 that freezing, threat avoidance and escape metrics in the shock grid assay are not significantly changed by prior exposure to the rat assay.

Figure R1. (Same as Figure 1, figure supplement 2). The order of threat exposure does not affect defensive behavior metrics. (A) Two cohorts of mice were exposed to the rat and shock grid threats in counterbalanced order, as specified in the yellow and green boxes. (B) The defensive behavioral metrics of these two cohorts were compared for the fear retrieval assay. None of the tested metrics were different between groups (Wilcoxon rank-sum test; each group, n=9 mice).

The manuscript concludes with an fMRI experiment in which the BOLD response to aversive images is reported to covary across the hypothalamus and PAG. It is intriguing that unpleasant pictures influence BOLD in regions that might be expected to contain circuits homologous to those demonstrated in rodents. It is important to note that viewing images is passive for the subjects of this experiment, and the data include no behavioral analogue of the escape responses that are the focus of the rest of the manuscript.

We agree with the Reviewer that there are many differences between the mouse and human behavioral tasks, and we have expanded the text highlighting these differences more clearly. One of our results, as highlighted by the Reviewer, is that inhibition of the PMd-dlPAG projection impairs escape from threats. Indeed, there is no escape in the human data, as stated by the Reviewer.

Now, we conducted new dual photometry recordings, in which we simultaneously monitor calcium transients in the PMd and the dlPAG in contralateral sides. Using these dual recordings, we show that mutual information between the PMd and the dlPAG in mice is higher during exposure to threats (rat and shock grid fear retrieval) than control assays (toy rat and pre-shock habituation) (Figure R2 and Figure 9 and Figure 9, figure supplement 1). Importantly, this analysis was also performed after excluding all time points that include escapes. Thus, the increase in PMd-dlPAG mutual information is independent of escapes, and is related to exposure to threats.

Similarly, the increase in activity in the human fMRI data in the hypothalamus-dlPAG pathway is also related to the exposure to aversive images, rather than specific defensive behaviors performed by the human subjects. This new finding of increased mutual information in the PMd-dlPAG circuit independently of escapes provides a better parallel to the human data.

In Figure R2 below we used mutual information instead of correlation because mutual information can capture both linear and non-linear correlation between two time-series. Figure R2E-G shows that the projection from PMd-cck cells to dlPAG is unilateral. Thus, in dual photometry recordings that were done contralaterally in the PMd and the dlPAG, the signals from the dlPAG are from local cell bodies, and are not contaminated by GCaMP signals from PMd-cck axon terminals.

Figure R2. (Panels from Figure 9(A-D) and from Figure 9, figure supplement 1 (panels E-G)) Dual fiber photometry signals from the PMd and dlPAG exhibit increased correlation and mutual information during threat exposure. (A) Scheme showing setup used to obtain dual fiber photometry recordings. (B) PMd-cck mice were injected with AAV9-Ef1a-DIO-GCaMP6s in the PMd and AAV9-syn-GCaMP6s in the dlPAG. (C) Expression of GCaMP6s in the PMd and dlPAG. (Scale bars: (left) 200 µm, (right) 150 µm) (D) Bars show the mutual information between the dual-recorded PMd and dlPAG signals, both including (left) and excluding (right) escape epochs, during exposure to threat and control. Mutual information is an information theory-derived metric reflecting the amount of information obtained for one variable by observing another variable. See Methods section for more details. (E) Cck-cre mice were injected with AAV9-Ef1a-DIO-YFP in the PMd in the left side. (F) Image shows the expression of YFP in PMd-cck cells in the left side. (scale bar: 200 µm) (G) PMd-cck axon terminals unilaterally express YFP in the dlPAG. (scale bar: 150µm). * p<0.05, ** p<0.01.

Reviewer #2 (Public Review):

The manuscript by Wang et al. addresses neuronal mechanisms underlying conserved escape behaviors. The study targets the midbrain periaqueductal grey, specifically the dorsolateral aspect (dlPAG), since previous research demonstrated that activation of dlPAG leads to escape behaviors in rodents and panic-related symptoms in humans. The hypothalamic dorsal premammillary nucleus (PMd) monosynaptically projects to the dlPAG and thus could play a role in escape behavior. The authors test whether cholecystokinin (CCK)-expressing PMd cells could be involved in escape behaviors from innate and conditioned threats using mainly two behavioral paradigms in mice: exposure to a live rat and electrical foot shocks.

Different approaches are used to test the main hypothesis. Using fiber photometry and microendoscopy calcium imaging in freely moving mice, the study finds that PMd CCK+ neurons were more active when mice are close to threats and during escape behaviors. Furthermore, PMD CCK+ activation patterns predicted escape behavior in a general linearized model. Chemogenetic inhibition of CCK+ PMd cells decreased escape speed from threats in both behavioral paradigms, while optogenetic activation of those cells lead to an increase in speed. Observation of c-fos expression after optogenetic activation revealed activation within two target areas of the PMd, the dlPAG and anteromedial ventral thalamus (AMv), in which cellular activity measured by fiber photometry also increased during escape behaviors. Interestingly, inhibition of PMd-to-dlPAG pathway, but not PMd-to-AMv, caused a decrease in escape velocity. Lastly, the authors investigated the response of several human participants to threatening images in an fMRI scan. These results suggest that similar to mice, an activation proportional to the threat intensity within a functional connection between hypothalamus and PAG pathway may occur in humans.

The authors conclude that a pathway from the PMd to the dlPAG, characterized by expression of CCK, control escape vigor and responsiveness to threat in mice, and that a similar pathway could be present in humans.

Overall, the comprehensive data from multiple approaches support a role of the identified pathway in escape behavior. However, an insufficient description of the used methods and experimental details makes it difficult to assess the validity and conclusivity of some findings. In addition, the strong interpretation emphasis on the functional specificity of the CCK+ PMd-dlPAG pathway appears not fully supported by the data.

1) The rationale for selection of CCK+ cells of the PMd is missing in the current manuscript. Despite methodological considerations, a clear description of these cells' role and characteristics from the existing literature is needed.

To address this point, we justify our choice of cck+ cells by discussing prior data showing that PMd cck cells are the major neuronal population of the PMd. Furthermore, cck is not strongly expressed in other adjacent hypothalamic nuclei, showing the high anatomical specificity of our manipulations targeting PMd-cck cells. We also discuss prior data (Wang et al., 2021) in the Introduction and Discussion about these cells.

2) The narrowness of the conclusions of the article is unnecessary. Although CCK+ PMd cells could play a role in regulating escape vigor, some of the presented results rather support the notion of a more general role of these cells in mediating defensive states. For example, the photometry data shows correlation of activity with other active defensive behavior. To address this point, a better analysis of the relation between neuronal activity and the general locomotor behavior of the animals is lacking. In addition, the already presented relation with the measured behaviors is not taken into account when interpreting the results (e.g. Fig 7 E). This description would be relevant to more comprehensively attributing functional roles for CCK + PMd cells.

At the Reviewer's request, we have included an analysis of the relationship between general locomotor behavior and PMd-cck df/F (Figure R3 and Figure 2, figure supplement 2). Interestingly, we found that the df/F increases monotonically with increasing ranges of speed and acceleration in the threat assays, while remaining fairly constant for matched ranges in the control assays.

We agree with the Reviewer that Figure 7E shows PMd-cck cells are activated not only during escape, but also other behaviors. However, the chemogenetic inhibition data show that PMd-cck cell activity only impaired escape speed, without altering freezing, approach or stretch-attend postures. Thus, the chemogenetic inhibition data indicates that the activity of these cells is only critical for escape, among the behaviors scored. Nevertheless, we discussed a “notion of a more general role of these cells in mediating defensive states” as suggested by Reviewer 2. However, Reviewer 1 provided the opposite feedback, stating that “It needs to be made clear that a specific role of PMd in quantitative measures of escape is the new result, instead of a broader role for this region”. Considering these opposing suggestions, we broadened the discussion on the role of the PMd, but did so conservatively.

Figure R3. (Same as Figure 2, figure supplement 2). Bars show the mean PMd-cck df/F (z-scored) for increasing ranges of (A) speed and (B) acceleration. (Wilcoxon signed-rank test; n=15) * p<0.05, ** p<0.01, *** p<0.001.

3) The imprecision of the methods description, especially the behavioral analysis is contributing to the previous point. In particular, the escape criterion itself seems to include a vague classification based on movement away from the threat- this should be more concretely defined (e.g. using angle of escape direction). In any case, the different behavioral context dimensions between the two paradigms would probably affect the escape criterion itself and thus have to be taken into account when interpreting the results.

The Reviewer makes an important point that the escape definition included in the Methods section was lacking in detail, specifying only a minimum directional speed. We had neglected to include two crucial criteria that were used as well: a minimum distance-from-threat at which escape must be initiated and a minimum distance traversed during escape. All escapes were therefore required to begin near the threat and lead to a substantial increase in mouse distance from the threat. These details are now included in the Methods section, as follows:

“'Escapes' were defined as epochs for which (1) the mouse speed away from the threat or control stimuli exceeded 2 cm/s for a minimum of 5 seconds continuously, (2) movement away from the threat was initiated at a maximum distance-from-threat of 30 cm and (3) the distance traversed from escape onset to offset was greater than 10 cm. Thus, escapes were required to begin near the threat and lead to a quick and substantial increase in distance from the threat.

'Escape duration' was defined as the amount of time that elapsed from escape onset to escape offset.

'Escape speed' was defined as the average speed from escape onset to offset.

'Escape angle' was defined as the cosine of the mouse head direction in radians, such that the values ranged from -1 (facing towards the threat) to 1 (facing away from the threat). Mouse head direction was determined by the angle of the line connecting a point midway between the ears and the nose.”

Using the escape definition above, a higher number of escapes and a higher average escape speed was observed in threat assays compared to control assays (Figure 1). This finding indicates that the definitions we used are capturing defensive evasion.

Both contexts have a length of 70 cm, so differences in the length of the contexts did not influence the definition of escape across contexts.

In response to the Reviewer's suggestion of an escape angle criterion, we have included Figure R4 which illustrates that, using the aforementioned escape definition, the resulting escape angle is quite stereotyped. The cosine of the escape angle shows very little variation, showing that only a narrow range of escape angles is used. Given this result, we opted to not include the angle of escape as part of the escape criteria to increase simplicity.

Figure R4. (A) Lines represent mouse position for all escapes that occurred during an example rat (top) and fear retrieval (bottom) session. Note that, while there is a diversity of escape routes, the escape angle is quite similar. (B) (left) Diagram provides a description of the escape angle metric, here calculated as the cosine of the head direction in radians. A value of 1 indicates an escape parallel with the long walls of the enclosure. (right) Bars represent the mean escape angle for all animals in Figure 1 during the rat and fear retrieval assays (n=32). As is apparent in (A), the mean escape angle cosine has little variability.

4) In line, more detailed descriptions of the animal's behavior are needed to support assessment of the results regarding the event-related fiber photometry results. Measures like frequency of escape, duration of freezing bouts and angle, duration and total speed of the escape bouts, and a better description of measures like Δ escape speed could be relevant for interpreting the results. In addition, there is no explanation of how the possible overlapping of behaviors in the broad time frame used in the experiments was regarded.

We have now included the requested measures as a supplement to Figure 2 (see also Fig. R5 below). Regarding overlapping behaviors, we have quantified the overlap between categorized behaviors in the fiber photometry assay and found that only a small fraction of behavioral timepoints were categorized as more than one behavior, primarily during behavioral transitions. This is quantified in Figure R6 below. Moreover, as is now described in the Methods, the analyses presented in Figure 2G-I (as well as Figure 7C-E, 7G-I) were performed only on behaviors that were separated from all other behaviors by a minimum of 5 seconds.

Figure R5. (Same as Figure 2, figure supplement 1) Behavioral metrics for the PMd fiber photometry cohort during threat exposure assays. (A) Diagram provides a description of the escape angle metric, here calculated as the cosine of the head direction in radians. A value of 1 indicates an escape parallel with the long walls of the enclosure. (B) Table shows pertinent defensive metrics during exposure to rat and fear retrieval assays for the PMd fiber photometry cohort. (n=15 mice).

Figure R6. The behavioral overlap between classified behaviors is minimal. The colormap depicts the fraction of behavioral timepoints for each of the four classified behaviors that was categorized as each of the remaining behaviors across all PMd fiber photometry assays (n=15 mice).

5) Part of the experimental results provide suboptimal evidence for the provided interpretation. That is, the lack of clear quantification and statistical analysis of the microendoscopy calcium imaging data on PMd-CCK+ cells makes it hard to reconcile this data with the photometry data. Furthermore, evidence through c-Fos staining after optogenetically stimulation of PMd-cck+ cells is insufficient evidence for the interpretation of broad, but functionally specific, recruitment of defensive networks. While the data on optogenetic inhibition of the PMd-CCK+ projection to the dlPAG seems to confirm the main hypothesis, both an intra-animal control and demonstration of statistical significance in the analysis are desirable to fully support that role.

We agree with the Reviewer that clear quantification and statistical analyses are essential in interpreting the microendoscopic analysis. However, we are not sure what is being requested, as we have applied both of these approaches to this dataset. For instance, in Figure 3, we quantify the percentage of cells that significantly encode each behavior as well as implement 5-fold logistic regression to determine how well these behaviors can be predicted. This accuracy is statistically compared to chance. Further quantification and statistical comparisons of speed and position decoding accuracy between threat and control assays are included in Figure 4. Concerning the Arch experiments, we have included an intra-animal control by comparing light off and on epochs, and we statistically compare the difference between these epochs with a control group.

Regarding the c-Fos experiment, we observe increased cfos expression in several nuclei known to be critical for defense, such as the bed nucleus of the stria terminalis and the ventromedial hypothalamus. This finding underlies our claim that optogenetic activation of the PMd recruits defensive networks. Nevertheless, it is entirely possible that naturalistic endogenous activation of the PMd does not recruit these nuclei. We added text addressing this caveat.

6) The provided fMRI data only provides circumstantial evidence to support a functionally specific hypothalamus to PAG pathway especially due to the technical characteristics and limitations of the experimental setup and behavioral paradigm.

The Reviewer makes an excellent point. Please see our response to Reviewer 1, point 6, where we provide a better parallel to the fMRI data in a new photometry analysis, as well as the added Figure 9.

Briefly, we now have conducted contralateral dual photometry recordings of the dlPAG and the PMd, and show an increase in mutual information between the neural activity of these two regions during exposure to threats. This result was found after removing all timepoints with escapes. Thus, the increase in mutual information is related broadly to threat exposure, rather than caused by specific moments during which escape occurs. We argue that this result more closely parallels the human data, as both the fMRI and mutual information from mice data show an increase in functional connectivity in the hypothalamus-dlPAG pathway during threat exposure, independently of escapes.

Reviewer #3 (Public Review):

This manuscript by Wang et al extends the Adhikari lab's earlier findings of the hypothalamic dorsal premammillary nucleus' role in defensive behavior. Using cell-type specific calcium imaging, the authors show that the activity of CCK-expressing PMd neurons precedes and predicts escape from both learned and unlearned threats. Optogenetic/chemogenetic inhibition revealed that the PMd-dlPAG pathway contributes to escape vigor. Additionally, optogenetic activation of CCK PMd neurons induces Fos in numerous brain regions implicated in fear and escape behaviors. Last, an analogous hypothalamic-PAG pathway in humans is shown to be activated by aversive images in humans.

Although these findings are potentially impactful, additional clarification and data are needed to strengthen and streamline the manuscript, as outlined below.

1) The results of the authors' recent publications (Wang et al Neuron 2021, Reis et al J. Neuro 2021) should be integrated into the manuscript. For example, the rationale for selectively manipulating CCK+ PMd neurons is not stated. Likewise, histological validation that the Cre-dependent GCaMP expression is restricted to CCK+ neurons should be shown or referenced. The authors should also provide discussion as to how the current results integrate with their other recent findings.

Following the Reviewer’s suggestions, we address these concerns by referencing our previous paper. Cck+ cells were chosen because this marker is expressed in over 90% of PMd cells (Wang et al., 2021), but not in adjacent nuclei (Mickelsen et al., 2020). These cells have also been shown to be important to control escape from innate threats, such as carbon dioxide (Wang et al., 2021). These are the justifications for selecting PMd-cck cells, as discussed in this revised submission. We also reference our prior work to indicate specific expression of GCaMP in PMd cck cells.

2) The authors used male and female mice in their experiments but there are no analyses of potential sex differences in threat responses or escape vigor. Were there any significant sex differences in the measurements presented in Figure 1? A supplementary figure showing data for male and female mice would be helpful. Also, for Figure 1, please display the individual data points so that the reader can appreciate the variability in the behavioral responses. How many approaches and escapes are observed in each test? What is the average duration of a freezing bout?

As the results reported in Figure 1 summarize data from a rather large cohort (n=32), we decided it best for clarity's sake to show the variability in behavioral responses as a histogram of the difference scores for each animal (threat - control), now included as Figure 1, Figure Supplement 1, as well as below (Figure R7). Showing 32 individual data points may make the figure difficult to visualize (but of course, we can instead plot these individual points if the Reviewer prefers that instead of the plots shown below). At the Reviewer's request, we have also included the number of approaches and escapes in Figure 1 and the supplement. The average duration of a freezing bout is 2.03s ± 0.15 and is now reported in the Results section. There were no significant sex differences in the Figure 1 measures, and this is stated in the text, as well as plotted below in Figure R8 (male n=17, female n=15; Wilcoxon rank-sum test, p>0.05).

Figure R7. (Also Figure 1, figure supplement 1) Distribution of the difference scores for threat - control assays. Histograms depict the difference scores for all mice, threat - control, for each behavioral metric in Figure 1. The dotted red line indicates zero, or no difference between threat and control (n=32 mice).

Figure R8 (Also Figure 1, Figure supplement 3). Distribution of the difference scores for threat - control assays for males and females. Histograms depict the difference scores for all mice, threat - control, for each behavioral metric in Figure 1, separately for males (green) and females (purple). The dotted red line indicates zero, or no difference between threat and control (male n=17, female n=15). No significant differences (p>0.05) were found between males and females in any of the metrics plotted.

3) In Fig. 2, there appears to be sustained activity of CCK+ neurons after the onset of threat approach, and ramping activity preceding stretch-attend. In-depth analysis of these responses may be beyond the scope of this study, but the findings should be discussed since the representation of approach-related behaviors indicates the PMd is involved in more general representation of threat proximity, rather than simply escape vigor.

We agree with the Reviewer that PMd-activity represents distance to both innate and conditioned threats. We also include new data showing that PMd-dlPAG mutual information increases in the presence of threats (Figure R2 and Figure 9). Taken together, these data show that PMd activity encodes more than just escape vigor. We have altered the text to emphasize these results. These dual-site recordings were done contralaterally, so that dlPAG-syn cell body GCaMP signals are not contaminated by GCaMP-expressing PMd-cck axon terminals in the dlPAG.

4) The authors state that PMd CCK neuronal activity regulates escape vigor. Although the authors show a correlation of the calcium signal amplitude and escape distance in Fig. 2I, a correlation with escape velocity would be a more convincing measure of vigor.

PMd-cck neural activity is related to escape speed, as shown by single cell miniaturized microscopy recordings. Figure 4D shows that PMd ensemble activity can predict escape speed from threats, but not control stimuli. These results were specific to escape, as PMd activity did not encode approach speed towards threats or control stimuli (Figure 4D). Furthermore, we performed new analysis and showed that a greater number of PMd cells show activity significantly correlated with escape from threats, compared to control stimuli. Finally, we have additionally shown that, for the cells whose activity is significantly correlated with escape speed, the mutual information between escape speed and df/F is significantly greater for threat than control. This has now been included as Figure 3I-K (same as Figure R9 below).

Figure R9. A higher fraction of PMd-cck cells are correlated with escape speed during exposure to threats. (Also Figure 3I-J) (A) Traces show the z-scored df/F (blue) and speed (gray) for one cell classified as a speed cell in the rat exposure assay (top) and one non-correlated cell from the toy rat assay (bottom). Individual escape epochs are indicated by red boxes. (B) Bars show the percent of cells that significantly correlate with escape speed. (Fisher's exact test; toy rat: n correlated = 56, n non-correlated = 405; rat: n correlated = 100, n non-correlated = 366; pre-shock: n correlated = 50, n non-correlated = 571; fear retrieval: n correlated = 122, n non-correlated = 391) (C) Bars show the mutual information in bits between escape speed and calcium activity for cells whose signals were significantly correlated with escape speed in (J). (Wilcoxon rank sum test; toy rat n=56, rat n=100; pre-shock n=50, fear retrieval n=122). p<0.001.

Unfortunately, the lower resolution provided by photometry did not reveal consistent correlations with escape velocity across assays. Despite this lack of single cell resolution, PMd-cck photometry amplitude correlated with escape velocity during exposure to the rat, but not the toy rat, as shown below (Figure R10). However, this result was not replicated in the fear retrieval assay. Taken together, these data show that PMd activity is indeed related to escape vigor.

Figure R10. Escape speed correlates with PMd-cck photometry amplitude during rat exposure. Bars depict the Spearman r-value of escape speed and PMd-cck photometry df/F (z-scored) amplitude during exposure to rat and toy rat. (n=9 mice) p<0.001.

5) The changes in prediction error from control to threat contexts in Figs. 4B and 4D are compelling, but the prediction error in the threat context seems high. Can the authors provide a basis for what constitutes a 'good' error score?

We have now included the chance error, calculated by training and testing the GLM on circularly permuted data across mice and indicated below with a dotted red line in Figure 4 and its supplement. The Methods have also been updated to reflect this new aspect of the analysis. A ‘good’ error would be a value that is significantly lower than the error expected by chance, which is indicated by the red dashed line in Figure R11.

Fig. R11. (Also Figure 4B, 4D and Figure 4, figure supplement 1) (A) Bars show the mean squared error (MSE) of the GLM-predicted location from the actual location. The MSE is significantly lower for threat than control assays (Wilcoxon signed-rank test; n=9 mice). The dotted red line indicates chance error, calculated by training and testing the GLM on circularly permuted data. Only threat assay error was significantly lower than chance (Wilcoxon signed-rank test; rat p<0.001, fear retrieval p=0.003). (B) Bars depict the MSE of the GLM-predicted velocity away from (left) and towards (right) the threat. The GLM more accurately decodes threat than control velocities for samples in which the mice move away from the threat (top). Only threat assay error was significantly lower than chance (Wilcoxon signed-rank test; rat p=0.004, fear retrieval p=0.012). (C) Bars depict the mean squared error of the GLM-predicted speed. The GLM more accurately decodes threat than control speeds. Only threat assay error was significantly lower than chance (rat p<0.020, fear retrieval p=0.040). (Wilcoxon signed-rank test; n=9 mice) p<0.01.

6) Off-target effects are a potential concern at the dose of CNO used (5 mg/kg). For example, the increased approach speed with CNO in the YFP control group (Fig. 5D) may be a result of the high CNO dose. How was the dose of CNO selected?

This dose was selected based on our prior experience using the same dose to study PMd-cck cells in our prior Neuron paper. Additionally, this is a common dose, used in many papers. Indeed, there are several recent neuroscience papers published in this journal, eLife, that use this exact dose of CNO (Chen et al., 2016; Halbout et al., 2019; Ito et al., 2020; Kwak and Jung, 2019; Li et al., 2020; Mukherjee et al., 2021; O’Hare et al., 2017; Patel et al., 2019).

Although in this particular case approach velocity trended higher after CNO treatment, this is not a consistent result. We ran another cohort of control mice (n=9 saline, 9 CNO 5 mg/kg) and show that no such trend in approach velocity to the shock grid was observed during fear retrieval (Figure R12).

Fig. R12. CNO has no effect on approach velocity in a separate control cohort. The experimental protocol was performed as described in Figure 1B for a control cohort. For this group, CNO injection had no significant effect on approach speed (Wilcoxon signed-rank test, n=9).

7) Given the visible trends in the data, the number of animals used in Fig. 6B is insufficient to make conclusions about the behavioral effect of optogenetic excitation of PMd CCK neurons. Either more animals should be added, or the analysis should be limited to the Fos staining.

At the Reviewer's request, we have increased the number of animals in this analysis and found the results unchanged. Figure 6B has been replaced in the main manuscript (same as Figure R13 below). The addition of these new animals also erased the previous non-significant trends seen with fewer animals.

Figure R13. (Also Figure 6B) Delivery of blue light increases speed in PMd-cck ChR2 mice, but not stretch-attend postures or freeze bouts. (PMd-cck YFP n=6, PMd-cck ChR2 n=8; Wilcoxon rank-sum test).

Author Response

Reviewer #1 (Public Review):

Briggs et al use a combination of mathematical modelling and experimental validation to tease apart the contributions of metabolic and electronic coupling to the pancreatic beta cell functional network. A number of recent studies have shown the existence of functional beta cell subpopulations, some of which are difficult to fully reconcile with established electrophysiological theory. More generally, the contribution of beta cell heterogeneity (metabolism, differentiation, proliferation, activity) to islet function cannot be explained by existing combined metabolic/electrical oscillator models. The present studies are thus timely in modelling the islet electrical (structural) and functional networks. Importantly, the authors show that metabolic coupling primarily drives the islet functional network, giving rise to beta cell subpopulations. The studies, however, do not diminish the critical role of electrical coupling in dictating glucose responsiveness, network extent as well as longer-range synchronization. As such, the studies show that islet structural and functional networks both act to drive islet activity, and that conclusions on the islet structural network should not be made using measures of the functional network (and vice versa).

Strengths:

• State-of-the-art multi-parameter modelling encompassing electrical and metabolic components.

• Experimental validation using advanced FRAP imaging techniques, as well as Ca2+ data from relevant gap junction KO animals.

• Well-balanced arguments that frame metabolic and electrical coupling as essential contributors to islet function.

• Likely to change how the field models functional connectivity and beta cell heterogeneity.

Weaknesses:

• Limitations of FRAP and electrophysiological gap junction measures not considered.

• Limitations of Cx36 (gap junction) KO animals not considered.

• Accuracy of citations should be improved in a few cases.

We thank reviewer 1 for their positive comments, including the many strengths in the approaches, arguments and impact. We do note the weaknesses raised by the reviewer and have addressed them following the comments below.

We would like to also note that when we refer to metabolic activity driving the functional network, we are not referring to metabolic coupling between beta cells. Rather we mean that two cells that show either high levels of metabolic activity (glycolytic flux) or that show similar levels metabolic activity will show increased synchronization and thus a functional network edge as compares to cells with elevated gap junction conductance. Increased metabolic activity would likely generate increased depolarizing currents that will provide an increased coupling current to drive synchronization; whereas similar metabolic activity would mean a given coupling current could more readily drive synchronized activity. We have substantially rewritten the manuscript to clarify this point.

Reviewer #2 (Public Review):

In their present work, Briggs et al. combine biophysical simulations and experimental recordings of beta cell activity with analyses of functional network parameters to determine the role played by gap-junctional coupling, metabolism, and KATP conductance in defining the functional roles that the cells play in the functional networks, assess the structure-function relationship, and to resolve an important current open question in the field on the role of so-called hub cells in islets of Langerhans.

Combining differential equation-based simulations on 1000 coupled cells with demanding calcium, NAPDH, and FRAP imaging, as well as with advanced network analyses, and then comparing the network metrics with simulated and experimentally determined properties is an achievement in its own right and a major methodological strength. The findings have the potential to help resolve the issue of the importance of hub cells in beta cell networks, and the methodological pipeline and data may prove invaluable for other researchers in the community.

However, methodologically functional networks may be based on different types of calcium oscillations present in beta cells, i.e., fast oscillations produced by bursts of electrical activity, slow oscillations produced by metabolic/glycolytic oscillations, or a mixture of both. At present, the authors base the network analyses on fast oscillations only in the case of simulated traces and on a mixture of fast and slow oscillations in the case of experimental traces. Since different networks may depend on the studied beta cell properties to a different extent (e.g., fast oscillation-based networks may, more importantly, depend on electrical properties and slow oscillationbased networks may more strongly depend on metabolic properties), it is important that in drawing the conclusions the authors separately address the influence of a cell's electrical and metabolic properties on its functional role in the network based on fast oscillations, slow oscillations, or a mixture of both.

We thank reviewer 2 for their positive comments, including addressing the importance of this study as it pertains to islet biology and acknowledging methodological complexities of this study. We also thank the reviewer for their careful reading and providing useful comments. We have integrated each comment into the manuscript. Most importantly, we have now extended our analysis to both fast and slow oscillations by incorporating an additional mathematical model of coupled slow oscillations and performing additional experimental analysis of fast, slow, and mixed oscillations.

Reviewer #3 (Public Review):

Over the past decade, novel approaches to understanding beta cell connectivity and how that contributes to the overall function of the pancreatic islet have emerged. The application of network theory to beta cell connectivity has been an extremely useful tool to understand functional hierarchies amongst beta cells within an islet. This helps to provide functional relevance to observations from structural and gene expression data that beta cells are not all identical.

There are a number of "controversies" in this field that have arisen from the mathematical and subsequent experimental identification of beta "hub" cells. These are small populations of beta cells that are very highly connected to other beta cells, as assessed by applying correlation statistics to individual beta cell calcium traces across the islet.

In this paper Briggs et al set out to answer the following areas of debate:

They use computational datasets, based on established models of beta cells acting in concert (electrically coupled) within an islet-like structure, to show that it is similarities in metabolic parameters rather than "structural" connections (ie proximity which subserves gap junction coupling) that drives functional network behaviour. Whilst the computational models are quite relevant, the fact that the parameters (eg connectivity coefficients) are quite different to what is measured experimentally, confirm the limitations of this model. Therefore it was important for the authors to back up this finding by performing both calcium and metabolic imaging of islet beta cells. These experimental data are reported to confirm that metabolic coupling was more strongly related to functional connectivity than gap junction coupling. However, a limitation here is that the metabolic imaging data confirmed a strong link between disconnected beta cells and low metabolic coupling but did not robustly show the opposite. Similarly, I was not convinced that the FRAP studies, which indirectly measured GJ ("structural") connections were powered well enough to be related to measures of beta cell connectivity.

The group goes on to provide further analytical and experimental data with a model of increasing loss of GJ connectivity (by calcium imaging islets from WT, heterozygous (50% GJ loss), and homozygous (100% loss). Given the former conclusion that it was metabolic not GJ connectivity that drives small world network behaviour, it was surprising to see such a great effect on the loss of hubs in the homs. That said, the analytical approaches in this model did help the authors confirm that the loss of gap junctions does not alter the preferential existence of beta cell connectivity and confirms the important contribution of metabolic "coupling". One perhaps can therefore conclude that there are two types of network behaviour in an islet (maybe more) and the field should move towards an understanding of overlapping network communities as has been done in brain networks.

Overall this is an extremely well-written paper which was a pleasure to read. This group has neatly and expertly provided both computational and experimental data to support the notion that it is metabolic but not "structural" ie GJ coupling that drives our observations of hubs and functional connectivity. However, there is still much work to do to understand whether this metabolic coupling is just a random epiphenomenon or somehow fated, the extent to which other elements of "structural" coupling - ie the presence of other endocrine cell types, the spatial distribution of paracrine hormone receptors, blood vessels and nerve terminals are also important.

We thank reviewer 3 for their positive comments, including the methodology, writing style, and the importance of this paper to the broader islet community. We thank the reviewer for their very in-depth and helpful comments. We have addressed each comment below and made significant changes to the manuscript according. We conducted more FRAP experiments and separated results into slow, fast, and mixed oscillations. We included analysis of an additional computational model that simulates slow calcium oscillations. Additionally, we substantially rewrote the paper to clarify that we are not referring to metabolic coupling and speak on the broader implications of network theory and our findings.

Reviewer #4 (Public Review):

This manuscript describes a complex, highly ambitious set of modeling and experimental studies that appear designed to compare the structural and functional properties of beta cell subpopulations within the islet network in terms of their influence on network synchronization. The authors conclude that the most functionally coupled cell subpopulations in the islet network are not those that are most structurally coupled via gap junctions but those that are most metabolically active.

Strengths of the paper include (1) its use of an interdisciplinary collection of methods including computer simulations, FRAP to monitor functional coupling by gap junctions, the monitoring of Ca2+ oscillations in single beta cells embedded in the network, and the use of sophisticated approaches from probability theory. Most of these methods have been used and validated previously. Unfortunately, however, it was not clear what the underlying premise of the paper actually is, despite many stated intentions, nor what about it is new compared to previous studies, an additional weakness.

Although the authors state that they are trying to answer 3 critical questions, it was not clear how important these questions are in terms of significance for the field. For example, they state that a major controversy in the field is whether network structure or network function mediates functional synchronization of beta cells within the islet. However, this question is not much debated. As an example, while it is known that there can be long-range functional coupling in islets, no workers in the field believe there is a physical structure within islets that mediates this, unlike the case for CNS neurons that are known to have long projections onto other neurons. Beta cells within the islets are locally coupled via gap junctions, as stated repeatedly by the authors but these mediate short-range coupling. Thus, there are clearly functional correlations over long ranges but no structures, only correlated activity. This weakness raises questions about the overall significance of the work, especially as it seems to reiterate ideas presented previously.

We thank reviewer 4 for their positive comments, including our multidisciplinary use of mathematical models and experimental imaging techniques. We have now included an additional model of slow oscillations (the Integrated Oscillator Model) to improve our conclusions. We also thank reviewer 4 for the insightful comments. We have carefully reviewed each comment and made significant changes to the manuscript accordingly. In particular, we have significantly rewritten the introduction and discussion attempting to clarify what is new in our manuscript and what is previously shown. Additionally, we agree with the reviewers’ sentiment that there is little debate over whether, for example, there are physical structures within the islet that mediate long-range functional connections. However, there is current debate over whether functional beta-cell subpopulations can dictate islet dynamics (see [11]–[13]). This debate can be framed by observing whether these functional subpopulations emerge from the islet due to physical connections (structural network) or something more nuisance (such as intrinsic dynamics). We have reframed the introduction and discussion to clarify this debate as well as more clearly state the premise of the paper.

Specific Comments

1). The authors state it is well accepted that the disruption of gap junctional coupling is a pathophysiological characteristic of diabetes, but this is not an opinion widely accepted by the field, although it has been proposed. The authors should scale back on such generalizations, or provide more compelling evidence to support such a claim.

Thank you for pointing this out, we have provided more specific citations and changes the wording from “well accepted” to “has been documented”. See Discussion page 13 lines 415-416.

2) The paper relies heavily on simulations performed using a version of the model of Cha et al (2011). While this is a reasonable model of fast bursting (e.g. oscillations having periods <1 min.), the Ca2+ oscillations that were recorded by the authors and shown in Fig. 2b of the manuscript are slow oscillations with periods of 5 min and not <1 min, which is a weakness of the model in the current context. Furthermore, the model outputs that are shown lack the well-known characteristics seen in real islets, such as fast-spiking occurring on prolonged plateaus, again as can be seen by comparing the simulated oscillations shown in Fig. 1d with those in Fig. 2b. It is recommended that the simulations be repeated using a more appropriate model of slow oscillations or at least using the model of Cha et al but employed to simulate in slower bursting.

The reviewer raises an important point and caveat associated with our simulated model and experimental data. This point was also made by other reviewers, and a similar response to this comment can be found elsewhere in response to reviewer 2 point 6. To address this comment, we have performed several additional experiments and analyses:

1) We collected additional Ca2+ (to identify the functional network and hubs) and FRAP data (to assess gap junction permeability) in islets which show either pure slow, pure fast, or mixed oscillations. We generated networks based on each time scale to compare with FRAP gap junction permeability data. We found that the conclusions of our first draft to be consistent across all oscillation types. There was no relationship between gap junction conductance, as approximated using FRAP, and normalized degree for slow (Figure 3j), fast (Figure 3 Supp 1d,e), or mixed (Figure 3 Supp 1g,h) oscillations. We also include discussion of these conclusions - See Results page 7 lines 184-186 and lines 188-191, Discussion page 12 lines 357-360.

2) We also performed additional simulations with a coupled ‘Integrated Oscillator Model’ which shows slow oscillations because of metabolic oscillations (Figure 2). We compared connectivity with gap junction coupling and underlying cell parameters. In this case, there is an association between functional and structural networks, with highly-connected hub cells showing higher gap junction conductance (Figure 2f) but also low KATP channel conductance (gKATP) (Figure 2e). However, there are some caveats to these findings – given the nature of the IOM model, we were limited to simulating smaller islets (260 cells) and less heterogeneity in the calcium traces was observed. Additional analysis suggests the greater association between functional and structural networks in this model was a result of the smaller islets, and the association was also dependent on threshold (unlike in the Cha-Noma fast oscillator model) robust. These limitations and results are discussed further (Discussion page 11 lines 344-354).

Additionally, in the IOM, the underlying cell dynamics of highly-connected hub cells are differentiated by KATP channel conductance (gKATP), which is different than in the fast oscillator model (differentiated by metabolism, kglyc). However this difference between models can be linked to differences in the way duty cycle is influenced by gKATP and kglyc (Figure 1h, Figure 2g). In each model there was a similar association between duty cycle and highly-connected hub cells. We also discuss these findings (Discussion page 11 lines 334-343).

Overall these results and discussion with respect to the coupled IOM oscillator model can be found in Figure 2, Results page 6 lines 128-156 and Discussion page 11 lines 332-354.

3) Much of the data analyzed whether obtained via simulation or through experiment seems to produce very small differences in the actual numbers obtained, as can be seen in the bar graphs shown in Figs. 1e,g for example (obtained from simulations), or Fig. 2j (obtained from experimental measurements). The authors should comment as to why such small differences are often seen as a result of their analyses throughout the manuscript and why also in many cases the observed variance is high. Related to the data shown, very few dots are shown in Figs. 1eg or Fig 4e and 4h even though these points were derived from simulations where 100s of runs could be carried out and many more points obtained for plotting. These are weaknesses unless specific and convincing explanations are provided.

We thank the reviewer for these comments, which are similar to those of reviewer 2 (point 4) and reviewer 3 (point 6). Indeed there is some variability between cells in both simulations and experiments related to the metabolic activity in hubs and non-hubs. The variability points to potentially other factors being involved in determining hubs beyond simply kglyc, including a minor role for gap junction coupling structural network and potentially cell position and other intrinsic factors. We now discuss this point – see Discussion page 12 lines 364-266.

The differences between hubs and nonhubs appear small because the value of kglyc is very small. For figure 1e, the average kglyc for nonhubs was 1.26x10-4 s-1 (which is the average of the distribution because most cells are non hubs) while the average kglyc for hubs was 1.4x10-4 s-1 which is about half of a standard deviation higher. The paired t-test controls for the small value of average kglyc.

For simulation data each of the 5 dots corresponds to a simulated islet averaged over 1000 cells (or 260 cells for coupled IOM). The computational resources are high to generate such data so it is not feasible to conduct 100s of runs. Again, we note the comparisons between hubs and non-hubs are paired, and we find statistically significant differences for kglyc in figure 1 using only 5 paired data points. That we find these differences indicates the substantial difference between hubs and non-hubs. This is further supported all effect sizes being much greater than 0.8 for all significantly different findings (Cha Noma - kglyc: 2.85, gcoup: 0.82) (IOM: gKATP: 1.27, gcoup: 2.94) – We have included these effect sizes in the captions see Figure 1 and 2 captions (pages 34, 36)

To consider all of the available data rather than the average across an entire islet, we created a kernel density estimate the kglyc for hubs and nonhubs created by concatenating every single cell in each of the five islets. A kstest results in a highly significant difference (P<0.0001) between these two distributions.

Author response image 1.

4) The data shown in Fig. 4i,j are intended to compare long-range synchronization at different distances along a string of coupled cells but the difference between the synchronized and unsynchronized cells for gcoup and Kglyc was subtle, very much so.

Thank you for pointing out these subtle differences. The y-axis scale for i and j is broad to allow us to represent all distances on a single plot. After correction for multiple comparison, the differences were still statistically significant. As the reviewer mentioned in point 3, each plot contains only five data points, each of which represent the average of a single simulated islet, therefore we are not concerned about statistical significance coming from too large of a sample size. We also checked the differences between synchronized and nonsynchronized cell pairs in figure 4 panels e and h (now figure 5 e, h). These are the same data as i and j but normalized such that all of the distances could be averaged together. We again found statistical significance between synchronized and non-synchronized cell pairs. As can be seen in Author response image 2 the difference between synchronized and non-synchronized cell pairs is greater than the variability between simulated islets. Thus, in this case the variability is not substantial.

Author response image 2.

5) The data shown in Fig. 5 for Cx36 knockout islets are used to assess the influence of gap junctional coupling, which is reasonable, but it would be reassuring to know that loss of this gene has no effects on the expression of other genes in the beta cell, especially genes involved with glucose metabolism.

This is an important point. Previous studies have assessed that no significant change in NAD(P)H is observed in Cx36 deficient islets – see Benninger et al J.Physiol 2011 [14]. Islet architecture is also retained. Further the insulin secretory response of dissociated Cx36 knockout beta cells is the same as that of dissociated wildtype beta cells, further indicating no significant defect in the intrinsic ability of the beta cell to release insulin – see Benninger et al J.Physiol 2011 [14]. We now Mention these findings in the discussion. See Discussion page 14 lines 459-464.

6) In many places throughout the paper, it is difficult to ascertain whether what is being shown is new vs. what has been shown previously in other studies. The paper would thus benefit strongly from added text highlighting the novelty here and not just restating what is known, for instance, that islets can exhibit small-world network properties. This detracts from the strengths of the paper and further makes it difficult to wade through. Even the finding here that metabolic characteristics of the beta cells can infer profound and influential functional coupling is not new, as the authors proposed as much many years ago. Again, this makes it difficult to distill what is new compared to what is mainly just being confirmed here, albeit using different methods.

Thank you for the suggestion, we have made significant modifications throughout the Introduction, Discussion and Results to be clearer about what is known from previous work and what is newly found in this manuscript.

Reviewer #5 (Public Review):

The authors use state-of-the-art computation, experiment, and current network analysis to try and disaggregate the impact of cellular metabolism driving cellular excitability and structural electrical connections through gap junctions on islet synchronization. They perform interesting simulations with a sophisticated mathematical model and compare them with closely associated experiments. This close association is impressive and is an excellent example of using mathematics to inform experiments and experimental results. The current conclusions, however, appear beyond the results presented. The use of functional connectivity is based on correlated calcium traces but is largely without an understood biophysical mechanism. This work aims to clarify such a mechanism between metabolism and structural connection and comes out on the side of metabolism driving the functional connectivity, but both are required and more nuanced conclusions should be drawn.

We thank reviewer 5 for their positive comments, including our multifaceted experimental and computational techniques. We also found the reviewers careful reading and thoughtful comments to be very helpful and we have worked to integrate each comment into our manuscript. It is evident from the reviewer comments that we did not clearly explain what was meant by our conclusions concerning the functional network reflecting metabolism rather than gap junctions. We have conducted significant rewriting to show that we are not concluding that communication (metabolic or electric) occurs due to conduits other than gap junctions. Rather, our data suggest that the functional network (which reflects calcium synchronization) reflects intrinsic dynamics of the cells, which include metabolic rates, more than individual gap junction connections.

References referred to in this response to reviewers document:

[1] A. Stožer et al., “Functional connectivity in islets of Langerhans from mouse pancreas tissue slices,” PLoS Comput Biol, vol. 9, no. 2, p. e1002923, 2013.

[2] N. L. Farnsworth, A. Hemmati, M. Pozzoli, and R. K. Benninger, “Fluorescence recovery after photobleaching reveals regulation and distribution of connexin36 gap junction coupling within mouse islets of Langerhans,” The Journal of physiology, vol. 592, no. 20, pp. 4431–4446, 2014.

[3] C.-L. Lei, J. A. Kellard, M. Hara, J. D. Johnson, B. Rodriguez, and L. J. Briant, “Beta-cell hubs maintain Ca2+ oscillations in human and mouse islet simulations,” Islets, vol. 10, no. 4, pp. 151–167, 2018.

[4] N. R. Johnston et al., “Beta cell hubs dictate pancreatic islet responses to glucose,” Cell metabolism, vol. 24, no. 3, pp. 389–401, 2016.

[5] V. Kravets et al., “Functional architecture of pancreatic islets identifies a population of first responder cells that drive the first-phase calcium response,” PLoS Biology, vol. 20, no. 9, p. e3001761, 2022.

[6] H. Ren et al., “Pancreatic α and β cells are globally phase-locked,” Nature Communications, vol. 13, no. 1, p. 3721, 2022.

[7] A. Stožer et al., “From Isles of Königsberg to Islets of Langerhans: Examining the function of the endocrine pancreas through network science,” Frontiers in Endocrinology, vol. 13, p. 922640, 2022.

[8] J. Zmazek et al., “Assessing different temporal scales of calcium dynamics in networks of beta cell populations,” Frontiers in physiology, vol. 12, p. 337, 2021.

[9] M. E. Corezola do Amaral et al., “Caloric restriction recovers impaired β-cell-β-cell gap junction coupling, calcium oscillation coordination, and insulin secretion in prediabetic mice,” American Journal of Physiology-Endocrinology and Metabolism, vol. 319, no. 4, pp. E709–E720, 2020.

[10] J. M. Dwulet, J. K. Briggs, and R. K. P. Benninger, “Small subpopulations of beta-cells do not drive islet oscillatory [Ca2+] dynamics via gap junction communication,” PLOS Computational Biology, vol. 17, no. 5, p. e1008948, May 2021, doi: 10.1371/journal.pcbi.1008948.

[11] B. E. Peercy and A. S. Sherman, “Do oscillations in pancreatic islets require pacemaker cells?,” Journal of Biosciences, vol. 47, no. 1, pp. 1–11, 2022.

[12] G. A. Rutter, N. Ninov, V. Salem, and D. J. Hodson, “Comment on Satin et al.‘Take me to your leader’: an electrophysiological appraisal of the role of hub cells in pancreatic islets. Diabetes 2020; 69: 830–836,” Diabetes, vol. 69, no. 9, pp. e10–e11, 2020.

[13] L. S. Satin and P. Rorsman, “Response to comment on satin et al.‘Take me to your leader’: An electrophysiological appraisal of the role of hub cells in pancreatic islets. Diabetes 2020; 69: 830–836,” Diabetes, vol. 69, no. 9, pp. e12–e13, 2020.

[14] R. K. Benninger, W. S. Head, M. Zhang, L. S. Satin, and D. W. Piston, “Gap junctions and other mechanisms of cell–cell communication regulate basal insulin secretion in the pancreatic islet,” The Journal of physiology, vol. 589, no. 22, pp. 5453–5466, 2011.

[15] R. Fried, Erectile dysfunction as a cardiovascular impairment. Academic Press, 2014. [16] T. Pipatpolkai, S. Usher, P. J. Stansfeld, and F. M. Ashcroft, “New insights into KATP channel gene mutations and neonatal diabetes mellitus,” Nature Reviews Endocrinology, vol. 16, no. 7, pp. 378–393, 2020.

[17] A. M. Notary, M. J. Westacott, T. H. Hraha, M. Pozzoli, and R. K. P. Benninger, “Decreases in Gap Junction Coupling Recovers Ca2+ and Insulin Secretion in Neonatal Diabetes Mellitus, Dependent on Beta Cell Heterogeneity and Noise,” PLOS Computational Biology, vol. 12, no. 9, p. e1005116, Sep. 2016, doi: 10.1371/journal.pcbi.1005116.

[18] J. V. Rocheleau, G. M. Walker, W. S. Head, O. P. McGuinness, and D. W. Piston, “Microfluidic glucose stimulation reveals limited coordination of intracellular Ca2+ activity oscillations in pancreatic islets,” Pro ceedings of the National Academy of Sciences, vol. 101, no. 35, pp. 12899–12903, 2004. [19] R. K. Benninger, M. Zhang, W. S. Head, L. S. Satin, and D. W. Piston, “Gap junction coupling and calcium waves in the pancreatic islet,” Biophysical journal, vol. 95, no. 11, pp. 5048–5061, 2008.

What an odd statement. They're basically lil broing them. The rulers were certainly not blindly loyal as we read in the head note. This statement seems intended to give the pretense that the British respect the people.

I looked at source [o11], the Vice article “Facebook Is Ignoring Moderators’ Trauma: ‘They Suggest Karaoke and Painting’”. The thing that really stuck in my head is how the company’s response sounds so shallow. These moderators are watching horrible stuff every day, like violence or hate, and then the solution is basically “do some hobbies, try karaoke.” It feel like treating serious psychological damage as just stress from a normal office job. For me this connects to the question in 15.3 about what support moderators should have. After reading this source, I think real support has to include proper therapy, higher pay, and maybe limits on how long someone can do this job. If a platform can afford global expansion and fancy offices, it should also afford not to ignore the people cleaning up its worst content.

Predictive, Oncogenic evidence:

Predictive: The study discusses how NOTCH1 mutations (NOTCH1MUT) correlate with sensitivity to PI3K inhibition, indicating that these mutations can influence the response to therapy. The mention of "inhibition of PI3K can selectively target NOTCH1 mutant HNSCC cells" supports this classification as it directly relates to treatment response.

Oncogenic: The abstract highlights that truncating and missense mutations in NOTCH1 are frequent in head and neck squamous cell carcinoma (HNSCC), suggesting that these somatic mutations contribute to tumor development or progression. The context of these mutations being prevalent in a cancer type supports the classification as oncogenic.

Predictive, Oncogenic evidence:

Predictive: The study discusses how NOTCH1 mutations (NOTCH1MUT) correlate with sensitivity to PI3K inhibition, indicating that these mutations can influence the response to therapy. The mention of "inhibition of PI3K can selectively target NOTCH1 mutant HNSCC cells" supports this classification as it directly relates to treatment response.

Oncogenic: The abstract highlights that truncating and missense mutations in NOTCH1 are frequent in head and neck squamous cell carcinoma (HNSCC), suggesting that these somatic mutations contribute to tumor development or progression. The context of these mutations being prevalent in a cancer type supports the classification as oncogenic.

Predictive, Oncogenic evidence:

Oncogenic: The study indicates that Rab18 functions as an oncoprotein in human head and neck squamous cell carcinoma (HNSCC), suggesting that it contributes to tumor development and progression. The mention of Rab18's role in regulating cell proliferation and invasion further supports its oncogenic potential.

Predictive: The abstract states that Rab18 influences cisplatin sensitivity in HNSCC, indicating a correlation between the variant and response to a specific therapy. This suggests that Rab18 may play a role in determining how well patients respond to cisplatin treatment.

Predictive, Oncogenic evidence:

Oncogenic: The study indicates that Rab18 functions as an oncoprotein in human head and neck squamous cell carcinoma (HNSCC), suggesting that it contributes to tumor development and progression. The mention of Rab18's role in regulating cell proliferation and invasion further supports its oncogenic potential.

Predictive: The abstract states that Rab18 influences cisplatin sensitivity in HNSCC, indicating a correlation between the variant and response to a specific therapy. This suggests that Rab18 may play a role in determining how well patients respond to cisplatin treatment.

Predictive, Oncogenic evidence:

Predictive: The study investigates how mutations in the AXL receptor, specifically tyrosine 821, contribute to resistance against cetuximab and radiation treatment in head and neck squamous cell carcinoma (HNSCC). The findings suggest that targeting AXL and c-ABL can overcome this resistance, indicating a correlation between the variant and treatment response.

Oncogenic: The research highlights that the signaling of AXL, particularly through the mutation of tyrosine 821, plays a role in the development of resistance in HNSCC, suggesting that this variant contributes to tumor progression and behavior in the context of cancer treatment.

Predictive, Oncogenic evidence:

Predictive: The study investigates how mutations in the AXL receptor, specifically tyrosine 821, contribute to resistance against cetuximab and radiation treatment in head and neck squamous cell carcinoma (HNSCC). The findings suggest that targeting AXL and c-ABL can overcome this resistance, indicating a correlation between the variant and treatment response.

Oncogenic: The research highlights that the signaling of AXL, particularly through the mutation of tyrosine 821, plays a role in the development of resistance in HNSCC, suggesting that this variant contributes to tumor progression and behavior in the context of cancer treatment.