• overheid moet gelijke gevallen gelijk behandelen, en ongelijke gevallen ongelijk
• burgers die in dezelfde situatie verkeren moeten hetzelfde behandeld worden door de overheid

• goed onderbouwd en overtuigend

houdt in dat - de argumenten inhoudelijk sterk genoeg zijn - ze logisch voortvloeien uit de feiten - en ze de beslissing echt kunnen dragen

• duidelijk zichtbaar en begrijpelijk voor de burger

houdt in dat: - de overheid de reden van haar besluit opschrijft in het besluit zelf - de burger dus kan zien op basis van welke argumenten en regels het besluit is genomen

• overheid moet haar besluiten goed onderbouwen en uitleggen
• overheid mag iets niet zomaar beslissen, ze moet de feiten, belangenafweging en regelgeving waarop het besluit gebaseerd is duidelijk vermelden

• overheid moet duidelijk, betrouwbaar en voorspelbaar handelen zodat burgers weten wat hun rechten en plichten zijn
• overheid moet duidelijk communiceren over regels en besluiten
• burgers moeten erop kunnen vertrouwen dat wat de overheid zegt of besluit blijft gelden

• de overheid mag haar bevoegdheden alleen gebruiken voor het doel waarvoor die bevoegdheid is gegeven: overheid mag haar macht niet gebruiken voor een ander doel dan waarvoor de wet haar die mag heeft gegeven

• overheid moet de verschillende belangen op een eerlijke en redelijke manier tegen elkaar afwegen voordat ze een besluit neemt

• de overheid mag niet een belang te zwaar laten wegen en een ander belang negeren

• gaat om het vinden van een goed balans tussen: het algemeen belang en de individuele belangen

overheid moet voorzichtig, zorgvuldig en volledig te werk gaan voordat ze een besluit neemt - ze moet goed nadenken, informatie verzamelen en de belangen van alle betrokkenen afwegen

formele zorgvuldigheid (in de voorbereiding van een besluit) - gaat over hoe de overheid tot een besluit komt - gaat dus om de procedure en de manier van werken

materiele zorgvuldigheid (inhoud van het besluit) - gaat over wat de overheid beslist - dus de inhoud en redelijkheid van het besluit zelf

• betekent dat de overheid eerlijk, open en onpartijdig moet handelen tegenover burgers
• burger moet ene eerlijke kans krijgen om zijn of haar standpunt te geven in een procedure

he’s drawing a line between everyday politeness and historically imposed performance. What begins as etiquette becomes survival strategy under domination. This “public transcript,” as he calls it, isn’t just a set of social cues it’s a choreography shaped by fear, prudence, and sometimes sheer exhaustion.

And Wald-Tinan’s anecdote is such a vivid example smiling at the landlord who dismissed your father, not because you respect him, but because appearing “amiable” protects you. It’s a performance built on unequal power, and it says so much about the psychic toll of domination.

there is no action possible without a little acting that’s powerful. It blurs the line between resistance and complicity, showing that even politeness can be strategic.

Últimas notícias de Itamar Franco

Invocation of the Muse

The legendary king of Mycenae and leader of the Greek army in the Trojan War of Homer's Illiad. Agamemnon is a great warrior but also a selfish ruler who famously upset his invincible champion Achilles, a feud that prolonged the war and suffering of his men. https://www.worldhistory.org/Agamemnon_(Person)/

eLife Assessment

This important manuscript provides compelling evidence that BK and CaV1.3 channels can co-localize as ensembles early in the biosynthetic pathway, including in the ER and Golgi. The findings, supported by a range of imaging and proximity assays, offer insights into channel organization in both heterologous and endogenous systems. While the data broadly support the central claims, mechanistic aspects remain unresolved, particularly regarding the determinants of mRNA co-localization, the temporal dynamics of ensemble trafficking, and the physiological implications of pre-assembly for channel function at the plasma membrane.

Reviewer #1 (Public review):

Summary:

This manuscript by Pournejati et al investigates how BK (big potassium) channels and CaV1.3 (a subtype of voltage-gated calcium channels) become functionally coupled by exploring whether their ensembles form early-during synthesis and intracellular trafficking-rather than only after insertion into the plasma membrane. To this end, the authors use the PLA technique to assess the formation of ion channel associations in the different compartments (ER, Golgi or PM), single-molecule RNA in situ hybridization (RNAscope), and super-resolution microscopy.

Strengths:

The manuscript is well written and addresses an interesting question, combining a range of imaging techniques. The findings are generally well-presented and offer important insights into the spatial organization of ion channel complexes, both in heterologous and endogenous systems.

Weaknesses:

The authors have improved their manuscript after revisions, and some previous concerns have been addressed. Still, the main concern about this work is that the current experiments do not quantitatively or mechanistically link the ensembles observed intracellularly (in the endoplasmic reticulum (ER) or Golgi) to those found at the plasma membrane (PM). As a result, it is difficult to fully integrate the findings into a coherent model of trafficking. Specifically, the manuscript does not address what proportion of ensembles detected at the PM originated in the ER. Without data on the turnover or half-life of these ensembles at the PM, it remains unclear how many persist through trafficking versus forming de novo at the membrane. The authors report the percentage of PLA-positive ensembles localized to various compartments, but this only reflects the distribution of pre-formed ensembles. What remains unknown is the proportion of total BK and CaV1.3 channels (not just those in ensembles) that are engaged in these complexes within each compartment. Without this, it is difficult to determine whether ensembles form in the ER and are then trafficked to the PM, or if independent ensemble formation also occurs at the membrane. To support the model of intracellular assembly followed by coordinated trafficking, it would be important to quantify the fraction of the total channel population that exists as ensembles in each compartment. A comparable ensemble-to-total ratio across ER and PM would strengthen the argument for directed trafficking of pre-assembled channel complexes.

Reviewer #2 (Public review):

Summary:

The co-localization of large conductance calcium- and voltage activated potassium (BK) channels with voltage-gated calcium channels (CaV) at the plasma membrane is important for the functional role of these channels in controlling cell excitability and physiology in a variety of systems.

An important question in the field is where and how do BK and CaV channels assemble as 'ensembles' to allow this coordinated regulation - is this through preassembly early in the biosynthetic pathway, during trafficking to the cell surface or once channels are integrated into the plasma membrane. These questions also have broader implications for assembly of other ion channel complexes.

Using an imaging based approach, this paper addresses the spatial distribution of BK-CaV ensembles using both overexpression strategies in tsa201 and INS-1 cells and analysis of endogenous channels in INS-1 cells using proximity ligation and superesolution approaches. In addition, the authors analyse the spatial distribution of mRNAs encoding BK and Cav1.3.

The key conclusion of the paper that BK and CaV1.3 are co-localised as ensembles intracellularly in the ER and Golgi is well supported by the evidence. However, whether they are preferentially co-translated at the ER, requires further work. Moreover, whether intracellular pre-assembly of BK-CaV complexes is the major mechanism for functional complexes at the plasma membrane in these models requires more definitive evidence including both refinement of analysis of current data as well as potentially additional experiments.

Strengths & Weaknesses

(1) Using proximity ligation assays of overexpressed BK and CaV1.3 in tsa201 and INS-1 cells the authors provide strong evidence that BK and CaV can exist as ensembles (ie channels within 40 nm) at both the plasma membrane and intracellular membranes, including ER and Golgi. They also provide evidence for endogenous ensemble assembly at the Golgi in INS-1 cells and it would have been useful to determine if endogenous complexes are also observe in the ER of INS-1 cells. There are some useful controls but the specificity of ensemble formation would be better determined using other transmembrane proteins rather than peripheral proteins (eg Golgi 58K).

(2) Ensemble assembly was also analysed using super-resolution (dSTORM) imaging in INS-1 cells. In these cells only 7.5% of BK and CaV particles (endogenous?) co-localise that was only marginally above chance based on scrambled images. More detailed quantification and validation of potential 'ensembles' needs to be made for example by exploring nearest neighbour characteristics (but see point 4 below) to define proportion of ensembles versus clusters of BK or Cav1.3 channels alone etc. For example, it is mentioned that a distribution of distances between BK and Cav is seen but data are not shown.

(3) The evidence that the intracellular ensemble formation is in large part driven by co-translation, based on co-localisation of mRNAs using RNAscope, requires additional critical controls and analysis. The authors now include data of co-localised BK protein that is suggestive but does not show co-translation. Secondly, while they have improved the description of some controls mRNA co-localisation needs to be measured in both directions (eg BK - SCN9A as well as SCN9A to BK) especially if the mRNAs are expressed at very different levels. The relative expression levels need to be clearly defined in the paper. Authors also use a randomized image of BK mRNA to show specificity of co-localisation with Cav1.3 mRNA, however the mRNA distribution would not be expected to be random across the cell but constrained by ER morphology if co-translated so using ER labelling as a mask would be useful?

(4) The authors attempt to define if plasma membrane assemblies of BK and CaV occur soon after synthesis. However, because the expression of BK and CaV occur at different times after transient transfection of plasmids more definitive experiments are required. For example, using inducible constructs to allow precise and synchronised timing of transcription. This would also provide critical evidence that co-assembly occurs very early in synthesis pathways - ie detecting complexes at ER before any complexes at Golgi or plasma membrane.

(5) While the authors have improved the definition of hetero-clusters etc it is still not clear in superesolution analysis, how they separate a BK tetramer from a cluster of BK tetramers with the monoclonal antibody employed ie each BK channel will have 4 binding sites (4 subunits in tetramer) whereas Cav1.3 has one binding site per channel. Thus, how do authors discriminate between a single BK tetramer (molecular cluster) with potential 4 antibodies bound compared to a cluster of 4 independent BK channels.

(6) The post-hoc tests used for one way ANOVA and ANOVA statistics need to be defined throughout

Reviewer #3 (Public review):

Summary:

The authors present a clearly written and beautifully presented piece of work demonstrating clear evidence to support the idea that BK channels and Cav1.3 channels can co-assemble prior to their assertion in the plasma membrane.

Strengths:

The experimental records shown back up their hypotheses and the authors are to be congratulated for the large number of control experiments shown in the ms.

Author response:

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

Recommendations for the Authors:

(1) Clarify Mechanistic Interpretations

(a) Provide stronger evidence or a more cautious interpretation regarding whether intracellular BK-CaV1.3 ensembles are precursors to plasma membrane complexes.

This is an important point. We adjusted the interpretation regarding intracellular BKCa_V1.3 hetero-clusters as precursors to plasma membrane complexes to reflect a more cautious stance, acknowledging the limitations of available data. We added the following to the manuscript.

“Our findings suggest that BK and Ca_V1.3 channels begin assembling intracellularly before reaching the plasma membrane, shaping their spatial organization and potentially facilitating functional coupling. While this suggests a coordinated process that may contribute to functional coupling, further investigation is needed to determine the extent to which these hetero-clusters persist upon membrane insertion.”

(b) Discuss the limitations of current data in establishing the proportion of intracellular complexes that persist on the cell surface.

We appreciate the suggestion. We expanded the discussion to address the limitations of current data in determining the proportion of intracellular complexes that persist on the cell surface. We added the following to the manuscript.

“Our findings highlight the intracellular assembly of BK-Ca_V1.3 hetero-clusters, though limitations in resolution and organelle-specific analysis prevent precise quantification of the proportion of intracellular complexes that ultimately persist on the cell surface. While our data confirms that hetero-clusters form before reaching the plasma membrane, it remains unclear whether all intracellular hetero-clusters transition intact to the membrane or undergo rearrangement or disassembly upon insertion. Future studies utilizing live cell tracking and high resolution imaging will be valuable in elucidating the fate and stability of these complexes after membrane insertion.”

(2) Refine mRNA Co-localization Analysis

(a) Include appropriate controls using additional transmembrane mRNAs to better assess the specificity of BK and CaV1.3 mRNA co-localization.

We agree with the reviewers that these controls are essential. We explain better the controls used to address this concern. We added the following to the manuscript. 

“To explore the origins of the initial association, we hypothesized that the two proteins are translated near each other, which could be detected as the colocalization of their mRNAs (Figure 5A and B). The experiment was designed to detect single mRNA molecules from INS-1 cells in culture. We performed multiplex in situ hybridization experiments using an RNAScope fluorescence detection kit to be able to image three mRNAs simultaneously in the same cell and acquired the images in a confocal microscope with high resolution. To rigorously assess the specificity of this potential mRNA-level organization, we used multiple internal controls. GAPDH mRNA, a highly expressed housekeeping gene with no known spatial coordination with channel mRNAs, served as a baseline control for nonspecific colocalization due to transcript abundance. To evaluate whether the spatial proximity between BK mRNA (KCNMA1) and Ca_V1.3 mRNA (CACNA1D) was unique to functionally coupled channels, we also tested for Na^V1.7 mRNA (SCN9A), a transmembrane sodium channel expressed in INS-1 cells but not functionally associated with BK. This allowed us to determine whether the observed colocalization reflected a specific biological relationship rather than shared expression context. Finally, to test whether this proximity might extend to other calcium sources relevant to BK activation, we probed the mRNA of ryanodine receptor 2 (RyR2), another Ca²⁺ channel known to interact structurally with BK channels [32]. Together, these controls were chosen to distinguish specific mRNA colocalization patterns from random spatial proximity, shared subcellular distribution, or gene expression level artifacts.”

(b) Quantify mRNA co-localization in both directions (e.g., BK with CaV1.3 and vice versa) and account for differences in expression levels.

We thank the reviewer for this suggestion. We chose to quantify mRNA co-localization in the direction most relevant to the formation of functionally coupled hetero-clusters, namely, the proximity of BK (KCNMA1) mRNA to Ca_V1.3 (CACNA1D) mRNA. Since BK channel activation depends on calcium influx provided by nearby Ca_V1.3 channels, this directional analysis more directly informs the hypothesis of spatially coordinated translation and channel assembly. To address potential confounding effects of transcript abundance, we implemented a scrambled control approach in which the spatial coordinates of KCNMA1 mRNAs were randomized while preserving transcript count. This control resulted in significantly lower colocalization with CACNA1D mRNA, indicating that the observed proximity reflects a specific spatial association rather than expressiondriven overlap. We also assessed colocalization of CACNA1D with both KCNMA1, GAPDH mRNAs and SCN9 (NaV1.7); as you can see in the graph below these data support t the same conclusion but were not included in the manuscript.

Author response image 1.

(c) Consider using ER labeling as a spatial reference when analyzing mRNA localization

We thank the reviewers for this suggestion. Rather than using ER labeling as a spatial reference, we assess BK and CaV1.3 mRNA localization using fluorescence in situ hybridization (smFISH) alongside BK protein immunostaining. This approach directly identifies BK-associated translation sites, ensuring that observed mRNA localization corresponds to active BK synthesis rather than general ER association. By evaluating BK protein alongside its mRNA, we provide a more functionally relevant measure of spatial organization, allowing us to assess whether BK is synthesized in proximity to CaV1.3 mRNA within micro-translational complexes. The results added to the manuscript is as follows.

“To further investigate whether KCNMA1 and CACNA1D are localized in regions of active translation (Figure 7A), we performed RNAScope targeting KCNMA1 and CACNA1D alongside immunostaining for BK protein. This strategy enabled us to visualize transcript-protein colocalization in INS-1 cells with subcellular resolution. By directly evaluating sites of active BK translation, we aimed to determine whether newly synthesized BK protein colocalized with CACNA1D mRNA signals (Figure 7A). Confocal imaging revealed distinct micro-translational complex where KCNMA1 mRNA puncta overlapped with BK protein signals and were located adjacent to CACNA1D mRNA (Figure 7B). Quantitative analysis showed that 71 ± 3% of all KCNMA1 colocalized with BK protein signal which means that they are in active translation. Interestingly, 69 ± 3% of the KCNMA1 in active translation colocalized with CACNA1D (Figure 7C), supporting the existence of functional micro-translational complexes between BK and Ca_V1.3 channels.”

(3) Improve Terminology and Definitions

(a) Clarify and consistently use terms like "ensemble," "cluster," and "complex," especially in quantitative analyses.

We agree with the reviewers, and we clarified terminology such as 'ensemble,' 'cluster,' and 'complex' and used them consistently throughout the manuscript, particularly in quantitative analyses, to enhance precision and avoid ambiguity.  

(b) Consider adopting standard nomenclature (e.g., "hetero-clusters") to avoid ambiguity.

We agree with the reviewers, and we adapted standard nomenclature, such as 'heteroclusters,' in the manuscript to improve clarity and reduce ambiguity.

(4) Enhance Quantitative and Image Analysis

(a) Clearly describe how colocalization and clustering were measured in super-resolution data.

We thank the reviewers for this suggestion. We have modified the Methods section to provide a clearer description of how colocalization and clustering were measured in our super-resolution data. Specifically, we now detail the image processing steps, including binary conversion, channel multiplication for colocalization assessment, and density-based segmentation for clustering analysis. These updates ensure transparency in our approach and improve accessibility for readers, and we added the following to the manuscript.

“Super-resolution imaging: 

Direct stochastic optical reconstruction microscopy (dSTORM) images of BK and 1.3 overexpressed in tsA-201 cells were acquired using an ONI Nanoimager microscope equipped with a 100X oil immersion objective (1.4 NA), an XYZ closed-loop piezo 736 stage, and triple emission channels split at 488, 555, and 640 nm. Samples were imaged at 35°C. For singlemolecule localization microscopy, fixed and stained cells were imaged in GLOX imaging buffer containing 10 mM β-mercaptoethylamine (MEA), 0.56 mg/ml glucose oxidase, 34 μg/ml catalase, and 10% w/v glucose in Tris-HCl buffer. Single-molecule localizations were filtered using NImOS software (v.1.18.3, ONI). Localization maps were exported as TIFF images with a pixel size of 5 nm. Maps were further processed in ImageJ (NIH) by thresholding and binarization to isolate labeled structures. To assess colocalization between the signal from two proteins, binary images were multiplied. Particles smaller than 400 nm² were excluded from the analysis to reflect the spatial resolution limit of STORM imaging (20 nm) and the average size of BK channels. To examine spatial localization preference, binary images of BK were progressively dilated to 20 nm, 40 nm, 60 nm, 80 nm, 100 nm, and 200 nm to expand their spatial representation. These modified images were then multiplied with the Ca_V1.3 channel to quantify colocalization and determine BK occupancy at increasing distances from Ca_V1.3. To ensure consistent comparisons across distance thresholds, data were normalized using the 200 nm measurement as the highest reference value, set to 1.”

(b) Where appropriate, quantify the proportion of total channels involved in ensembles within each compartment.

We thank the reviewers for this comment. However, our method does not allow for direct quantification of the total number of BK and Ca_V1.3 channels expressed within the ER or ER exit sites, as we rely on proximity-based detection rather than absolute fluorescence intensity measurements of individual channels. Traditional methods for counting total channel populations, such as immunostaining or single-molecule tracking, are not applicable to our approach due to the hetero-clusters formation process. Instead, we focused on the relative proportion of BK and Ca_V1.3 hetero-clusters within these compartments, as this provides meaningful insights into trafficking dynamics and spatial organization. By assessing where hetero-cluster preferentially localize rather than attempting to count total channel numbers, we can infer whether their assembly occurs before plasma membrane insertion. While this approach does not yield absolute quantification of ER-localized BK and Ca_V1.3 channels, it remains a robust method for investigating hetero-cluster formation and intracellular trafficking pathways. To reflect this limitation, we added the following to the manuscript.

“Finally, a key limitation of this approach is that we cannot quantify the proportion of total BK or Ca_V1.3 channels engaged in hetero-clusters within each compartment. The PLA method provides proximity-based detection, which reflects relative localization rather than absolute channel abundance within individual organelles”.

(5) Temper Overstated Claims

(a) Revise language that suggests the findings introduce a "new paradigm," instead emphasizing how this study extends existing models.

We agree with the reviewers, and we have revised the language to avoid implying a 'new paradigm.' The following is the significance statement.

“This work examines the proximity between BK and Ca_V1.3 molecules at the level of their mRNAs and newly synthesized proteins to reveal that these channels interact early in their biogenesis. Two cell models were used: a heterologous expression system to investigate the steps of protein trafficking and a pancreatic beta cell line to study the localization of endogenous channel mRNAs. Our findings show that BK and Ca_V1.3 channels begin assembling intracellularly before reaching the plasma membrane, revealing new aspects of their spatial organization. This intracellular assembly suggests a coordinated process that contributes to functional coupling.”

(b) Moderate conclusions where the supporting data are preliminary or correlative.

We agree with the reviewers, and we have moderated conclusions in instances where the supporting data are preliminary or correlative, ensuring a balanced interpretation. We added the following to the manuscript. 

“This study provides novel insights into the organization of BK and Ca_V1.3 channels in heteroclusters, emphasizing their assembly within the ER, at ER exit sites, and within the Golgi. Our findings suggest that BK and Ca_V1.3 channels begin assembling intracellularly before reaching the plasma membrane, shaping their spatial organization, and potentially facilitating functional coupling. While this suggests a coordinated process that may contribute to functional coupling, further investigation is needed to determine the extent to which these hetero-clusters persist upon membrane insertion. While our study advances the understanding of BK and Ca_V1.3 heterocluster assembly, several key questions remain unanswered. What molecular machinery drives this colocalization at the mRNA and protein level? How do disruptions to complex assembly contribute to channelopathies and related diseases? Additionally, a deeper investigation into the role of RNA binding proteins in facilitating transcript association and localized translation is warranted”.

(6) Address Additional Technical and Presentation Issues

(a) Include clearer figure annotations, especially for identifying PLA puncta localization (e.g., membrane vs. intracellular).

We agree with the reviewers, and we have updated the figures to include clearer annotations that distinguish PLA puncta localized at the membrane versus those within intracellular compartments.

(b) Reconsider the scale and arrangement of image panels to better showcase the data.

We agree with the reviewers, and we have adjusted the scale and layout of the image panels to enhance data visualization and readability. Enlarged key regions now provide better clarity of critical features.

(c) Provide precise clone/variant information for BK and CaV1.3 channels used.

We thank the reviewers for their suggestion, and we now provide precise information regarding the BK and Ca_V1.3 channel constructs used in our experiments, including their Addgene plasmid numbers and relevant variant details. These have been incorporated into the Methods section to ensure reproducibility and transparency. We added the following to the manuscript. 

“The Ca_V1.3 α subunit construct used in our study corresponds to the rat Ca_V1.3e splice variant containing exons 8a, 11, 31b, and 42a, with a deletion of exon 32. The BK channel construct used in this study corresponds to the VYR splice variant of the mouse BKα subunit (KCNMA1)”.

(d) Correct typographical errors and ensure proper figure/supplementary labeling throughout.

Typographical errors have been corrected, and figure/supplementary labeling has been reviewed for accuracy throughout the manuscript.

(7) Expand the Discussion

(a) Include a brief discussion of findings such as BK surface expression in the absence of CaV1.3.

We thank the reviewers for their suggestion. We expanded the Discussion to include a brief analysis of BK surface expression in the absence of Ca_V1.3. We included the following in the manuscript. 

“BK Surface Expression and Independent Trafficking Pathways

BK surface expression in the absence of Ca_V1.3 indicates that its trafficking does not strictly rely on Ca_V1.3-mediated interactions. Since BK channels can be activated by multiple calcium sources, their presence in intracellular compartments suggests that their surface expression is governed by intrinsic trafficking mechanisms rather than direct calcium-dependent regulation. While some BK and Ca_V1.3 hetero-clusters assemble into signaling complexes intracellularly, other BK channels follow independent trafficking pathways, demonstrating that complex formation is not obligatory for all BK channels. Differences in their transport kinetics further reinforce the idea that their intracellular trafficking is regulated through distinct mechanisms. Studies have shown that BK channels can traffic independently of Ca_V1.3, relying on alternative calcium sources for activation [13, 41]. Additionally, Ca_V1.3 exhibits slower synthesis and trafficking kinetics than BK, emphasizing that their intracellular transport may not always be coordinated. These findings suggest that BK and Ca_V1.3 exhibit both independent and coordinated trafficking behaviors, influencing their spatial organization and functional interactions”.

(b) Clarify why certain colocalization comparisons (e.g., ER vs. ER exit sites) are not directly interpretable.

We thank the reviewer for their suggestion. A clarification has been added to the result section and discussion of the manuscript explaining why colocalization comparisons, such as ER versus ER exit sites, are not directly interpretable. We included the following in the manuscript.

“Result:

ER was not simply due to the extensive spatial coverage of ER labeling, we labeled ER exit sites using Sec16-GFP and probed for hetero-clusters with PLA. This approach enabled us to test whether the hetero-clusters were preferentially localized to ER exit sites, which are specialized trafficking hubs that mediate cargo selection and direct proteins from the ER into the secretory pathway. In contrast to the more expansive ER network, which supports protein synthesis and folding, ER exit sites ensure efficient and selective export of proteins to their target destinations”.

“By quantifying the proportion of BK and Ca_V1.3 hetero-clusters relative to total channel expression at ER exit sites, we found 28 ± 3% colocalization in tsA-201 cells and 11 ± 2% in INS-1 cells (Figure 3F). While the percentage of colocalization between hetero-clusters and the ER or ER exit sites alone cannot be directly compared to infer trafficking dynamics, these findings reinforce the conclusion that hetero-clusters reside within the ER and suggest that BK and Ca_V1.3 channels traffic together through the ER and exit in coordination”.

“Colocalization and Trafficking Dynamics

The colocalization of BK and Ca_V1.3 channels in the ER and at ER exit sites before reaching the Golgi suggests a coordinated trafficking mechanism that facilitates the formation of multi-channel complexes crucial for calcium signaling and membrane excitability [37, 38]. Given the distinct roles of these compartments, colocalization at the ER and ER exit sites may reflect transient proximity rather than stable interactions. Their presence in the Golgi further suggests that posttranslational modifications and additional assembly steps occur before plasma membrane transport, providing further insight into hetero-cluster maturation and sorting events. By examining BK-Ca_V1.3 hetero-cluster distribution across these trafficking compartments, we ensure that observed colocalization patterns are considered within a broader framework of intracellular transport mechanisms [39]. Previous studies indicate that ER exit sites exhibit variability in cargo retention and sorting efficiency [40], emphasizing the need for careful evaluation of colocalization data. Accounting for these complexities allows for a robust assessment of signaling complexes formation and trafficking pathways”.

Reviewer #1 (Recommendations for the authors):

In addition to the general aspects described in the public review, I list below a few points with the hope that they will help to improve the manuscript: 

(1) Page 3: "they bind calcium delimited to the point of entry at calcium channels", better use "sources" 

We agree with the reviewer. The phrasing on Page 3 has been updated to use 'sources' instead of 'the point of entry at calcium channels' for clarity.

(2) Page 3 "localized supplies of intracellular calcium", I do not like this term, but maybe this is just silly.

We agree with the reviewer. The term 'localized supplies of intracellular calcium' on Page 3 has been revised to “Localized calcium sources”

(3) Regarding the definitions stated by the authors: How do you distinguish between "ensembles" corresponding to "coordinated collection of BK and Cav channels" and "assembly of BK clusters with Cav clusters"? I believe that hetero-clusters is more adequate. The nomenclature does not respond to any consensus in the protein biology field, and I find that it introduces bias more than it helps. I would stick to heteroclusters nomenclature that has been used previously in the field. Moreover, in some discussion sections, the term "ensemble" is used in ways that border on vague, especially when talking about "functional signaling complexes" or "ensembles forming early." It's still acceptable within context but could benefit from clearer language to distinguish ensemble (structural proximity) from complex (functional consequence).

We agree with the reviewer, and we recognize the importance of precise nomenclature and have adopted hetero-clusters instead of ensembles to align with established conventions in the field. This term specifically refers to the spatial organization of BK and Ca_V1.3 channels, while functional complexes denote mechanistic interactions. We have revised sections where ensemble was used ambiguously to ensure clear distinction between structure and function.

The definition of "cluster" is clearly stated early but less emphasized in later quantitative analyses (e.g., particle size discussions in Figure 7). Figure 8 is equally confusing, graphs D and E referring to "BK ensembles" and "Cav ensembles", but "ensembles" should refer to combinations of both channels, whereas these seem to be "clusters". In fact, the Figure legend mentions "clusters".

We agree with the reviewer. Terminology has been revised throughout the manuscript to ensure consistency, with 'clusters' used appropriately in quantitative analyses and figure descriptions.

(4) Methods: how are clusters ("ensembles") analysed from the STORM data? What is the logarithm used for? More info about this is required. Equally, more information and discussion about how colocalization is measured and interpreted in superresolution microscopy are required.

We thank the reviewer for their suggestion, and additional details have been incorporated into the Methods section to clarify how clusters ('ensembles') are analyzed from STORM data, including the role of the logarithm in processing. Furthermore, we have expanded the discussion to provide more information on how colocalization is measured and interpreted in super resolution microscopy. We include the following in the manuscript.

“Direct stochastic optical reconstruction microscopy (dSTORM) images of BK and Ca_V1.3 overexpressed in tsA-201 cells were acquired using an ONI Nanoimager microscope equipped with a 100X oil immersion objective (1.4 NA), an XYZ closed-loop piezo 736 stage, and triple emission channels split at 488, 555, and 640 nm. Samples were imaged at 35°C. For singlemolecule localization microscopy, fixed and stained cells were imaged in GLOX imaging buffer containing 10 mM β-mercaptoethylamine (MEA), 0.56 mg/ml glucose oxidase, 34 μg/ml catalase, and 10% w/v glucose in Tris-HCl buffer. Single-molecule localizations were filtered using NImOS software (v.1.18.3, ONI). Localization maps were exported as TIFF images with a pixel size of 5 nm. Maps were further processed in ImageJ (NIH) by thresholding and binarization to isolate labeled structures. To assess colocalization between the signal from two proteins, binary images were multiplied. Particles smaller than 400 nm² were excluded from the analysis to reflect the spatial resolution limit of STORM imaging (20 nm) and the average size of BK channels. To examine spatial localization preference, binary images of BK were progressively dilated to 20 nm, 40 nm, 60 nm, 80 nm, 100 nm, and 200 nm to expand their spatial representation. These modified images were then multiplied with the Ca_V1.3 channel to quantify colocalization and determine BK occupancy at increasing distances from Ca_V1.3. To ensure consistent comparisons across distance thresholds, data were normalized using the 200 nm measurement as the highest reference value, set to 1”.

(5) Related to Figure 2:

(a) Why use an antibody to label GFP when PH-PLCdelta should be a membrane marker? Where is the GFP in PH-PKC-delta (intracellular, extracellular? Images in Figure 2E are confusing, there is a green intracellular signal.

We thank the reviewer for their feedback. To clarify, GFP is fused to the N-terminus of PH-PLCδ and primarily localizes to the inner plasma membrane via PIP2 binding. Residual intracellular GFP signal may reflect non-membrane-bound fractions or background from anti-GFP immunostaining. We added a paragraph explaining the use of the antibody anti GFP in the Methods section Proximity ligation assay subsection. 

(b) The images in Figure 2 do not help to understand how the authors select the PLA puncta located at the plasma membrane. How do the authors do this? A useful solution would be to indicate in Figure 2 an example of the PLA signals that are considered "membrane signals" compared to another example with "intracellular signals". Perhaps this was intended with the current Figure, but it is not clear.

We agree with the reviewer. We have added a sentence to explain how the number of PLA puncta at the plasma membrane was calculated. 

“We visualized the plasma membrane with a biological sensor tagged with GFP (PHPLCδ-GFP) and then probed it with an antibody against GFP (Figure 2E). By analyzing the GFP signal, we created a mask that represented the plasma membrane. The mask served to distinguish between the PLA puncta located inside the cell and those at the plasma membrane, allowing us to calculate the number of PLA puncta at the plasma membrane”.

(c) Figure 2C: What is the negative control? Apologies if it is described somewhere, but I seem not to find it in the manuscript.

We thank the reviewer for their suggestion. For the negative control in Figure 2C, BK was probed using the primary antibody without co-staining for Ca_V1.3 or other proteins, ensuring specificity and ruling out non-specific antibody binding or background fluorescence. A sentence clarifying the negative control for Figure 2C has been added to the Results section, specifying that BK was probed using the primary antibody without costaining for Ca_V1.3 or other proteins to ensure specificity. 

“To confirm specificity, a negative control was performed by probing only for BK using the primary antibody, ensuring that detected signals were not due to non-specific binding or background fluorescence”.

(d) What is the resolution in z of the images shown in Figure 2? This is relevant for the interpretation of signal localization.

The z-resolution of the images shown in Figure 2 was approximately 270–300 nm, based on the Zeiss Airyscan system’s axial resolution capabilities. Imaging was performed with a step size of 300 nm, ensuring adequate sampling for signal localization while maintaining optimal axial resolution.

“In a different experiment, we analyzed the puncta density for each focal plane of the cell (step size of 300 nm) and compared the puncta at the plasma membrane to the rest of the cell”.

(e) % of total puncta in PM vs inside cell are shown for transfected cells, what is this proportion in INS-1 cells?

This quantification was performed for transfected cells; however, we have not conducted the same analysis in INS-1 cells. Future experiments could address this to determine potential differences in puncta distribution between endogenous and overexpressed conditions.

(6) Related to Figure 3:

(a) Figure 3B: is this antibody labelling or GFP fluorescence? Why do they use GFP antibody labelling, if the marker already has its own fluorescence? This should at least be commented on in the manuscript.

We thank the reviewer for their concern. In Figure 3B, GFP was labeled using an antibody rather than relying on its intrinsic fluorescence. This approach was necessary because GFP fluorescence does not withstand the PLA protocol, resulting in significant fading. Antibody labeling provided stronger signal intensity and improved resolution, ensuring optimal signal-to-noise ratio for accurate analysis.

A clarification regarding the use of GFP antibody labeling in Figure 3B has been added to the Methods section, explaining that intrinsic GFP fluorescence does not endure the PLA protocol, necessitating antibody-based detection for improved signal and resolution.We added the following to the manuscript. 

“For PLA combined with immunostaining, PLA was followed by a secondary antibody incubation with Alexa Fluor-488 at 2 μg/ml for 1 hour at 21˚C. Since GFP fluorescence fades significantly during the PLA protocol, resulting in reduced signal intensity and poor image resolution, GFP was labeled using an antibody rather than relying on its intrinsic fluorescence”.

(b) Why is it relevant to study the ER exit sites? Some explanation should be included in the main text (page 11) for clarification to non-specialized readers. Again, the quantification should be performed on the proportion of clusters/ensembles out of the total number of channels expressed at the ER (or ER exit sites).

We thank the reviewer for their feedback. We have modified this section to include a more detailed explanation of the relevance of ER exit sites to protein trafficking. ER exit sites serve as specialized sorting hubs that regulate the transition of proteins from the ER to the secretory pathway, distinguishing them from the broader ER network, which primarily facilitates protein synthesis and folding. This additional context clarifies why studying ER exit sites provides valuable insights into ensemble trafficking dynamics.

Regarding quantification, our method does not allow for direct measurement of the total number of BK and Ca_V1.3 channels expressed at the ER or ER exit sites. Instead, we focused on the proportion of hetero-clusters localized within these compartments, which provides insight into trafficking pathways despite the limitation in absolute channel quantification. We included the following in the manuscript in the Results section. 

“To determine whether the observed colocalization between BK–Ca_V1.3 hetero-clusters and the ER was not simply due to the extensive spatial coverage of ER labeling, we labeled ER exit sites using Sec16-GFP and probed for hetero-clusters with PLA. This approach enabled us to test whether the hetero-clusters were preferentially localized to ER exit sites, which are specialized trafficking hubs that mediate cargo selection and direct proteins from the ER into the secretory pathway. In contrast to the more expansive ER network, which supports protein synthesis and folding, ER exit sites ensure efficient and selective export of proteins to their target destinations”.

“By quantifying the proportion of BK and Ca_V1.3 hetero-clusters relative to total channel expression at ER exit sites, we found 28 ± 3% colocalization in tsA-201 cells and 11 ± 2% in INS-1 cells (Figure 3F). While the percentage of colocalization between hetero-clusters and the ER or ER exit sites alone cannot be directly compared to infer trafficking dynamics, these findings reinforce the conclusion that hetero-clusters reside within the ER and suggest that BK and Ca_V1.3 channels traffic together through the ER and exit in coordination”.

(7) Related to Figure 4:

A control is included to confirm that the formation of BK-Cav1.3 ensembles is not unspecific. Association with a protein from the Golgi (58K) is tested. Why is this control only done for Golgi? No similar experiment has been performed in the ER. This aspect should be commented on.

We thank the reviewer for their suggestion. We selected the Golgi as a control because it represents the final stage of protein trafficking before proteins reach their functional destinations. If BK and Ca_V1.3 hetero-cluster formation is specific at the Golgi, this suggests that their interaction is maintained throughout earlier trafficking steps, including within the ER. While we did not perform an equivalent control experiment in the ER, the Golgi serves as an effective checkpoint for evaluating specificity within the broader protein transport pathway. We included the following in the manuscript.

“We selected the Golgi as a control because it represents the final stage of protein trafficking, ensuring that hetero-cluster interactions observed at this point reflect specificity maintained throughout earlier trafficking steps, including within the ER”.

(8) How is colocalization measured, eg, in Figure 6? Are the images shown in Figure 6 representative? This aspect would benefit from a clearer description.

We thank the reviewer for their suggestion. A section clarifying colocalization measurement and the representativeness of Figure 6 images has been added to the Methods under Data Analysis. We included the following in the manuscript.

For PLA and RNAscope experiments, we used custom-made macros written in ImageJ. Processing of PLA data included background subtraction. To assess colocalization, fluorescent signals were converted into binary images, and channels were multiplied to identify spatial overlap.

(9) The text should be revised for typographical errors, for example:

(a) Summary "evidence of" (CHECK THIS ONE)

We agree with the reviewer, and we corrected the typographical errors

(b) Table 1, row 3: "enriches" should be "enrich"

We agree with the reviewer. The term 'enriches' in Table 1, row 3 has been corrected to 'enrich'.

(c) Figure 2B "priximity"

We agree with the reviewer. The typographical errors in Figure 2B has been corrected from 'priximity' to 'proximity'.

(d) Legend of Figure 7 (C) "size of BK and Cav1.3 channels". Does this correspond to individual channels or clusters?

We agree with the reviewer. The legend of Figure 7C has been clarified to indicate that 'size of BK and Cav1.3 channels' refers to clusters rather than individual channels.

(e) Methods: In the RNASCOPE section, "Fig.4-supp1" should be "Fig. 5-supp1"

(f) Page 15, Figure 5B is cited, should be Figure 6B

We agree with the reviewer. The reference in the RNASCOPE section has been updated from 'Fig.4-supp1' to 'Fig. 5-supp1,' and the citation on Page 15 has been corrected from Figure 5B to Figure 6B.

Reviewer #2 (Recommendations for the authors):

(1) The abstract could be more accessible for a wider readership with improved flow.

We thank the reviewer for their suggestion. We modified the summary as follows to provide a more coherent flow for a wider readership. 

“Calcium binding to BK channels lowers BK activation threshold, substantiating functional coupling with calcium-permeable channels. This coupling requires close proximity between different channel types, and the formation of BK–Ca_V1.3 hetero-clusters at nanometer distances exemplifies this unique organization. To investigate the structural basis of this interaction, we tested the hypothesis that BK and Ca_V1.3 channels assemble before their insertion into the plasma membrane. Our approach incorporated four strategies: (1) detecting interactions between BK and Ca_V1.3 proteins inside the cell, (2) identifying membrane compartments where intracellular hetero-clusters reside, (3) measuring the proximity of their mRNAs, and (4) assessing protein interactions at the plasma membrane during early translation. These analyses revealed that a subset of BK and Ca_V1.3 transcripts are spatially close in micro-translational complexes, and their newly synthesized proteins associate within the endoplasmic reticulum (ER) and Golgi. Comparisons with other proteins, transcripts, and randomized localization models support the conclusion that BK and Ca_V1.3 hetero-clusters form before their insertion at the plasma membrane”.

(2) Figure 2B - spelling of proximity.

We agree with the reviewer. The typographical errors in Figure 2B has been corrected from 'priximity' to 'proximity'.

Reviewer #3 (Recommendations for the authors):

Minor issues to improve the manuscript:

(1) For completeness, the authors should include a few sentences and appropriate references in the Introduction to mention that BK channels are regulated by auxiliary subunits.

We agree with the reviewer. We have revised the Introduction to include a brief discussion of how BK channel function is modulated by auxiliary subunits and provided appropriate references to ensure completeness. These additions highlight the broader regulatory mechanisms governing BK channel activity, complementing the focus of our study. We included the following in the manuscript. 

“Additionally, BK channels are modulated by auxiliary subunits, which fine-tune BK channel gating properties to adapt to different physiological conditions. β and γ subunits regulate BK channel kinetics, altering voltage sensitivity and calcium responsiveness [18]. These interactions ensure precise control over channel activity, allowing BK channels to integrate voltage and calcium signals dynamically in various cell types. Here, we focus on the selective assembly of BK channels with Ca_V1.3 and do not evaluate the contributions of auxiliary subunits to BK channel organization.”

(2) Insert a space between 'homeostasis' and the square bracket at the end of the Introduction's second paragraph.

We agree with the reviewer. A space has been inserted between 'homeostasis' and the square bracket in the second paragraph of the Introduction for clarity.

(3) The images presented in Figures 2-5 should be increased in size (if permitted by the Journal) to allow the reader to clearly see the puncta in the fluorescent images. This would necessitate reconfiguring the figures into perhaps a full A4 page per figure, but I think the quality of the images presented really do deserve to "be seen". For example, Panels A & B could be at the top of Figure 2, with C & D presented below them. However, I'll leave it up to the authors to decide on the most aesthetically pleasing way to show these.

We agree with the reviewer. We have increased the size of Figures 2–8 to enhance the visibility of fluorescent puncta, as suggested. To accommodate this, we reorganized the panel layout for each figure—for example, in Figure 2, Panels A and B are now placed above Panels C and D to support a more intuitive and aesthetically coherent presentation. We believe this revised configuration highlights the image quality and improves readability while conforming to journal layout constraints.

(4) I think that some of the sentences could be "toned down"

(a) eg, in the first paragraph below Figure 2, the authors state "that 46(plus minus)3% of the puncta were localised on intracellular membranes" when, at that stage, no data had been presented to confirm this. I think changing it to "that 46(plus minus)3% of the puncta were localised intracellularly" would be more precise.

(b) Similarly, please consider replacing the wording of "get together at membranes inside the cell" to "co-localise intracellularly".

(c) In the paragraph just before Figure 5, the authors mention that "the abundance of KCNMA1 correlated more with the abundance of CACNA1D than ... with GAPDH." Although this is technically correct, the R2 value was 0.22, which is exceptionally poor. I don't think that the paper is strengthened by sentences such as this, and perhaps the authors might tone this down to reflect this.

(d) The authors clearly demonstrate in Figure 8 that a significant number of BK channels can traffic to the membrane in the absence of Cav1.3. Irrespective of the differences in transcription/trafficking time between the two channel types, the authors should insert a few lines into their discussion to take this finding into account.

We appreciate the reviewer’s feedback regarding the clarity and precision of our phrasing.

Our responses for each point are below.

(a) We have modified the statement in the first paragraph below Figure 2, changing '46 ± 3% of the puncta were localized on intracellular membranes' to '46 ± 3% of the puncta were localized ‘intracellularly’ to ensure accuracy in the absence of explicit data confirming membrane association.

(b) Similarly, we have replaced 'get together at membranes inside the cell' with 'colocalize intracellularly' to maintain clarity and avoid unintended implications. 

(c) Regarding the correlation between KCNMA1 and CACNA1D abundance, we recognize that the R² value of 0.22 is relatively low. To reflect this appropriately, we have revised the phrasing to indicate that while a correlation exists, it is modest. We added the following to the manuscript. 

“Interestingly, the abundance of KCNMA1 transcripts correlated more with the abundance of CACNA1D transcripts than with the abundance of GAPDH, a standard housekeeping gene, though with a modest R² value.”

(d) To incorporate the findings from Figure 8, we have added discussion acknowledging that a substantial number of BK channels traffic to the membrane independently of Ca_V1.3. This addition provides context for potential trafficking mechanisms that operate separately from ensemble formation.

(5) For clarity, please insert the word "total" in the paragraph after Figure 3 "..."63{plus minus}3% versus 50%{plus minus}6% of total PLA puncta were localised at the ER". I know this is explicitly stated later in the manuscript, but I think it needs to be clarified earlier.

We agree with the reviewer. The word 'total' has been inserted in the paragraph following Figure 3 to clarify the percentage of PLA puncta localized at the ER earlier in the manuscript

(6) In the discussion, I think an additional (short) paragraph needs to be included to clarify to the reader why the % "colocalization between ensembles and the ER or the ER exit sites can't be compared or used to understand the dynamics of the ensembles". This may permit the authors to remove the last sentence of the paragraph just before the results section, "BK and Cav1.3 ensembles go through the Golgi."

We thank the reviewer for their suggestion. We have added a short paragraph in the discussion to clarify why colocalization percentages between ensembles and the ER or ER exit sites cannot be compared to infer ensemble dynamics. This allowed us to remove the final sentence of the paragraph preceding the results section ('BK and Cav1.3 ensembles go through the Golgi).

(7) In the paragraph after Figure 6, Figure 5B is inadvertently referred to. Please correct this to Figure 6B.

We agree with the reviewer. The reference to Figure 5B in the paragraph after Figure 6 has been corrected to Figure 6B.

(8) In the discussion under "mRNA co-localisation and Protein Trafficking", please insert a relevant reference illustrating that "disruption in mRNA localization... can lead to ion channel mislocalization".

We agree with the reviewer. We have inserted a relevant reference under 'mRNA Colocalization and Protein Trafficking' to illustrate that disruption in mRNA localization can lead to ion channel mislocalization.

(9) The supplementary Figures appear to be incorrectly numbered. Please correct and also ensure that they are correctly referred to in the text.

We agree with the reviewer. The numbering of the supplementary figures has been corrected, and all references to them in the text have been updated accordingly.

(10) The final panels of the currently labelled Figure 5-Supplementary 2 need to have labels A-F included on the image.

We agree with the reviewer. Labels A-F have been added to the final panels of Figure 5-Supplementary 2.

References

(1) Shah, K.R., X. Guan, and J. Yan, Structural and Functional Coupling of Calcium-Activated BK Channels and Calcium-Permeable Channels Within Nanodomain Signaling Complexes. Frontiers in Physiology, 2022. Volume 12 - 2021.

(2) Chen, A.L., et al., Calcium-Activated Big-Conductance (BK) Potassium Channels Traffic through Nuclear Envelopes into Kinocilia in Ray Electrosensory Cells. Cells, 2023. 12(17): p. 2125.

(3) Berkefeld, H., B. Fakler, and U. Schulte, Ca2+-activated K+ channels: from protein complexes to function. Physiol Rev, 2010. 90(4): p. 1437-59.

(4) Loane, D.J., P.A. Lima, and N.V. Marrion, Co-assembly of N-type Ca2+ and BK channels underlies functional coupling in rat brain. J Cell Sci, 2007. 120(Pt 6): p. 98595.

(5) Boncompain, G. and F. Perez, The many routes of Golgi-dependent trafficking. Histochemistry and Cell Biology, 2013. 140(3): p. 251-260.

(6) Kurokawa, K. and A. Nakano, The ER exit sites are specialized ER zones for the transport of cargo proteins from the ER to the Golgi apparatus. The Journal of Biochemistry, 2019. 165(2): p. 109-114.

(7) Chen, G., et al., BK channel modulation by positively charged peptides and auxiliary γ subunits mediated by the Ca2+-bowl site. Journal of General Physiology, 2023. 155(6).

self

from https://jonudell.info/h/facet/?max=50&exactTagSearch=true&expanded=true&addQuoteContext=true&any=groan%2Bzone

Sleep-time Compute: Beyond Inference Scaling at Test-time

Core Concept

• Sleep-time compute allows models to "think" offline about contexts before queries are presented, reducing test-time compute requirements by ~5× on benchmark tasks

"by anticipating what queries users might ask and pre-computing useful quantities, we can significantly reduce the compute requirements at test-time"

• The approach works by processing context c during idle time to create an enhanced representation c', which is then used at test-time: S(c) → c', followed by Tb(q, c') → a

"In practice, this is achieved by prompting the model to generate a new context consisting of inferences about the existing context, which may be potentially useful for answering test-time queries"

Key Results

• Performance improvements: Sleep-time compute reduces test-time compute needed to achieve same accuracy by ~5× on Stateful GSM-Symbolic and Stateful AIME

"Sleep-time compute produces a pareto improvement in the test-time compute vs. accuracy curve, reducing the test-time compute needed to achieve the same accuracy by ∼ 5×"

• Scaling benefits: By scaling up sleep-time compute, accuracy increases by up to 13% on Stateful GSM-Symbolic and 18% on Stateful AIME

• Cost amortization: When multiple queries share the same context, average cost per query decreases by 2.5×

"By amortizing sleep-time compute across related queries about the same context using Multi-Query GSM-Symbolic, we can decrease the average cost per query by 2.5×"

Datasets Introduced

• Stateful GSM-Symbolic: Modified from GSM-Symbolic (P1: 5000 examples, P2: 2500 examples) by splitting problems into context and question

"We introduce two datasets to study applying sleep-time compute in stateful settings, Stateful GSM-Symbolic, and Stateful AIME – by splitting the existing problems in these datasets into a context and a question"

• Stateful AIME: Contains 60 questions from AIME 2024 and 2025, split into context and query components

• Multi-Query GSM-Symbolic: Extends GSM-Symbolic with multiple related queries per context (P1: 12,043 questions, 1,095 contexts; P2: 5,497 questions, 500 contexts)

• SWE-Features: Software engineering benchmark for multi-file feature implementation tasks (33 examples from Aider-AI/aider and ComfyUI repositories)

Models Evaluated

• Non-reasoning models: GPT-4o-mini and GPT-4o on GSM-Symbolic tasks

• Reasoning models: OpenAI's o1, o3-mini, Anthropic's Claude Sonnet 3.7 Extended Thinking, and DeepSeek-R1 on AIME tasks

• Test-time compute scaled both sequentially (varying verbosity/reasoning effort) and in parallel (pass@k sampling)

Effectiveness Analysis

• Query predictability correlation: Sleep-time compute is most effective when queries are predictable from context

"sleep-time compute is more effective in settings where the query is more easily predictable from the context"

• Predictability measured using log-probability of question given context under Llama2-70B base model

• Accuracy gap between sleep-time and test-time compute widens for more predictable questions (binned analysis across 5 quantiles)

Implementation Details

• Sleep-time compute implemented via function calling with two functions: - rethink_memory: Takes new string input and replaces current context - finish_rethinking: Terminates sleep-time compute process

• Models allowed up to 10 calls to rethink_memory function

• Cost modeling assumes test-time tokens are 10× more expensive than sleep-time tokens (t=10) due to latency optimization

"Since at test-time, there are strict latency constraints, and latency optimized inference can be roughly 10× more expensive, we model the total cost of inference between both sleep-time and test-time, by up-weighing the cost of test-time tokens"

Comparison to Baselines

• Pass@k parallel scaling: Sleep-time compute consistently outperforms pass@k at same test-time token budget

"sleep-time compute consistently outperforms pass@k parallel scaling at the same test-time token budget, demonstrating that sleep-time compute can be a more effective way to scale inference-time compute than standard parallel test-time scaling"

• Context-only baseline: Sleep-time compute significantly outperforms models that only receive context and must guess the question, demonstrating questions are not trivially predictable

SWE-Features Case Study

• At lower test-time budgets, sleep-time compute achieves ~1.5× reduction in test-time tokens with higher F1 scores

• At higher budgets, standard test-time compute performs better, with higher precision but comparable recall

• Hypothesis: sleep-time compute explores more files, leading to editing more files and slightly lower precision

Related Work & Context

• Builds on recent test-time scaling approaches: sequential (OpenAI o1, DeepSeek-R1) and parallel (pass@k, best-of-N)

• Connection to speculative decoding (Leviathan et al., 2023): Both speculate on user queries, but sleep-time compute uses generated tokens as input regardless of actual query

• Connection to pre-computation in systems: Similar to memory caches (Smith, 1982) and data cubes for OLAP workloads (Gray et al., 1997)

• Resembles representation learning but operates in natural language space rather than parameter/activation space

Limitations & Future Directions

• Sleep-time compute less effective when queries are unpredictable or unrelated to context

• Current approach assumes simple two-phase interaction (sleep-time and test-time), but real-world scenarios involve multiple interaction rounds

• Future work: Optimal allocation of compute between sleep-time and test-time based on query predictability

• Potential application to synthetic data generation at scale for pretraining

Authors & Affiliation

Kevin Lin, Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica, Joseph E. Gonzalez (Letta & UC Berkeley)

Code and data: https://github.com/letta-ai/sleep-time-compute

Scaling Context Requires Rethinking Attention

Core Thesis

• Neither transformers nor sub-quadratic architectures are well-suited for long-context training

  "the cost of processing the context is too expensive in the former, too inexpensive in the latter"

• Power attention introduced as solution: A linear-cost sequence modeling architecture with independently adjustable state size > "an architectural layer for linear-cost sequence modeling whose state size can be adjusted independently of parameters"

Three Requirements for Long-Context Architectures

1. Balanced Weight-to-State Ratio (WSFR)

• Weight-state FLOP ratio should approach 1:1 for compute-optimal models

  "for compute-optimal models, the WSFR should be somewhat close to 1:1"

• Exponential attention becomes unbalanced at long contexts

• At 65,536 context: WSFR is 1:8

• At 1,000,000 context: WSFR is 1:125

  "exponential attention is balanced for intermediate context lengths, but unbalanced for long context lengths, where it does far more state FLOPs than weight FLOPs"

• Linear attention remains unbalanced at all context lengths

• WSFR stays at 30:1 regardless of context length

  "Linear attention...is unbalanced at all context lengths in the opposite direction: far more weight FLOPs than state FLOPs"

2. Hardware-Aware Implementation

• Must admit efficient implementation on tensor cores
• Power attention achieves 8.6x faster throughput than Flash Attention at 64k context (head size 32)
• 3.3x speedup at head size 64

3. Strong In-Context Learning (ICL)

• Large state size improves ICL performance

  "state scaling improves performance"

• Windowed attention fails ICL beyond window size

  "no in-context learning occurs beyond 100 tokens for window-32 attention"

• Linear attention maintains ICL across entire sequence

  "linear attention...demonstrate consistent in-context learning across the entire sequence"

Power Attention Technical Details

Mathematical Foundation

• Power attention formula: Uses p-th power instead of exponential

  "attnᵖₚₒw(Q, K, V)ᵢ = Σⱼ₌₁ⁱ (QᵢᵀKⱼ)ᵖVⱼ"

• Symmetric power expansion (SPOW) reduces state size vs tensor power (TPOW)

• At p=2, d=64: SPOW uses 2,080 dimensions vs TPOW's 4,096 (49% savings)
• At p=4, d=64: 95% size reduction

  "SPOWₚ is a state expansion that increases the state size by a factor of (ᵈ⁺ᵖ⁻¹ₚ)/d without introducing any parameters"

Implementation Innovation

• Fused expand-MMA kernel: Expands tiles on-the-fly during matrix multiplication

  "a matrix multiplication where the tiles of one operand are expanded on-the-fly"

• Tiled symmetric power expansion (TSPOW): Interpolates between TPOW and SPOW

• Provides GPU-friendly structure while reducing data duplication

• Optimal tile size: d-tile = 8 for p=2, d-tile = 4 for p=3

• Chunked form enables practical efficiency

  "The chunked form interpolates between the recurrent form and the attention form, capturing benefits of both"

• Cost: O(tDv + tcd) where c is chunk size

Experimental Results

In-Context Learning Performance

• Power attention dominates windowed attention at equal state sizes across all context lengths
• All scaling axes improve ICL: gradient updates, batch size, parameter count, context length

  "In all cases, the ICL curve becomes steeper as we scale the respective axis"

Long-Context Training (65,536 tokens)

• Power attention (p=2) outperforms both exponential and linear attention in loss-per-FLOP
• RWKV with power attention shows near-zero ICL benefit beyond 2,000 tokens
• Power attention enables RWKV to ICL "nearly as well as exponential attention"

Compute-Optimal Under Latency Constraints

• When inference latency constrains parameter count and state size:
• Window-1k attention: loss 1.638
• Standard attention: loss 1.631
• Power attention (p=2): loss 1.613 (best)

Dataset and Experimental Setup

LongCrawl64

• 6.66M documents, each 65,536 tokens (435B total tokens)
• Sourced from Common Crawl, filtered for long sequences
• Critical for ICL research

  "Most sequences in OpenWebText have length less than 1k"

Architectures Tested

• Base architectures: GPT-2, RWKV (RWKV7), GLA, RetNet
• Attention variants: Exponential, linear, windowed, power (p=2)
• Training: LongCrawl64, AdamW, bf16, learning rate 3e-4 with warmup and cosine decay

Key Limitations and Future Work

Current Limitations

1. Experiments limited to natural language NLL - no other domains/modalities tested
2. Compute-optimal context grows slowly in natural language

   "autoregressive prediction of natural language is largely dominated by short-context dependencies"

3. p=2 only - normalization requires positive inner products (even powers only)
4. Triton implementation - not yet optimized to CUDA level

Future Directions

• Explore domains with long-term dependencies: chain-of-thought reasoning, audio, video
• Scaling laws research for state size, context size, and ICL
• CUDA implementation for further speedups beyond current Triton kernels
• Alternative normalization to support odd powers
• Comprehensive comparison to hybrid models, sparse attention, MQA, latent attention

Key References and Tools

Implementations

• Open-source kernels: https://github.com/m-a-n-i-f-e-s-t/power-attention
• FLA codebase: Flash Linear Attention [Yang and Zhang, 2024]
• PyTorch framework [Paszke et al., 2019]

Related Techniques

• Flash Attention [Dao, 2023]: Operator fusion to avoid materializing attention matrix
• Linear attention [Katharopoulos et al., 2020]: Enables recurrent formulation
• Gating [Lin et al., 2025]: Learned mechanism to avoid attending to old data
• Sliding window attention [Child et al., 2019]: Truncates KV cache

Key Papers

• Transformers [Vaswani et al., 2023]
• Mamba [Gu and Dao, 2024]: Modern RNN architecture
• RWKV [Peng et al., 2023]: Reinventing RNNs for transformer era
• Scaling laws [Kaplan et al., 2020]

Technical Contributions

1. Framework for evaluating long-context architectures (balance, efficiency, ICL)
2. Power attention architecture with parameter-free state size adjustment
3. Symmetric power expansion theory and implementation
4. Hardware-efficient kernels with operation fusion
5. Empirical validation on 435B token dataset

The Prompt Report: A Systematic Survey of Prompting Techniques

Overview & Scope

• Comprehensive taxonomy: "We establish a structured understanding of prompt engineering by assembling a taxonomy of prompting techniques and analyzing their applications. We present a detailed vocabulary of 33 vocabulary terms, a taxonomy of 58 LLM prompting techniques, and 40 techniques for other modalities."

• Scope limitation: "We limit our study to focus on prefix prompts rather than cloze prompts, because modern LLM transformer architectures widely employ prefix prompts"

• Focus on hard prompts: "Additionally, we refined our focus to hard (discrete) prompts rather than soft (continuous) prompts and leave out papers that make use of techniques using gradient-based updates (i.e. fine-tuning). Hard prompts contain only tokens (vectors) that correspond to words in the model's vocabulary"

Key Definitions

Prompt & Prompting

• Prompt definition: "A prompt is an input to a Generative AI model, that is used to guide its output"

• Prompt template: "A prompt template is a function that contains one or more variables which will be replaced by some media (usually text) to create a prompt"

• Prompting: "Prompting is the process of providing a prompt to a GenAI, which then generates a response"

Prompt Engineering

• Consolidated definition: "Prompt engineering is the iterative process of developing a prompt by modifying or changing the prompting technique that you are using"

• Process description: "The Prompt Engineering Process consists of three repeated steps 1) performing inference on a dataset 2) evaluating performance and 3) modifying the prompt template"

Core Prompt Components

Essential Elements

• Directive: "Many prompts issue a directive in the form of an instruction or question. This is the core intent of the prompt"

• Examples/Exemplars: "Examples, also known as exemplars or shots, act as demonstrations that guide the GenAI to accomplish a task"

• Output formatting: "It is often desirable for the GenAI to output information in certain formats, for example, CSV, Markdown, XML, or even custom formats"

• Style instructions: "Style instructions are a type of output formatting used to modify the output stylistically rather than structurally"

• Role/Persona: "A Role, also known as a persona, is a frequently discussed component that can improve writing and style text"

Systematic Review Methodology

PRISMA Process

• Approach: "We conducted a machine-assisted systematic review grounded in the PRISMA process to identify 58 different text-based prompting techniques"

• Data sources: "Our main data sources were arXiv, Semantic Scholar, and ACL. We query these databases with a list of 44 keywords narrowly related to prompting and prompt engineering"

• Pipeline: "We retrieve papers from arXiv based on a simple set of keywords and boolean rules. Then, human annotators label a sample of 1,661 articles"

• Inter-rater reliability: "A set of 300 articles are reviewed independently by two annotators, with 92% agreement (Krippendorff's α = Cohen's κ = 81%)"

• Final dataset: "The combined human and LLM annotations generate a final set of 1,565 papers"

Major Technique Categories

In-Context Learning (ICL)

• Definition: "ICL refers to the ability of GenAIs to learn skills and tasks by providing them with exemplars and or relevant instructions within the prompt, without the need for weight updates/retraining"

• Few-Shot Prompting: "Brown et al. (2020) is the paradigm seen in Figure 2.4, where the GenAI learns to complete a task with only a few examples (exemplars)"

Design Decisions for Few-Shot Prompting

• Exemplar quantity: "Increasing the quantity of exemplars in the prompt generally improves model performance, particularly in larger models. However, in some cases, the benefits may diminish beyond 20 exemplars"

• Exemplar ordering: "The order of exemplars affects model behavior. On some tasks, exemplar order can cause accuracy to vary from sub-50% to 90%+"

• Label distribution impact: "As in traditional supervised machine learning, the distribution of exemplar labels in the prompt affects behavior"

• Label quality: "Despite the general benefit of multiple exemplars, the necessity of strictly valid demonstrations is unclear. Some work suggests that the accuracy of labels is irrelevant—providing models with exemplars with incorrect labels may not negatively diminish performance"

• Exemplar format: "The formatting of exemplars also affects performance. One of the most common formats is 'Q: {input}, A: {label}', but the optimal format may vary across tasks"

• Exemplar similarity: "Selecting exemplars that are similar to the test sample is generally beneficial for performance. However, in some cases, selecting more diverse exemplars can improve performance"

Few-Shot Techniques

• K-Nearest Neighbor (KNN): "Liu et al. (2021) is part of a family of algorithms that selects exemplars similar to test samples to boost performance"

• Vote-K: "Su et al. (2022) is another method to select similar exemplars to the test sample... Vote-K also ensures that newly added exemplars are sufficiently different than existing ones to increase diversity"

• Self-Generated In-Context Learning (SG-ICL): "Kim et al. (2022) leverages a GenAI to automatically generate exemplars. While better than zero-shot scenarios when training data is unavailable, the generated samples are not as effective as actual data"

• Prompt Mining: "Jiang et al. (2020) is the process of discovering optimal 'middle words' in prompts through large corpus analysis"

Zero-Shot Techniques

• Role Prompting: "Wang et al. (2023j); Zheng et al. (2023d), also known as persona prompting, assigns a specific role to the GenAI in the prompt"

• Style Prompting: "Lu et al. (2023a) involves specifying the desired style, tone, or genre in the prompt to shape the output"

• Emotion Prompting: "Li et al. (2023a) incorporates phrases of psychological relevance to humans (e.g., 'This is important to my career') into the prompt, which may lead to improved LLM performance"

• System 2 Attention (S2A): "Weston and Sukhbaatar (2023) first asks an LLM to rewrite the prompt and remove any information unrelated to the question therein"

• Rephrase and Respond (RaR): "Deng et al. (2023) instructs the LLM to rephrase and expand the question before generating the final answer"

• Re-reading (RE2): "Xu et al. (2023) adds the phrase 'Read the question again:' to the prompt in addition to repeating the question"

• Self-Ask: "Press et al. (2022) prompts LLMs to first decide if they need to ask follow up questions for a given prompt"

Thought Generation

• Chain-of-Thought (CoT): "Wei et al. (2022b) leverages few-shot prompting to encourage the LLM to express its thought process before delivering its final answer"

• Zero-Shot-CoT: "The most straightforward version of CoT contains zero exemplars. It involves appending a thought inducing phrase like 'Let's think step by step.' to the prompt"

• Step-Back Prompting: "Zheng et al. (2023c) is a modification of CoT where the LLM is first asked a generic, high-level question about relevant concepts or facts before delving into reasoning"

• Thread-of-Thought (ThoT): "Zhou et al. (2023) consists of an improved thought inducer for CoT reasoning. Instead of 'Let's think step by step,' it uses 'Walk me through this context in manageable parts step by step, summarizing and analyzing as we go.'"

• Tabular Chain-of-Thought (Tab-CoT): "Jin and Lu (2023) consists of a Zero-Shot CoT prompt that makes the LLM output reasoning as a markdown table"

Few-Shot CoT Variants

• Contrastive CoT: "Chia et al. (2023) adds both exemplars with incorrect and correct explanations to the CoT prompt in order to show the LLM how not to reason"

• Complexity-based Prompting: "Fu et al. (2023b) involves two major modifications to CoT. First, it selects complex examples for annotation and inclusion in the prompt... Second, during inference, it samples multiple reasoning chains"

• Active Prompting: "Diao et al. (2023) starts with some training questions/exemplars, asks the LLM to solve them, then calculates uncertainty (disagreement in this case) and asks human annotators to rewrite the exemplars with highest uncertainty"

• Memory-of-Thought: "Li and Qiu (2023b) leverage unlabeled training exemplars to build Few-Shot CoT prompts at test time"

• Automatic Chain-of-Thought (Auto-CoT): "Zhang et al. (2022b) uses Wei et al. (2022b)'s Zero-Shot prompt to automatically generate chains of thought. These are then used to build a Few-Shot CoT prompt"

Decomposition

• Least-to-Most Prompting: "Zhou et al. (2022a) starts by prompting a LLM to break a given problem into sub-problems without solving them. Then, it solves them sequentially, appending model responses to the prompt each time"

• Decomposed Prompting (DECOMP): "Khot et al. (2022) Few-Shot prompts a LLM to show it how to use certain functions. These might include things like string splitting or internet searching"

• Plan-and-Solve Prompting: "Wang et al. (2023f) consists of an improved Zero-Shot CoT prompt, 'Let's first understand the problem and devise a plan to solve it. Then, let's carry out the plan and solve the problem step by step'"

• Tree-of-Thought (ToT): "Yao et al. (2023b), also known as Tree of Thoughts, creates a tree-like search problem by starting with an initial problem then generating multiple possible steps in the form of thoughts"

• Program-of-Thoughts: "Chen et al. (2023d) uses LLMs like Codex to generate programming code as reasoning steps. A code interpreter executes these steps to obtain the final answer"

• Skeleton-of-Thought: "Ning et al. (2023) focuses on accelerating answer speed through parallelization. Given a problem, it prompts an LLM to create a skeleton of the answer"

Ensembling

• Demonstration Ensembling (DENSE): "Khalifa et al. (2023) creates multiple few-shot prompts, each containing a distinct subset of exemplars from the training set. Next, it aggregates over their outputs"

• Self-Consistency: "Wang et al. (2022) is based on the intuition that multiple different reasoning paths can lead to the same answer. This method first prompts the LLM multiple times to perform CoT, crucially with a non-zero temperature"

• Universal Self-Consistency: "Chen et al. (2023e) is similar to Self-Consistency except that rather than selecting the majority response by programmatically counting how often it occurs, it inserts all outputs into a prompt template"

• DiVeRSe: "Li et al. (2023i) creates multiple prompts for a given problem then performs Self-Consistency for each, generating multiple reasoning paths"

• Prompt Paraphrasing: "Jiang et al. (2020) transforms an original prompt by changing some of the wording, while still maintaining the overall meaning"

Self-Criticism

• Self-Calibration: "Kadavath et al. (2022) first prompts an LLM to answer a question. Then, it builds a new prompt that includes the question, the LLM's answer, and an additional instruction asking whether the answer is correct"

• Self-Refine: "Madaan et al. (2023) is an iterative framework where, given an initial answer from the LLM, it prompts the same LLM to provide feedback on the answer, and then prompts the LLM to improve the answer based on the feedback"

• Self-Verification: "Weng et al. (2022) generates multiple candidate solutions with Chain-of-Thought (CoT). It then scores each solution by masking certain parts of the original question"

• Chain-of-Verification (COVE): "Dhuliawala et al. (2023) first uses an LLM to generate an answer to a given question. Then, it creates a list of related questions that would help verify the correctness of the answer"

Prompt Engineering Automation

Meta Prompting

• Definition: "Meta Prompting is the process of prompting a LLM to generate or improve a prompt or prompt template"

Automated Techniques

• AutoPrompt: "Shin et al. (2020b) uses a frozen LLM as well as a prompt template that includes some 'trigger tokens', whose values are updated via backpropagation at training time"

• Automatic Prompt Engineer (APE): "Zhou et al. (2022b) uses a set of exemplars to generate a Zero-Shot instruction prompt. It generates multiple possible prompts, scores them, then creates variations of the best ones"

• Gradientfree Instructional Prompt Search (GrIPS): "Prasad et al. (2023) is similar to APE, but uses a more complex set of operations including deletion, addition, swapping, and paraphrasing"

• RLPrompt: "Deng et al. (2022) uses a frozen LLM with an unfrozen module added. It uses this LLM to generate prompt templates, scores the templates on a dataset, and updates the unfrozen module using Soft Q-Learning"

Answer Engineering

Core Concept

• Definition: "Answer engineering is the iterative process of developing or selecting among algorithms that extract precise answers from LLM outputs"

Three Design Decisions

• Answer Shape: "The shape of an answer is its physical format. For example, it could be a token, span of tokens, or even an image or video"

• Answer Space: "The space of an answer is the domain of values that its structure may contain. This may simply be the space of all tokens, or in a binary labeling task, could just be two possible tokens"

• Answer Extractor: "In cases where it is impossible to entirely control the answer space... a rule can be defined to extract the final answer. This rule is often a simple function (e.g. a regular expression)"

Extraction Methods

• Verbalizer: "Often used in labeling tasks, a verbalizer maps a token, span, or other type of output to a label and vice-versa (injective)"

• Regex: "Regexes are often used to extract answers. They are usually used to search for the first instance of a label"

• Separate LLM: "Sometimes outputs are so complicated that regexes won't work consistently. In this case, it can be useful to have a separate LLM evaluate the output and extract an answer"

Multilingual Prompting

Core Challenges

• Performance disparity: "State-of-the-art GenAIs have often been predominately trained with English dataset, leading to a notable disparity in the output quality in languages other than English, particularly low-resource languages"

Key Techniques

• Translate First Prompting: "Shi et al. (2022) is perhaps the simplest strategy and first translates non-English input examples into English"

• Cross-Lingual Thought (XLT): "Huang et al. (2023a) utilizes a prompt template composed of six separate instructions, including role assignment, cross-lingual thinking, and CoT"

• Cross-Lingual Self Consistent Prompting (CLSP): "Qin et al. (2023a) introduces an ensemble technique that constructs reasoning paths in different languages to answer the same question"

Prompt Language Selection

• English advantage: "Constructing the prompt template in English is often more effective than in the task language for multilingual tasks. This is likely due to the predominance of English data during LLM pre-training"

• Native language rationale: "In contrast, many multilingual prompting benchmarks such as BUFFET or LongBench use task language prompts for language-specific use cases"

Machine Translation Techniques

• Multi-Aspect Prompting and Selection (MAPS): "He et al. (2023b) mimics the human translation process, which involves multiple preparatory steps to ensure high-quality output"

• Chain-of-Dictionary (CoD): "Lu et al. (2023b) first extracts words from the source phrase, then makes a list of their meanings in multiple languages, automatically via retrieval from a dictionary"

• Interactive-Chain-Prompting (ICP): "Pilault et al. (2023) deals with potential ambiguities in translation by first asking the GenAI to generate sub-questions about any ambiguities in the phrase to be translated"

Multimodal Prompting

Image Prompting

• Prompt Modifiers: "are simply words appended to a prompt to change the resultant image. Components such as Medium (e.g. 'on canvas') or Lighting (e.g. 'a well lit scene') are often used"

• Negative Prompting: "allows users to numerically weight certain terms in the prompt so that the model considers them more/less heavily than others"

Multimodal ICL

• Paired-Image Prompting: "shows the model two images: one before and one after some transformation. Then, present the model with a new image for which it will perform the demonstrated conversion"

• Image-as-Text Prompting: "Hakimov and Schlangen (2023) generates a textual description of an image. This allows for the easy inclusion of the image (or multiple images) in a text-based prompt"

Multimodal CoT

• Duty Distinct Chain-of-Thought (DDCoT): "Zheng et al. (2023b) extends Least-to-Most prompting to the multimodal setting, creating subquestions, then solving them and combining the answers"

• Chain-of-Images (CoI): "Meng et al. (2023) is a multimodal extension of Chain-of-Thought prompting, that generates images as part of its thought process"

Other Modalities

• Audio: "Experiments with audio ICL have generated mixed results, with some open source audio models failing to perform ICL. However, other results do show an ICL ability in audio models"

• Video: "Prompting has also been extended to the video modality, for use in text-to-video generation, video editing, and video-to-text generation"

• 3D: "Prompting can also be used in 3D modalities, for example in 3D object synthesis, 3D surface texturing, and 4D scene generation"

Agents

Definition

• Agent concept: "In the context of GenAI, we define agents to be GenAI systems that serve a user's goals via actions that engage with systems outside the GenAI itself"

Tool Use Agents

• Modular Reasoning, Knowledge, and Language (MRKL) System: "Karpas et al. (2022) is one of the simplest formulations of an agent. It contains a LLM router providing access to multiple tools"

• Self-Correcting with Tool-Interactive Critiquing (CRITIC): "Gou et al. (2024a) first generates a response to the prompt, with no external calls. Then, the same LLM criticizes this response for possible errors"

Code-Generation Agents

• Program-aided Language Model (PAL): "Gao et al. (2023b) translates a problem directly into code, which is sent to a Python interpreter to generate an answer"

• Tool-Integrated Reasoning Agent (ToRA): "Gou et al. (2024b) is similar to PAL, but instead of a single code generation step, it interleaves code and reasoning steps for as long as necessary"

Observation-Based Agents

• Reasoning and Acting (ReAct): "Yao et al. (2022) generates a thought, takes an action, and receives an observation (and repeats this process) when given a problem to solve"

• Reflexion: "Shinn et al. (2023) builds on ReAct, adding a layer of introspection. It obtains a trajectory of actions and observations, then is given an evaluation of success/failure"

Lifelong Learning

• Voyager: "Wang et al. (2023a) is composed of three parts. First, it proposes tasks for itself to complete in order to learn more about the world. Second, it generates code to execute these actions. Finally, it saves these actions to be retrieved later"

• Ghost in the Minecraft (GITM): "Zhu et al. (2023) starts with an arbitrary goal, breaks it down into subgoals recursively, then iteratively plans and executes actions by producing structured text"

Retrieval Augmented Generation (RAG)

• Core concept: "RAG is a paradigm in which information is retrieved from an external source and inserted into the prompt. This can enhance performance in knowledge intensive tasks"

• Verify-and-Edit: "Zhao et al. (2023a) improves on self-consistency by generating multiple chains-of-thought, then selecting some to be edited. They do this by retrieving relevant (external) information"

• Interleaved Retrieval guided by Chain-of-Thought (IRCoT): "Trivedi et al. (2023) is a technique for multi-hop question answering that interleaves CoT and retrieval"

Evaluation

Prompting Techniques for Evaluation

• In-Context Learning: "is frequently used in evaluation prompts, much in the same way it is used in other applications"

• Role-based Evaluation: "is a useful technique for improving and diversifying evaluations. By creating prompts with the same instructions for evaluation, but different roles, it is possible to effectively generate diverse evaluations"

• Chain-of-Thought: "prompting can further improve evaluation performance"

• Model-Generated Guidelines: "Liu et al. (2023d, h) prompt an LLM to generate guidelines for evaluation. This reduces the insufficient prompting problem arising from ill-defined scoring guidelines"

Output Formats

• Styling: "Formatting the LLM's response using XML or JSON styling has also been shown to improve the accuracy of the judgment generated by the evaluator"

• Linear Scale: "A very simple output format is a linear scale (e.g. 1-5). Many works use ratings of 1-10, 1-5, or even 0-1"

• Binary Score: "Prompting the model to generate binary responses like Yes or No and True or False is another frequently used output format"

• Likert Scale: "Prompting the GenAI to make use of a Likert Scale can give it a better understanding of the meaning of the scale"

Evaluation Frameworks

• LLM-EVAL: "Lin and Chen (2023) is one of the simplest evaluation frameworks. It uses a single prompt that contains a schema of variables to evaluate"

• G-EVAL: "Liu et al. (2023d) is similar to LLM-EVAL, but includes an AutoCoT steps in the prompt itself"

• ChatEval: "Chan et al. (2024) uses a multi-agent debate framework with each agent having a separate role"

Other Methodologies

• Batch Prompting: "For improving compute and cost efficiency, some works employ batch prompting for evaluation where multiple instances are evaluated at once"

• Pairwise Evaluation: "Chen et al. (2023g) find that directly comparing the quality of two texts may lead to suboptimal results and that explicitly asking LLM to generate a score for individual summaries is the most effective"

Security & Safety

Prompt Hacking

• Definition: "Prompt hacking refers to a class of attacks which manipulate the prompt in order to attack a GenAI"

• Prompt Injection: "is the process of overriding original developer instructions in the prompt with user input"

• Jailbreaking: "is the process of getting a GenAI model to do or say unintended things through prompting"

Security Risks

• Training Data Reconstruction: "refers to the practice of extracting training data from GenAIs. A straightforward example of this is Nasr et al. (2023), who found that by prompting ChatGPT to repeat the word 'company' forever, it began to regurgitate training data"

• Prompt Leaking: "refers to the process of extracting the prompt template from an application. Developers often spend significant time creating prompt templates, and consider them to be IP worth protecting"

• Package Hallucination: "occurs when LLM-generated code attempts to import packages that do not exist. After discovering what package names are frequently hallucinated by LLMs, hackers could create those packages, but with malicious code"

Defense Mechanisms

• Prompt-based Defenses: "Multiple prompt-based defenses have been proposed, in which instructions are included in the prompt to avoid prompt injection. However, Schulhoff et al. (2023) ran a study with hundreds of thousands of malicious prompts and found that no prompt-based defense is fully secure"

• Detectors: "are tools designed to detect malicious inputs and prevent prompt hacking. Many companies have built such detectors, which are often built using fine-tuned models trained on malicious prompts"

• Guardrails: "are rules and frameworks for guiding GenAI outputs. Guardrails often make use of detectors, but not always. Guardrails are more concerned with the general dialogue flow in an application"

Alignment Issues

Prompt Sensitivity

• Small changes impact: "Several works show that LLMs are highly sensitive to the input prompt, i.e., even subtle changes to a prompt such as exemplar order can result in vastly different outputs"

• Task format variation: "describes different ways to prompt an LLM to execute the same task... Zhao et al. (2021b) show that these minor changes can alter the accuracy of GPT-3 by up to 30%"

• Prompt Drift: "Chen et al. (2023b) occurs when the model behind an API changes over time, so the same prompt may produce different results on the updated model"

Calibration Issues

• Overconfidence: "LLMs are often overconfident in their answers, especially when prompted to express their own confidence in words, which may lead to user overreliance on model outputs"

• Sycophancy: "refers to the concept that LLMs will often express agreement with the user, even when that view contradicts the model's own initial output"

Bias & Fairness

• Vanilla Prompting: "Si et al. (2023b) simply consists of an instruction in the prompt that tells the LLM to be unbiased. This technique has also been referred to as moral self-correction"

• Cultural Awareness: "Yao et al. (2023a) can be injected into prompts to help LLMs with cultural adaptation"

• AttrPrompt: "Yu et al. (2023) is a prompting technique designed to avoid producing text biased towards certain attributes when generating synthetic data"

Ambiguity Handling

• Ambiguous Demonstrations: "Gao et al. (2023a) are examples that have an ambiguous label set. Including them in a prompt can increase ICL performance"

• Question Clarification: "Rao and Daumé III (2019) allows the LLM to identify ambiguous questions and generate clarifying questions to pose to the user"

Benchmarking Results

MMLU Evaluation

• Performance trends: "Performance generally improved as techniques grew more complex. However, Zero-Shot-CoT dropped precipitously from Zero-Shot. Although it had a wide spread, for all variants, Zero-Shot performed better"

• Best performer: "Few-Shot CoT performs the best, and unexplained performance drops from certain techniques need further research"

• Self-Consistency impact: "Both cases of Self-Consistency, naturally had lower spread since they repeated a single technique, but it only improved accuracy for Zero-Shot prompts"

Case Study: Suicide Crisis Detection

• Problem domain: "Our illustrative problem involves detection of signal that is predictive of crisis-level suicide risk in text written by a potentially suicidal individual"

• Target construct: "We focus here on the most important predictive factor in Suicide Crisis Syndrome assessments, referred to in the literature as either frantic hopelessness or entrapment"

• Dataset: "Two coders trained on the recognition of the factors in Suicide Crisis Syndrome coded a set of 221 posts for presence or absence of entrapment, achieving solid inter-coder reliability (Krippendorff's alpha = 0.72)"

Prompt Engineering Process

• Development effort: "The exercise proceeded through 47 recorded development steps, cumulatively about 20 hours of work. From a cold start with 0% performance, performance was boosted to an F1 of 0.53"

• Best manual approach: "10-Shot AutoDiCoT prompt includes 15 exemplars (without CoT reasoning) and one bootstrapped reasoning demonstration"

• DSPy comparison: "The best resulting prompt... achieves 0.548 F1 (and 0.385 / 0.952 precision / recall) on the test set, without making any use of the professor's email nor the incorrect instruction about the explicitness of entrapment"

Key Takeaways

• Sensitivity to details: "prompt engineering is fundamentally different from other ways of getting a computer to behave the way you want it to: these systems are being cajoled, not programmed, and... can be incredibly sensitive to specific details in prompts without there being any obvious reason those details should matter"

• Domain expertise crucial: "the third and most important take-away is that prompt engineering should involve engagement between the prompt engineer, who has expertise in how to coax LLMs to behave in desired ways, and domain experts, who understand what those desired ways are and why"

• Automation value: "Ultimately we found that there was significant promise in an automated method for exploring the prompting space, but also that combining that automation with human prompt engineering/revision was the most successful approach"

Most-Used Techniques & Models

Popular Techniques (by citations)

• Top techniques: "The prevalence of citations for Few-Shot and Chain-of-Thought prompting is unsurprising and helps to establish a baseline for understanding the prevalence of other techniques"

Popular Models (by citations in dataset)

• Top models cited include: GPT-3, GPT-4, ChatGPT, PaLM, LLaMA families

Popular Benchmarks

• Top datasets: MMLU, GSM8K, various arithmetic and commonsense reasoning benchmarks

Future Directions & Recommendations

For Beginners

• Start simple: "To those just beginning in prompt engineering, our recommendations resemble what one would recommend in any machine learning setting: understand the problem you are trying to solve (rather than just focusing on input/output and benchmark scores)"

• Stay skeptical: "It is better to start with simpler approaches first, and to remain skeptical of claims about method performance"

For Practitioners

• Contextual understanding: "To those already engaged in prompt engineering, we hope that our taxonomy will shed light on the relationships between existing techniques"

For Researchers

• Situate new work: "To those developing new techniques, we encourage situating new methods within our taxonomy, as well as including ecologically valid case studies and illustrations of those techniques"

Key References & Tools

Foundational Papers

• Brown et al. (2020) - GPT-3 and Few-Shot Learning
• Wei et al. (2022b) - Chain-of-Thought Prompting
• Kojima et al. (2022) - Zero-Shot Chain-of-Thought
• Wang et al. (2022) - Self-Consistency
• Zhou et al. (2022a) - Least-to-Most Prompting
• Yao et al. (2023b) - Tree-of-Thought
• Liu et al. (2023b) - Pre-training Survey

Agent Frameworks

• Karpas et al. (2022) - MRKL Systems
• Yao et al. (2022) - ReAct
• Shinn et al. (2023) - Reflexion
• Wang et al. (2023a) - Voyager

Tools & Platforms

• LearnPrompting.org - Maintained terminology and techniques list
• HuggingFace Dataset - Full paper corpus
• GitHub Repository - Code and resources

Evaluation & Safety

• Schulhoff et al. (2023) - HackAPrompt competition on prompt injection
• Liu et al. (2023d) - G-EVAL framework
• Chan et al. (2024) - ChatEval multi-agent debate

Multilingual & Multimodal

• Shi et al. (2022) - Multilingual Chain-of-Thought
• Zhang et al. (2023d) - Multimodal Chain-of-Thought
• Asai et al. (2023) - BUFFET multilingual benchmark

Automated Prompt Engineering

• Zhou et al. (2022b) - APE (Automatic Prompt Engineer)
• Khattab et al. (2023) - DSPy framework
• Deng et al. (2022) - RLPrompt

Dataset & Methodology Details

Dataset Composition

• Final corpus: "The dataset contains 1,565 research papers in PDF format. Any duplicate papers were removed automatically, though some could exist"

• Time frame: "The dataset was curated the duration of the research paper, primarily in February of 2024"

• Source distribution: "We wrote scripts to automatically query the APIs of Arxiv and Semantic Scholar"

Quality Control

• Human validation: "After collecting data from different sources, we removed duplicate papers and did a manual and semi-automated review of papers to ensure they were all relevant"

• LLM-assisted review: "We develop a prompt using gpt-4-1106-preview to classify the remaining articles. We validate the prompt against 100 ground-truth annotations, achieving 89% precision and 75% recall (for an F1 of 81%)"

Search Keywords (Selected Examples)

• Core terms: "jailbreak prompt", "prompt engineering", "few-shot learning", "in-context learning"
• Technique-specific: "chain-of-thought", "zero-shot prompting", "prompt optimization"
• Domain-specific: "llm prompting", "transformer model prompts", "multimodal prompting"

Critical Insights & Limitations

Nature of Prompting

• Black art acknowledgment: "This can be interpreted both optimistically and pessimistically. Optimistically, it demonstrates how improvements can arise through exploration and fortuitous discovery. On the pessimistic side, the value of duplicating the email in the prompt highlights the extent to which prompting remains a difficult to explain black art"

• Emergent vs discovered: "Many of the techniques described here have been called 'emergent', but it is perhaps more appropriate to say that they were discovered—the result of thorough experimentation, analogies from human reasoning, or pure serendipity"

Validation Challenges

• Lack of standardization: "The field is new, and evaluation is variable and unstandardized—even the most meticulous experimentation may suffer from unanticipated shortcomings, and model outputs themselves are sensitive to meaning-preserving changes in inputs"

• Transfer uncertainty: "As a result, we encourage the reader to avoid taking any claims at face value and to recognize that techniques may not transfer to other models, problems, or datasets"

Scope Limitations

• Focus restrictions: "To keep the work approachable to less technical readers and maintain a manageable scope... we only study task-agnostic techniques"

• Exclusions: "These decisions keep the work approachable to less technical readers and maintain a manageable scope"

Practical Implementation Notes

Prompt Template Best Practices

• Variable replacement: "A prompt template is a function that contains one or more variables which will be replaced by some media (usually text) to create a prompt"

• Context preservation: "It is often necessary to include additional information in the prompt... Additional Information is sometimes called 'context', though we discourage the use of this term as it is overloaded with other meanings in the prompting space"

Answer Extraction Strategies

• Verbalizer design: "For example, if we wish for a model to predict whether a Tweet is positive or negative, we could prompt it to output either '+' or '-' and a verbalizer would map these token sequences to the appropriate labels"

• Regex patterns: "Regexes are often used to extract answers. They are usually used to search for the first instance of a label. However, depending on the output format and whether CoTs are generated, it may be better to search for the last instance"

• Cascading approaches: "Sometimes outputs are so complicated that regexes won't work consistently. In this case, it can be useful to have a separate LLM evaluate the output and extract an answer"

Model Selection Considerations

• Guardrails interference: "A take-away from this initial phase is that the 'guard rails' associated with some large language models may interfere with the ability to make progress on a prompting task, and this could influence the choice of model for reasons other than the LLM's potential quality"

• Temperature settings: "For the two Self-Consistency results, we set temperature to 0.5, following Wang et al. (2022)'s guidelines. For all other prompts, a temperature of 0 was used"

Terminology Disambiguation

Conflicting Usages

• In-Context Learning ambiguity: "Note that the word 'learn' is misleading. ICL can simply be task specification–the skills are not necessarily new, and can have already been included in the training data"

• Brown et al. definitions: "Brown et al. (2020) seemingly offer two different definitions for ICL... However, they explicitly state that ICL does not necessarily involve learning new tasks"

• Prompt vs Prompt Template: "Brown et al. (2020) consider the word 'llama' to be the prompt, while 'Translate English to French:' is the 'task description'. More recent papers, including this one, refer to the entire string passed to the LLM as the prompt"

Hard vs Soft Prompts

• Hard (discrete): "These prompts only contain tokens that directly correspond to words in the LLM vocabulary"

• Soft (continuous): "These prompts contain tokens that may not correspond to any word in the vocabulary... Soft prompts can be used when fine-tuning is desired, but modifying the weights of the full model is prohibitively expensive"

Prefix vs Cloze

• Prefix prompts: "In Prefix prompts, the token to be predicted is at the end of the prompt. This is usually the case with modern GPT-style models"

• Cloze prompts: "In Cloze prompts, the token(s) to be predicted are presented as 'slots to fill', usually somewhere in the middle of the prompt. This is usually the case for earlier transformer models such as BERT"

Advanced Technique Details

AutoDiCoT (Novel Contribution)

• Algorithm description: "We call the algorithm in Figure 6.12 Automatic Directed CoT (AutoDiCoT), since it automatically directs the CoT process to reason in a particular way"

• Process: "For each pair (qi, ai) in training data: Label qi as entrapment or not using the model. If correct, prompt with 'Why?' to generate reasoning. If incorrect, prompt 'It is actually [is/is not] entrapment, please explain why.'"

• Generalizability: "This technique can be generalized to any labeling task. It combines the automatic generation of CoTs with showing the LLM examples of bad reasoning, as in the case of Contrastive CoT"

Design Decision Framework

• Six critical factors: "We highlight six separate design decisions, including the selection and order of exemplars that critically influence the output quality"

• Tradeoffs: "Although effective, employing KNN during prompt generation may be time and resource intensive"

Iterative Retrieval

• FLARE approach: "Forward-Looking Active REtrieval augmented generation (FLARE) and Imitate, Retrieve, Paraphrase (IRP) perform retrieval multiple times during long-form generation"

• Three-step process: "1) generating a temporary sentence to serve as a content plan; 2) retrieving external knowledge using the temporary sentence as a query; 3) injecting the retrieved knowledge into the temporary sentence"

• Query quality: "These temporary sentences have been shown to be better search queries compared to the document titles provided in long-form generation tasks"

Meta-Analysis Statistics

Citation Patterns

• Most cited techniques: "The prevalence of citations for Few-Shot and Chain-of-Thought prompting is unsurprising and helps to establish a baseline for understanding the prevalence of other techniques"

• Model usage: Citation analysis shows GPT family dominates research, followed by PaLM and open-source alternatives

• Dataset popularity: MMLU, GSM8K, and arithmetic reasoning benchmarks most frequently used

Research Trends

• Paper growth: 1,565 relevant papers identified from broader corpus of 4,247 unique records

• Quality metrics: Inter-annotator agreement of 92% (Krippendorff's α = Cohen's κ = 81%) for relevance labeling

• LLM assistance: "We validate the prompt against 100 ground-truth annotations, achieving 89% precision and 75% recall (for an F1 of 81%)" for automated paper screening

Formal Definitions

Mathematical Formulation

• Basic prompt conditioning: "p(A|T,Q) = ∏(i=1 to |A|) p_LM(ai|T,Q,a1:i-1)" where T is prompt template, Q is question, A is answer

• Few-shot extension: "p(A|T(X,x)) = ∏(i=1 to |A|) p_LM(ai|T(X,x),a1:i-1)" where X is set of training exemplars

• Optimization objective: "T* = argmax_T E_{xi,yi~D}[S(p_LM(A|T(xi)),yi)]" maximizing scoring function S over dataset D

• Answer engineering: "A ~ p_LM(A|T(xi),yi); T* = argmax_{T,E} E_{xi,yi~D}[S(E(A),yi)]" where E is extraction function

Storage & Implementation Constraints

Browser Environment

• Critical restriction: "NEVER use localStorage, sessionStorage, or ANY browser storage APIs in artifacts. These APIs are NOT supported and will cause artifacts to fail in the Claude.ai environment"

• Alternatives: "Instead, you MUST: Use React state (useState, useReducer) for React components; Use JavaScript variables or objects for HTML artifacts; Store all data in memory during the session"

Library Availability (React Artifacts)

• Available libraries include: lucide-react, recharts, MathJS, lodash, d3, Plotly, Three.js (r128), Papaparse, SheetJS, shadcn/ui, Chart.js, Tone, mammoth, tensorflow
• Important limitation: "NO OTHER LIBRARIES ARE INSTALLED OR ABLE TO BE IMPORTED"
• Three.js caveat: "IMPORTANT: Do NOT use THREE.CapsuleGeometry as it was introduced in r142. Use alternatives like CylinderGeometry, SphereGeometry, or create custom geometries instead"

Contributions & Authorship

Team Structure

• Lead authors: Sander Schulhoff (lead), Michael Ilie (co-lead)
• Principal investigator: Philip Resnik
• Total contributors: 58 authors from 13 institutions

Major Section Leads

• Benchmarking: Konstantine Kahadze
• Agents: Ashay Srivastava
• Alignment: Nishant Balepur
• Security: Sevien Schulhoff
• Multilingual: Dayeon Ki
• Evaluation: Sweta Agrawal

Domain Expertise

• SCS labeling: Megan L. Rogers, Inna Goncearenco, Giuseppe Sarli, Igor Galynker provided clinical expertise
• Multilingual guidance: Marine Carpuat framed and reviewed multilingual section

Additional Resources

Maintained Resources

• Live terminology: "We maintain an up-to-date list of terms and techniques at LearnPrompting.org"
• Dataset access: Available on HuggingFace with full datasheet
• Code repository: GitHub with systematic review pipeline

Future Updates

• Iterative taxonomy: "We expect this to be the first iteration of terminologies that will develop over time"
• Community contribution: "If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so? Yes, anyone is free to use/modify the data"

Citation Information

• Preferred citation: Schulhoff et al. (2024), "The Prompt Report: A Systematic Survey of Prompting Techniques"
• Contact: sanderschulhoff@gmail.com for dataset inquiries
• Funding acknowledgment: "$10,000 in API credits given by OpenAI"

RAG is Dead, Context Engineering is King — with Jeff Huber of Chroma

Core Thesis

• Context Engineering over RAG: "RAG" as a term is fundamentally flawed and confusing

  "RAG. We never use the term rag. I hate the term rag... retrieval, augmented generation. Are three concepts put together into one thing? Like, that's just really confusing."

• Context Engineering Definition: The job of determining optimal context window contents

  "Context engineering is the job of figuring out what should be in the context window any given LLM generation step. And there's both an inner loop, which is setting up the, you know, what should be in the context window this time. And there's the outer loop, which is how do you get better over time at filling the context window with only the relevant information."

Context Rot Research

• Models degrade with longer contexts: Performance is not invariant to token count

  "The performance of LLMs is not invariant to how many tokens you use. As you use more and more tokens, the model can pay attention to less and then also can reason sort of less effectively."

• Needle in Haystack is misleading: Lab benchmarks don't reflect real-world usage

  "There was this bit of, like, this sort of implication where, like, oh, look, our model is perfect on this task, needle in a haystack. Therefore, the context window you can use for whatever you want. There was an implication there. And, well, I hope that that is true someday. That is not the case."

• Claude Sonnet 4.5 performs best: Based on area under curve for context utilization

  "I don't have much commentary. That is what we found for this particular task... I think it shows here if this is true, that's a big explanation for why" developers love Claude

Retrieval System Architecture

First-Stage Retrieval (Hybrid Approach)

• Multiple signals for initial culling: Dense vectors, lexical search, metadata filtering

  "One pattern is to use what a lot of people call first stage retrieval to do a big cull down... using signals like vector search, like full text search, like metadata filtering, metadata search, and others to go from, let's say 10,000 down to 300."

• LLMs can handle more than 10 results: Unlike traditional search for humans

  "You don't have to give an LLM 10 blue links. You can brute force a lot more."

Re-ranking with LLMs

• LLM re-ranking is cost-effective and emerging: From 300 candidates down to 30

  "Using an LLM as a re-ranker and brute forcing from 300 down to 30, I've seen now emerging a lot, like a lot of people are doing this and it actually is like way more cost effective than I think a lot of people realize I've heard of people that are running models themselves that are getting like a penny per million input tokens"

• Purpose-built re-rankers will decline: Like specialized hardware, only needed at extreme scale

  "I actually think that like probably purpose built re-rankers will go away. And the same way that like purpose built... if you're at extreme scale, extreme cost, yes, you'll care to optimize that... the same way that if you're running with hardware... you're just going to use a CPU or GPU. Unless you absolutely have to."

Context Assembly Best Practices

• Structured ingestion matters: Extract metadata and signals at write time

  "As much structured information as you can put into your write or your ingestion pipeline, you should. So all of the metadata you can extract, do it at ingestion. All of the chunk rewriting you can do, do it at ingestion."

• Chunk rewriting for code: Generate natural language descriptions

  "Instead of just embedding the code, you first have an LLM generate like a natural language description of like what this code is doing. And either you embed like just the natural language description or you embed that and the code"

Code Retrieval Strategies

• Regex remains dominant: 85-90% of queries satisfied, but embeddings add value

  "My guess is that like for code today, it's something like 90% of queries or 85% of queries can be satisfactorily run with Regex... But you maybe can get like 15% or 10% or 5% improvement by also using embeddings."

• Chroma supports native regex search: With indexing for scale

  "We've actually worked on now inside of Chroma, both single load and distributed, we support regex search natively. So you can do regex search inside of Chroma because we've seen that as like a very powerful tool for code search."

• Fork-based indexing for versioning: Fast branch/commit-specific indexes

  "Another feature we added to Chroma is the ability to do forking. So you can take an existing index and you can create a copy of that index in under a hundred milliseconds for pennies... you now can like have an index for like different each commit."

Generative Benchmarking

• Small golden datasets are highly valuable: Few hundred examples sufficient

  "The returns to a very high-quality small label data set are so high. Everybody thinks you have to have, like, a million examples or whatever. No. Actually, just, like, a couple hundred even, like, high-quality examples is extremely beneficial."

• Generate synthetic QA pairs: When you have chunks but need queries

  "We did a whole technical report around how do you teach an LLM to write good queries from chunks? Because, again, you want, like, chunk query pairs. And so if you have the chunks, you need the queries."

• Data labeling parties work: Simple, practical approach

  "Thursday night, we're all going to be in the conference room. We're ordering pizza. And we're just going to have a data labeling party for a few hours. That's all it takes to bootstrap this."

Memory and Context Engineering

• Memory is context engineering's benefit: Same problem, different framing

  "Memory again is like the memory is the term that like everybody can understand... but what is memory under the hood? It's still just context engineering... the domain of how do you put the right information into the context window?"

• Compaction enables offline improvement: Re-indexing and refinement

  "Offline processing is helpful, and I think that is also helpful in this case... You're taking data. You're like, oh, maybe those two data points should be merged. Maybe they should be split. Maybe they should be, like, rewritten."

Future of Retrieval Systems

• Stay in latent space: Avoid natural language round-trip

  "Why are we going back to natural language? Why aren't we just like passing the embeddings like directly to the models who are just going to functionally like re put it into latent space."

• Continuous retrieval during generation: Not just one-shot before generation

  "For the longest time we've done one retrieval per generation... why are we not continually retrieving as we need to"

• Current approaches are crude: Will seem primitive in 5-10 years

  "I think when we look back in things, this was like, like hilariously crude, the way we do things today."

Chroma Product Philosophy

• Developer experience is paramount: Zero-config, serverless approach

  "In the same way that you could run pip install ChromaDB and be up and running in five seconds... That same story had to be true for the cloud... It needed to be like zero config, zero knobs to tune. It should just be always fast, always very cost-effective and always fresh without you having to do or think about anything."

• Usage-based billing: True serverless pricing

  "We only charge you for the minimal slice of compute that you use and like nothing more, which not all serverless databases can claim"

• Slow, intentional hiring: Culture and craft over speed

  "The slope of our future growth is entirely dependent on the people that are here in this office... we've just really decided to hire very slowly and be really picky."

Key Technical Reports

1. Context Rot: LLM performance degradation with context length
2. Generative Benchmarking: Synthetic QA pair generation for evaluation

Referenced Papers/Technologies

• REALM: Retrieval-Augmented Language Model
• RETRO: Retrieval-Enhanced Transformer
• Voyager paper: Used Chroma in implementation
• Lance Martin's Context Engineering: Key thought leadership
• TLA Plus, Raft, Paxos: Distributed systems foundations

Company Details

• Downloads: 5M+ monthly, 70M+ all-time on PyPI
• GitHub: 21,000+ stars
• Architecture: Rust-based, fully multi-tenant, separation of storage/compute
• Open Source: Apache 2 license for core and distributed versions
• Cloud: Serverless, usage-based, $5 free credits (~100K docs + queries)

gradually decline in effectiveness or vigour due to underuse or neglect.

(of a person) tending to spend much time seated; somewhat inactive.

体内平衡的

eLife Assessment

This useful study reports a method to detect and analyze a novel post-translational modification, lysine acetoacetylation (Kacac), finding it regulates protein metabolism pathways. The study unveils epigenetic modifiers involved in placing this mark, including key histone acetyltransferases such as p300, and concomitant HDACs, which remove the mark. Proteomic and bioinformatics analysis identified many human proteins with Kacac sites, potentially suggesting broad effects on cellular processes and disease mechanisms. The data presented are solid, although some concerns persist regarding inconsistencies in molecular weight of the enzyme used. The study will be of interest to those studying protein and metabolic regulation.

Reviewer #2 (Public review):

In the manuscript by Fu et al., the authors developed a chemo-immunological method for the reliable detection of Kacac, a novel post-translational modification, and demonstrated that acetoacetate and AACS serve as key regulators of cellular Kacac levels. Furthermore, the authors identified the enzymatic addition of the Kacac mark by acyltransferases GCN5, p300, and PCAF, as well as its removal by deacetylase HDAC3. These findings indicate that AACS utilizes acetoacetate to generate acetoacetyl-CoA in the cytosol, which is subsequently transferred into the nucleus for histone Kacac modification. A comprehensive proteomic analysis has identified 139 Kacac sites on 85 human proteins. Bioinformatics analysis of Kacac substrates and RNA-seq data reveal the broad impacts of Kacac on diverse cellular processes and various pathophysiological conditions. This study provides valuable additional insights into the investigation of Kacac and would serve as a helpful resource for future physiological or pathological research.

Comments on revised version:

The authors have made efforts to revise this manuscript and address my concerns. The revisions are appropriate and have improved the quality of the manuscript.

Reviewer #3 (Public review):

Summary:

This paper presents a timely and significant contribution to the study of lysine acetoacetylation (Kacac). The authors successfully demonstrate a novel and practical chemo-immunological method using the reducing reagent NaBH4 to transform Kacac into lysine β-hydroxybutyrylation (Kbhb).

Strengths:

This innovative approach enables simultaneous investigation of Kacac and Kbhb, showcasing its potential in advancing our understanding of post-translational modifications and their roles in cellular metabolism and disease.

Weaknesses:

The experimental evidence presented in the article is insufficient to fully support the authors' conclusions. In the in vitro assays, the proteins used appear to be highly inconsistent with their expected molecular weights, as shown by Coomassie Brilliant Blue staining (Figure S3A). For example, p300, which has a theoretical molecular weight of approximately 270 kDa, appeared at around 37 kDa; GCN5/PCAF, expected to be ~70 kDa, appeared below 20 kDa. Other proteins used in the in vitro experiments also exhibited similarly large discrepancies from their predicted sizes. These inconsistencies severely compromise the reliability of the in vitro findings. Furthermore, the study lacks supporting in vivo data, such as gene knockdown experiments, to validate the proposed conclusions at the cellular level.

Author response:

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

Reviewer #1 (Public review):

Summary

Lysine acetoacetylation (Kacac) is a recently discovered histone post-translational modification (PTM) connected to ketone body metabolism. This research outlines a chemo-immunological method for detecting Kacac, eliminating the requirement for creating new antibodies. The study demonstrates that acetoacetate acts as the precursor for Kacac, which is catalyzed by the acyltransferases GCN5, p300, and PCAF, and removed by the deacetylase HDAC3. AcetoacetylCoA synthetase (AACS) is identified as a central regulator of Kacac levels in cells. A proteomic analysis revealed 139 Kacac sites across 85 human proteins, showing the modification's extensive influence on various cellular functions. Additional bioinformatics and RNA sequencing data suggest a relationship between Kacac and other PTMs, such as lysine βhydroxybutyrylation (Kbhb), in regulating biological pathways. The findings underscore Kacac's role in histone and non-histone protein regulation, providing a foundation for future research into the roles of ketone bodies in metabolic regulation and disease processes.

Strengths 

(1) The study developed an innovative method by using a novel chemo-immunological approach to the detection of lysine acetoacetylation. This provides a reliable method for the detection of specific Kacac using commercially available antibodies.

(2) The research has done a comprehensive proteome analysis to identify unique Kacac sites on 85 human proteins by using proteomic profiling. This detailed landscape of lysine acetoacetylation provides a possible role in cellular processes.

(3) The functional characterization of enzymes explores the activity of acetoacetyltransferase of key enzymes like GCN5, p300, and PCAF. This provides a deeper understanding of their function in cellular regulation and histone modifications.

(4) The impact of acetyl-CoA and acetoacetyl-CoA on histone acetylation provides the differential regulation of acylations in mammalian cells, which contributes to the understanding of metabolic-epigenetic crosstalk.

(5) The study examined acetoacetylation levels and patterns, which involve experiments using treatment with acetohydroxamic acid or lovastatin in combination with lithium acetoacetate, providing insights into the regulation of SCOT and HMGCR activities.

We thank all the reviewers for their positive, insightful comments which have helped us improve our manuscript. We have revised the manuscript as suggested by the reviewers.

Weakness 

(1) There is a limitation to functional validation, related to the work on the biological relevance of identified acetoacetylation sites. Hence, the study requires certain functional validation experiments to provide robust conclusions regarding the functional implications of these modifications on cellular processes and protein function. For example, functional implications of the identified acetoacetylation sites on histone proteins would aid the interpretation of the results.

We agree with the reviewer that investigating the functional role of individual histone Kacac sites is essential for understanding the epigenetic impact of Kacac marks on gene expression, signaling pathways, and disease mechanisms. This topic is out of the scope of this paper which focuses on biochemical studies and proteomics. Functional elucidation in specific pathways will be a critical direction for future investigation, ideally with the development of site-specific anti-Kacac antibodies.

(2) The authors could have studied acetoacetylation patterns between healthy cells and disease models like cancer cells to investigate potential dysregulation of acetoacetylation in pathological conditions, which could provide insights into their PTM function in disease progression and pathogenesis.

We appreciate the reviewer’s valuable suggestion. In our study, we measured Kacac levels in several types of cancer cell lines, including HCT116 (Fig. 2B), HepG2 (Supplementary Fig. S2), and HeLa cells (data not shown in the manuscript), and found that acetoacetate-mediated Kacac is broadly present in all these cancer cell lines. Our proteomics analysis linked Kacac to critical cellular functions, e.g. DNA repair, RNA metabolism, cell cycle regulation, and apoptosis, and identified promising targets that are actively involved in cancer progression such as p53, HDAC1, HMGA2, MTA2, LDHA. These findings suggest that Kacac has significant, non-negligible effects on cancer pathogenesis. We concur that exploring the acetoacetylation patterns in cancer patient samples with comparison with normal cells represents a promising direction for next-step research. We plan to investigate these big issues in future studies. 

(3) The time-course experiments could be performed following acetoacetate treatment to understand temporal dynamics, which can capture the acetoacetylation kinetic change, thereby providing a mechanistic understanding of the PTM changes and their regulatory mechanisms.

As suggested, time-course experiments were performed, and the data have been included in the revised manuscript (Supplementary Fig. S2A).

(4) Though the discussion section indeed provides critical analysis of the results in the context of existing literature, further providing insights into acetoacetylation's broader implications in histone modification. However, the study could provide a discussion on the impact of the overlap of other post-translational modifications with Kacac sites with their implications on protein functions.

We appreciate the reviewer’s helpful suggestion. We have added more discussions on the impact of the Kacac overlap with other post-translational modifications in the discussion section of the revised manuscript.

Impact

The authors successfully identified novel acetoacetylation sites on proteins, expanding the understanding of this post-translational modification. The authors conducted experiments to validate the functional significance of acetoacetylation by studying its impact on histone modifications and cellular functions.

We appreciate the reviewer’s comments.

Reviewer #2 (Public review):

In the manuscript by Fu et al., the authors developed a chemo-immunological method for the reliable detection of Kacac, a novel post-translational modification, and demonstrated that acetoacetate and AACS serve as key regulators of cellular Kacac levels. Furthermore, the authors identified the enzymatic addition of the Kacac mark by acyltransferases GCN5, p300, and PCAF, as well as its removal by deacetylase HDAC3. These findings indicate that AACS utilizes acetoacetate to generate acetoacetyl-CoA in the cytosol, which is subsequently transferred into the nucleus for histone Kacac modification. A comprehensive proteomic analysis has identified 139 Kacac sites on 85 human proteins. Bioinformatics analysis of Kacac substrates and RNA-seq data reveals the broad impacts of Kacac on diverse cellular processes and various pathophysiological conditions. This study provides valuable additional insights into the investigation of Kacac and would serve as a helpful resource for future physiological or pathological research.

The following concerns should be addressed:

(1) A detailed explanation is needed for selecting H2B (1-26) K15 sites over other acetylation sites when evaluating the feasibility of the chemo-immunological method.

The primary reason for selecting the H2B (1–26) K15acac peptide to evaluate the feasibility of our chemo-immunological method is that H2BK15acac was one of the early discovered modification sites in our preliminary proteomic screening data. The panKbhb antibody used herein is independent of peptide sequence so different modification sites on histones can all be recognized. We have added the explanation to the manuscript.

(2) In Figure 2(B), the addition of acetoacetate and NaBH4 resulted in an increase in Kbhb levels. Specifically, please investigate whether acetoacetylation is primarily mediated by acetoacetyl-CoA and whether acetoacetate can be converted into a precursor of β-hydroxybutyryl (bhb-CoA) within cells. Additional experiments should be included to support these conclusions.

We appreciate the reviewer’s valuable comments. In our paper, we had the data showing that acetoacetate treatment had very little effect on histone Kbhb levels in HEK293T cells, as observed in lanes 1–4 of Fig. 2A, demonstrating that acetoacetate minimally contributes to Kbhb generation. We drew the conclusion that histone Kacac is primarily mediated by acetoacetyl-CoA based on multiple pieces of evidence: first, we observed robust Kacac formation from acetoacetyl-CoA upon incubation with HATs and histone proteins or peptides, as confirmed by both western blotting (Figs. 3A, 3B; Supplementary Figs. S3C– S3F) and MALDI-MS analysis (Supplementary Fig. S4A). Second, treatment with hymeglusin—a specific inhibitor of hydroxymethylglutaryl-CoA synthase, which catalyzes the conversion of acetoacetyl-CoA to HMG-CoA—led to increased Kacac levels in HepG2 cells (PMID: 37382194). Third, we demonstrated that AACS whose function is to convert acetoacetate into acetoacetyl-CoA leads to marked histone Kacac upregulation (Fig. 2E). Collectively, these findings strongly support the conclusion that acetoacetate promotes Kacac formation primarily via acetoacetyl-CoA.

(3) In Figure 2(E), the amount of pan-Kbhb decreased upon acetoacetate treatment when SCOT or AACS was added, whereas this decrease was not observed with NaBH4 treatment. What could be the underlying reason for this phenomenon?

In the groups without NaBH₄ treatment (lanes 5–8, Figure 2E), the Kbhb signal decreased upon the transient overexpression of SCOT or AACS, owing to protein loading variation in these two groups (lanes 7 and 8). Both Ponceau staining and anti-H3 results showed a lower amount of histones in the AACS- or SCOT-treated samples. On the other hand, no decrease in the Kbhb signal was observed in the NaBH₄-treated groups (lanes 1–4), because NaBH₄ treatment elevated Kacac levels, thereby compensating for the reduced histone loading. The most important conclusion from this experiment is that AACS overexpression increased Kacac levels, whereas SCOT overexpression had no/little effect on histone Kacac levels in HEK293T cells.

(4) The paper demonstrates that p300, PCAF, and GCN5 exhibit significant acetoacetyltransferase activity and discusses the predicted binding modes of HATs (primarily PCAF and GCN5) with acetoacetyl-CoA. To validate the accuracy of these predicted binding models, it is recommended that the authors design experiments such as constructing and expressing protein mutants, to assess changes in enzymatic activity through western blot analysis.

We appreciate the reviewer’s valuable suggestion. Our computational modeling shows that acetoacetyl-CoA adopts a binding mode similar to that of acetyl-CoA in the tested HATs. This conclusion is supported by experimental results showing that the addition of acetyl-CoA significantly competed for the binding of acetoacetyl-CoA to HATs, leading to reduced enzymatic activity in mediating Kacac (Fig. 3C). Further structural biology studies to investigate the key amino acid residues involved in Kacac binding within the GCN5/PCAF binding pocket, in comparison to Kac binding—will be a key direction of future studies.

(5) HDAC3 shows strong de-acetoacetylation activity compared to its de-acetylation activity. Specific experiments should be added to verify the molecular docking results. The use of HPLC is recommended, in order to demonstrate that HDAC3 acts as an eraser of acetoacetylation and to support the above conclusions. If feasible, mutating critical amino acids on HDAC3 (e.g., His134, Cys145) and subsequently analyzing the HDAC3 mutants via HPLC and western blot can further substantiate the findings.

We appreciate the reviewer’s helpful suggestion. In-depth characterizations of HDAC3 and other HDACs is beyond this manuscript. We plan in the future to investigate the enzymatic activity of recombinant HDAC3, including the roles of key amino acid residues and the catalytic mechanism underlying Kacac removal, and to compare its activity with that involved in Kac removal.

(6) The resolution of the figures needs to be addressed in order to ensure clarity and readability.

Edits have been made to enhance figure resolutions in the revised manuscript.

Reviewer #3 (Public review):

Summary:

This paper presents a timely and significant contribution to the study of lysine acetoacetylation (Kacac). The authors successfully demonstrate a novel and practical chemo-immunological method using the reducing reagent NaBH4 to transform Kacac into lysine β-hydroxybutyrylation (Kbhb).

Strengths:

This innovative approach enables simultaneous investigation of Kacac and Kbhb, showcasing their potential in advancing our understanding of post-translational modifications and their roles in cellular metabolism and disease.

Weaknesses:

The paper's main weaknesses are the lack of SDS-PAGE analysis to confirm HATs purity and loading consistency, and the absence of cellular validation for the in vitro findings through knockdown experiments. These gaps weaken the evidence supporting the conclusions.

We appreciate the reviewer’s positive comments on the quality of this work and the importance to the field. The SDS-PAGE results of HAT proteins (Supplementary Fig. S3A) was added in the revised manuscript. The cellular roles of p300 and GCN5 as acetoacetyltransferases were confirmed in a recent study (PMID: 37382194). Their data are consistent with our studies herein and provide further support for our conclusion. We agree that knockdown experiments are essential to further validate the activities of these enzymes and plan to address this in future studies.

Reviewer #1 (Recommendations for the authors):

This study conducted the first comprehensive analysis of lysine acetoacetylation (Kacac) in human cells, identifying 139 acetoacetylated sites across 85 proteins in HEK293T cells. Kacac was primarily localized to the nucleus and associated with critical processes like chromatin organization, DNA repair, and gene regulation. Several previously unknown Kacac sites on histones were discovered, indicating its widespread regulatory role. Key enzymes responsible for adding and removing Kacac marks were identified: p300, GCN5, and PCAF act as acetoacetyltransferases, while HDAC3 serves as a remover. The modification depends on acetoacetate, with AACS playing a significant role in its regulation. Unlike Kbhb, Kacac showed unique cellular distribution and functional roles, particularly in gene expression pathways and metabolic regulation. Acetoacetate demonstrated distinct biological effects compared to βhydroxybutyrate, influencing lipid synthesis, metabolic pathways, and cancer cell signaling. The findings suggest that Kacac is an important post-translational modification with potential implications for disease, metabolism, and cellular regulation.

Major Concerns

(1) The authors could expand the study by including different cell lines and also provide a comparative study by using cell lines - such as normal vs disease (eg. Cancer cell like) - to compare and to increase the variability of acetoacetylation patterns across cell types. This could broaden the understanding of the regulation of PTMs in pathological conditions.

We sincerely appreciate the reviewer’s valuable suggestions. We concur that a

deeper investigation into Kacac patterns in cancer cell lines would significantly enhance understanding of Kacac in the human proteome. Nevertheless, due to constraints such as limited resource availability, we are currently unable to conduct very extensive explorations as proposed. Nonetheless, as shown in Fig. 2A, Fig. 2B, and Supplementary Fig. S2, our present data provide strong evidence for the widespread occurrence of acetoacetatemediated Kacac in both normal and cancer cell lines. Notably, our proteomic profiling identified several promising targets implicated in cancer progression, including p53, HDAC1, HMGA2, MTA2, and LDHA. We plan to conduct more comprehensive explorations of acetoacetylation patterns in cancer samples in future studies.

(2) The paper lacks inhibition studies silencing the enzyme genes or inhibiting the enzyme using available inhibitors involved in acetoacetylation or using aceto-acetate analogues to selectively modulate acetoacetylation levels. This can validate their impact on downstream cellular pathways in cellular regulation.

We appreciate the reviewer’s valuable suggestions. Our study, along with the previous research, has conducted initial investigations into the inhibition of key enzymes involved in the Kacac pathway. For example, inhibition of HMGCS, which catalyzes the conversion of acetoacetyl-CoA to HMG-CoA, was shown to enhance histone Kacac levels (PMID: 37382194). In our study, we examined the inhibitory effects of SCOT and HMGCR, both of which potentially influence cellular acetoacetyl-CoA levels. However, their respective inhibitors did not significantly affect histone Kacac levels. We also investigated the role of acetyl-CoA, which competes with acetoacetyl-CoA for binding to HAT enzymes and can function as a competitive inhibitor in histone Kacac generation. Furthermore, inhibition of HDAC activity by SAHA led to increased histone Kacac levels in HepG2 cells (PMID: 37382194), supporting our conclusion that HDAC3 functions as the eraser responsible for Kacac removal. These inhibition studies confirmed the functions of these enzymes and provided insights into their regulatory roles in modulating Kacac and its downstream pathways. Further in-depth investigations will explore the specific roles of these enzymes in regulating Kacac within cellular pathways.

(3) The authors could validate the functional impact of pathways using various markers through IHC/IFC or western blot to confirm their RNA-seq analysis, since pathways could be differentially regulated at the RNA vs protein level.

We agree that pathways can be differentially regulated at the RNA and protein levels. It is our future plan to select and fully characterize one or two gene targets to elaborate the presence and impact of Kacac marks on their functional regulation at both the gene expression and protein level.

(4) Utilize in vitro reconstitution assays to confirm the direct effect of acetoacetylation on histone modifications and nucleosome assembly, establishing a causal relationship between acetoacetylation and chromatin regulation.

We appreciate this suggestion, and this will be a very fine biophysics project for us and other researchers for the next step. We plan to do this and related work in a future paper to characterize the impact of lysine acetoacetylation on chromatin structure and gene expression. Technique of site-specific labelling will be required. Also, we hope to obtain monoclonal antibodies that directly recognize Kacac in histones to allow for ChIP-seq assays in cells.

(5) The authors could provide a site-directed mutagenesis experiment by mutating a particular site, which can validate and address concerns regarding the specificity of a particular site involved in the mechanism.

We agree that validating and characterizing the specificity of individual Kacac sites and understanding their functional implications are important for elucidating the mechanisms by which Kacac affects these substrate proteins. Such work will involve extensive biochemical and cellular studies. It is our future goal to select and fully characterize one or two gene targets in detail and in depth to elaborate the presence and impact of Kacac on their function regulation using comprehensive techniques (transfection, mutation, pulldown, and pathway analysis, etc.).

(6) If possible, the authors could use an in vivo model system, such as mice, to validate the physiological relevance of acetoacetylation in a more complex system.  

We currently do not have access to resources of relevant animal models. We will conduct in vivo screening and characterization of protein acetoacetylation in animal models and clinical samples in collaboration with prospective collaborators.

Minor Concerns

(1) The authors could discuss the overlap of Kacac sites with other post-translational modifications and their implications on protein functions. They could provide comparative studies with other PTMs, which can improvise a comprehensive understanding of acetoacetylation function in epigenetic regulation.

We have expanded the discussion in the revised manuscript to address the overlap between Kacac and other post-translational modifications, along with their potential functional implications.

(2) The authors could provide detailed information on the implications of their data, which would enhance the impact of the research and its relevance to the scientific community. Specifically, they could clarify the acetoacetylation (Kacac) significance in nucleosome assembly and its correlation with RNA processing.

In the revised manuscript, we have added more elaborations on the implication and significance of Kacac in nucleosome assembly and RNA processing.

Reviewer #3 (Recommendations for the authors):

Major Comments:

(1) Figures 3A, 3B, Supplementary Figures S3A-D

I could not find the SDS-PAGE analysis results for the purified HATs used in the in vitro assay. It is imperative to display these results to confirm consistent loading amounts and sufficient purity of the HATs across experimental groups. Additionally, I did not observe any data on CBP, even though it was mentioned in the results section. If CBP-related experiments were not conducted, please remove the corresponding descriptions.

We appreciate the reviewer’s valuable suggestion. The SDS-PAGE results for the HAT proteins have been included, and the part in the results section discussing CBP has been updated according to the reviewer’s suggestion in the revised manuscript.

(2) Knockdown of Selected HATs and HDAC3 in cells

The authors should perform gene knockdown experiments in cells, targeting the identified HATs and HDAC3, followed by Western blot and mass spectrometry analysis of Kacac expression levels. This would validate whether the findings from the in vitro assays are biologically relevant in cellular contexts.

We appreciate the reviewer’s valuable suggestion. Our identified HATs, including p300 and GCN5, were reported as acetoacetyltransferases in cellular contexts by a recent study (PMID: 37382194). Their findings are precisely consistent with our biochemical results, providing additional evidence that p300 and GCN5 mediate Kacac both in vitro and in vivo. In addition, inhibition of HDAC activity by SAHA greatly increased histone Kacac levels in HepG2 cells (PMID: 37382194), supporting the role of HDAC3 as an eraser responsible for Kacac removal. We plan to further study these enzymes’ contributions to Kacac through gene knockdown experiments and investigate the specific functions of enzyme-mediated Kacac under some pathological contexts.

Minor Comments:

(1) Abstract accuracy

In the Abstract, the authors state, "However, regulatory elements, substrate proteins, and epigenetic functions of Kacac remain unknown." Please revise this statement to align with the findings in Reference 22 and describe these elements more appropriately. If similar issues exist in other parts of the manuscript, please address them as well.

The issues have been addressed in the revised manuscript based on the reviewer's comments.

(2) Terminology issue

GCN5 and PCAF are both members of the GNAT family. It is not accurate to describe "GCN5/PCAF/HAT1" as one family. Please refine the terminology to reflect the classification accurately.

The description has been refined in the revised manuscript to accurately reflect the classification, in accordance with the reviewer's suggestion.

(3) Discussion on HBO1

Reference 22 has already established HBO1 as an acetoacetyltransferase. This paper should include a discussion of HBO1 alongside the screened p300, PCAF, and GCN5 to provide a more comprehensive perspective.

More discussion on HBO1 alongside the other screened HATs has been added in the revised manuscript.

eLife Assessment

This useful study explores the role of RAP2A in asymmetric cell division (ACD) regulation in glioblastoma stem cells (GSCs), drawing parallels to Drosophila ACD mechanisms and proposing that an imbalance toward symmetric divisions drives tumor progression. While findings on RAP2A's role in GSC expansion are promising, and the reviewers found the study innovative and technically sound, the study is nevertheless still considered incomplete because of its reliance on neurosphere models without in vivo confirmation and insufficient mechanistic validation. Addressing those gaps would substantiate the study's claims.

Reviewer #1 (Public review):

Summary:

The authors validate the contribution of RAP2A to GB progression. RAp2A participates in asymetric cell division, and the localization of several cell polarity markers including cno and Numb.

Strengths:

The use of human data, Drosophila models and cell culture or neurospheres is a good scenario to validate the hypothesis using complementary systems.

Moreover, the mechanisms that determine GB progression, and in particular glioma stem cells biology, are relevant for the knowledge on glioblastoma and opens new possibilities to future clinical strategies.

Weaknesses:

While the manuscript presents a well-supported investigation into RAP2A's role in GBM, some methodological aspects could benefit from further validation. The major concern is the reliance on a single GB cell line (GB5), including multiple GBM lines, particularly primary patient-derived 3D cultures with known stem-like properties, would significantly enhance the study's robustness.

Several specific points raised in previous reviews have improved this version of the manuscript:

• The specificity of Rap2l RNAi has been further confirmed by using several different RNAi tools.

• Quantification of phenotypic penetrance and survival rates in Rap2l mutants would help determine the consistency of ACD defects. The authors have substantially increased the number of samples analyzed including three different RNAi lines (both the number of NB lineages and the number of different brains analyzed) to confirm the high penetrance of the phenotype.

• The observations on neurosphere size and Ki-67 expression require normalization (e.g., Ki-67+ cells per total cell number or per neurosphere size). This is included in the manuscript and now clarified in the text.

• The discrepancy in Figures 6A and 6B requires further discussion. The authors have included a new analysis and further explanations and they can conclude that in 2 cell-neurospheres there are more cases of asymmetric divisions in the experimental condition (RAP2A) than in the control.

• Live imaging of ACD events would provide more direct evidence. Live imaging was not done due to technical limitations. Despite being a potential contribution to the manuscript, the current conclusions of the manuscript are supported by the current data, and live experiments can be dispensable

• Clarification of terminology and statistical markers (e.g., p-values) in Figure 1A would improve clarity. This has been improved.

Comments on revisions:

The manuscript has improved the clarity in general, and I think that it is suitable for publication. However, for future experiments and projects, I would like to insist in the relevance of validating the results in vivo using xenografts with 3D-primary patient-derived cell lines or GB organoids.

Reviewer #2 (Public review):

This study investigates the role of RAP2A in regulating asymmetric cell division (ACD) in glioblastoma stem cells (GSCs), bridging insights from Drosophila ACD mechanisms to human tumor biology. They focus on RAP2A, a human homolog of Drosophila Rap2l, as a novel ACD regulator in GBM is innovative, given its underexplored role in cancer stem cells (CSCs). The hypothesis that ACD imbalance (favoring symmetric divisions) drives GSC expansion and tumor progression introduces a fresh perspective on differentiation therapy. However, the dual role of ACD in tumor heterogeneity (potentially aiding therapy resistance) requires deeper discussion to clarify the study's unique contributions against existing controversies.

Comments on revisions:

More experiments as suggested in the original assessment of the submission are needed to justify the hypothesis drawn in the manuscript.

Author response:

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

Reviewer #1 (Public review):

Summary:

The authors validate the contribution of RAP2A to GB progression. RAp2A participates in asymmetric cell division, and the localization of several cell polarity markers, including cno and Numb.

Strengths:

The use of human data, Drosophila models, and cell culture or neurospheres is a good scenario to validate the hypothesis using complementary systems.

Moreover, the mechanisms that determine GB progression, and in particular glioma stem cells biology, are relevant for the knowledge on glioblastoma and opens new possibilities to future clinical strategies.

Weaknesses:

While the manuscript presents a well-supported investigation into RAP2A's role in GBM, several methodological aspects require further validation. The major concern is the reliance on a single GB cell line (GB5), which limits the generalizability of the findings. Including multiple GBM lines, particularly primary patient-derived 3D cultures with known stem-like properties, would significantly enhance the study's relevance.

Additionally, key mechanistic aspects remain underexplored. Further investigation into the conservation of the Rap2l-Cno/aPKC pathway in human cells through rescue experiments or protein interaction assays would be beneficial. Similarly, live imaging or lineage tracing would provide more direct evidence of ACD frequency, complementing the current indirect metrics (odd/even cell clusters, Numb asymmetry).

Several specific points require attention:

(1) The specificity of Rap2l RNAi needs further confirmation. Is Rap2l expressed in neuroblasts or intermediate neural progenitors? Can alternative validation methods be employed?

There are no available antibodies/tools to determine whether Rap2l is expressed in NB lineages, and we have not been able either to develop any. However, to further prove the specificity of the Rap2l phenotype, we have now analyzed two additional and independent RNAi lines of Rap2l along with the original RNAi line analyzed. We have validated the results observed with this line and found a similar phenotype in the two additional RNAi lines now analyzed. These results have been added to the text ("Results section", page 6, lines 142-148) and are shown in Supplementary Figure 3.

(2) Quantification of phenotypic penetrance and survival rates in Rap2l mutants would help determine the consistency of ACD defects.

In the experiment previously mentioned (repetition of the original Rap2l RNAi line analysis along with two additional Rap2l RNAi lines) we have substantially increased the number of samples analyzed (both the number of NB lineages and the number of different brains analyzed). With that, we have been able to determine that the penetrance of the phenotype was 100% or almost 100% in the 3 different RNAi lines analyzed (n>14 different brains/larvae analyzed in all cases). Details are shown in the text (page 6, lines 142-148), in Supplementary Figure 3 and in the corresponding figure legend.

(3) The observations on neurosphere size and Ki-67 expression require normalization (e.g., Ki-67+ cells per total cell number or per neurosphere size). Additionally, apoptosis should be assessed using Annexin V or TUNEL assays.

The experiment of Ki-67+ cells was done considering the % of Ki-67+ cells respect the total cell number in each neurosphere. In the "Materials and methods" section it is well indicated: "The number of Ki67+ cells with respect to the total number of nuclei labelled with DAPI within a given neurosphere were counted to calculate the Proliferative Index (PI), which was expressed as the % of Ki67+ cells over total DAPI+ cells"

Perhaps it was not clearly showed in the graph of Figure 5A. We have now changed it indicating: "% of Ki67+ cells/ neurosphere" in the "Y axis". 

Unfortunately, we currently cannot carry out neurosphere cultures to address the apoptosis experiments. 

(4) The discrepancy in Figures 6A and 6B requires further discussion.

We agree that those pictures can lead to confusion. In the analysis of the "% of neurospheres with even or odd number of cells", we included the neurospheres with 2 cells both in the control and in the experimental condition (RAP2A). The number of this "2 cell-neurospheres" was very similar in both conditions (27,7 % and 27 % of the total neurospheres analyzed in each condition), and they can be the result of a previous symmetric or asymmetric division, we cannot distinguish that (only when they are stained with Numb, for example, as shown in Figure 6B). As a consequence, in both the control and in the experimental condition, these 2-cell neurospheres included in the group of "even" (Figure 6A) can represent symmetric or asymmetric divisions. However, in the experiment shown in Figure 6B, it is shown that in these 2 cellneurospheres there are more cases of asymmetric divisions in the experimental condition (RAP2A) than in the control.

Nevertheless, to make more accurate and clearer the conclusions, we have reanalyzed the data taking into account only the neurospheres with 3-5-7 (as odd) or 4-6-8 (as even) cells. Likewise, we have now added further clarifications regarding the way the experiment has been analyzed in the methods.

(5) Live imaging of ACD events would provide more direct evidence.

We agree that live imaging would provide further evidence. Unfortunately, we currently cannot carry out neurosphere cultures to approach those experiments.

(6) Clarification of terminology and statistical markers (e.g., p-values) in Figure 1A would improve clarity.

We thank the reviewer for pointing out this issue. To improve clarity, we have now included a Supplementary Figure (Fig. S1) with the statistical parameters used. Additionally, we have performed a hierarchical clustering of genes showing significant or not-significant changes in their expression levels.

(7) Given the group's expertise, an alternative to mouse xenografts could be a Drosophila genetic model of glioblastoma, which would provide an in vivo validation system aligned with their research approach.

The established Drosophila genetic model of glioblastoma is an excellent model system to get deep insight into different aspects of human GBM. However, the main aim of our study was to determine whether an imbalance in the mode of stem cell division, favoring symmetric divisions, could contribute to the expansion of the tumor. We chose human GBM cell lines-derived neurospheres because in human GBM it has been demonstrated the existence of cancer stem cells (glioblastoma or glioma stem cells -GSCs--). And these GSCs, as all stem cells, can divide symmetric or asymmetrically. In the case of the Drosophila model of GBM, the neoplastic transformation observed after overexpressing the EGF receptor and PI3K signaling is due to the activation of downstream genes that promote cell cycle progression and inhibit cell cycle exit. It has also been suggested that the neoplastic cells in this model come from committed glial progenitors, not from stem-like cells.

With all, it would be difficult to conclude the causes of the potential effects of manipulating the Rap2l levels in this Drosophila system of GBM. We do not discard this analysis in the future (we have all the "set up" in the lab). However, this would probably imply a new project to comprehensively analyze and understand the mechanism by which Rap2l (and other ACD regulators) might be acting in this context, if it is having any effect. 

However, as we mentioned in the Discussion, we agree that the results we have obtained in this study must be definitely validated in vivo in the future using xenografts with 3D-primary patient-derived cell lines.

Reviewer #2 (Public review):

This study investigates the role of RAP2A in regulating asymmetric cell division (ACD) in glioblastoma stem cells (GSCs), bridging insights from Drosophila ACD mechanisms to human tumor biology. They focus on RAP2A, a human homolog of Drosophila Rap2l, as a novel ACD regulator in GBM is innovative, given its underexplored role in cancer stem cells (CSCs). The hypothesis that ACD imbalance (favoring symmetric divisions) drives GSC expansion and tumor progression introduces a fresh perspective on differentiation therapy. However, the dual role of ACD in tumor heterogeneity (potentially aiding therapy resistance) requires deeper discussion to clarify the study's unique contributions against existing controversies. Some limitations and questions need to be addressed.

(1) Validation of RAP2A's prognostic relevance using TCGA and Gravendeel cohorts strengthens clinical relevance. However, differential expression analysis across GBM subtypes (e.g., MES, DNA-methylation subtypes ) should be included to confirm specificity.

We have now included a Supplementary figure (Supplementary Figure 2), in which we show the analysis of RAP2A levels in the different GBM subtypes (proneural, mesenchymal and classical) and their prognostic relevance (i.e. the proneural subtype that presents RAP2A levels significantly higher than the others is the subtype that also shows better prognostic).

(2) Rap2l knockdown-induced ACD defects (e.g., mislocalization of Cno/Numb) are well-designed. However, phenotypic penetrance and survival rates of Rap2l mutants should be quantified to confirm consistency.

We have now analyzed two additional and independent RNAi lines of Rap2l along with the original RNAi line. We have validated the results observed with this line and found a similar phenotype in the two additional RNAi lines now analyzed. To determine the phenotypic penetrance, we have substantially increased the number of samples analyzed (both the number of NB lineages and the number of different brains analyzed). With that, we have been able to determine that the penetrance of the phenotype was 100% or almost 100% in the 3 different Rap2l RNAi lines analyzed (n>14 different brains/larvae analyzed in all cases). These results have been added to the text ("Results section", page 6, lines 142-148) and are shown in Supplementary Figure 3 and in the corresponding figure legend. 

(3) While GB5 cells were effectively used, justification for selecting this line (e.g., representativeness of GBM heterogeneity) is needed. Experiments in additional GBM lines (especially the addition of 3D primary patient-derived cell lines with known stem cell phenotype) would enhance generalizability.

We tried to explain this point in the paper (Results). As we mentioned, we tested six different GBM cell lines finding similar mRNA levels of RAP2A in all of them, and significantly lower levels than in control Astros (Fig. 3A). We decided to focus on the GBM cell line called GB5 as it grew well (better than the others) in neurosphere cell culture conditions, for further analyses. We agree that the addition of at least some of the analyses performed with the GB5 line using other lines (ideally in primary patientderive cell lines, as the reviewer mentions) would reinforce the results. Unfortunately, we cannot perform experiments in cell lines in the lab currently. We will consider all of this for future experiments.

(4) Indirect metrics (odd/even cell clusters, NUMB asymmetry) are suggestive but insufficient. Live imaging or lineage tracing would directly validate ACD frequency.

We agree that live imaging would provide further evidence. Unfortunately, we cannot approach those experiments in the lab currently.

(5) The initial microarray (n=7 GBM patients) is underpowered. While TCGA data mitigate this, the limitations of small cohorts should be explicitly addressed and need to be discussed.

We completely agree with this comment. We had available the microarray, so we used it as a first approach, just out of curiosity of knowing whether (and how) the levels of expression of those human homologs of Drosophila ACD regulators were affected in this small sample, just as starting point of the study. We were conscious of the limitations of this analysis and that is why we followed up the analysis in the datasets, on a bigger scale. We already mentioned the limitations of the array in the Discussion:

"The microarray we interrogated with GBM patient samples had some limitations. For example, not all the human genes homologs of the Drosophila ACD regulators were present (i.e. the human homologs of the determinant Numb). Likewise, we only tested seven different GBM patient samples. Nevertheless, the output from this analysis was enough to determine that most of the human genes tested in the array presented altered levels of expression"[....] In silico analyses, taking advantage of the existence of established datasets, such as the TCGA, can help to more robustly assess, in a bigger sample size, the relevance of those human genes expression levels in GBM progression, as we observed for the gene RAP2A."

(6) Conclusions rely heavily on neurosphere models. Xenograft experiments or patient-derived orthotopic models are critical to support translational relevance, and such basic research work needs to be included in journals.

We completely agree. As we already mentioned in the Discussion, the results we have obtained in this study must be definitely validated in vivo in the future using xenografts with 3D-primary patient-derived cell lines.

(7) How does RAP2A regulate NUMB asymmetry? Is the Drosophila Rap2l-Cno/aPKC pathway conserved? Rescue experiments (e.g., Cno/aPKC knockdown with RAP2A overexpression) or interaction assays (e.g., Co-IP) are needed to establish molecular mechanisms.

The mechanism by which RAP2A is regulating ACD is beyond the scope of this paper. We do not even know how Rap2l is acting in Drosophila to regulate ACD. In past years, we did analyze the function of another Drosophila small GTPase, Rap1 (homolog to human RAP1A) in ACD, and we determined the mechanism by which Rap1 was regulating ACD (including the localization of Numb): interacting physically with Cno and other small GTPases, such as Ral proteins, and in a complex with additional ACD regulators of the "apical complex" (aPKC and Par-6). Rap2l could be also interacting physically with the "Ras-association" domain of Cno (domain that binds small GTPases, such as Ras and Rap1). We have added some speculations regarding this subject in the Discussion:

"It would be of great interest in the future to determine the specific mechanism by which Rap2l/RAP2A is regulating this process. One possibility is that, as it occurs in the case of the Drosophila ACD regulator Rap1, Rap2l/RAP2A is physically interacting or in a complex with other relevant ACD modulators."

(8) Reduced stemness markers (CD133/SOX2/NESTIN) and proliferation (Ki-67) align with increased ACD. However, alternative explanations (e.g., differentiation or apoptosis) must be ruled out via GFAP/Tuj1 staining or Annexin V assays.

We agree with these possibilities.  Regarding differentiation, the potential presence of increased differentiation markers would be in fact a logic consequence of an increase in ACD divisions/reduced stemness markers. Unfortunately, we cannot approach those experiments in the lab currently.

(9) The link between low RAP2A and poor prognosis should be validated in multivariate analyses to exclude confounding factors (e.g., age, treatment history).

We have now added this information in the "Results section" (page 5, lines 114-123).

(10) The broader ACD regulatory network in GBM (e.g., roles of other homologs like NUMB) and potential synergies/independence from known suppressors (e.g., TRIM3) warrant exploration.

The present study was designed as a "proof-of-concept" study to start analyzing the hypothesis that the expression levels of human homologs of known Drosophila ACD regulators might be relevant in human cancers that contain cancer stem cells, if those human homologs were also involved in modulating the mode of (cancer) stem cell division. 

To extend the findings of this work to the whole ACD regulatory network would be the logic and ideal path to follow in the future.

We already mentioned this point in the Discussion:

"....it would be interesting to analyze in the future the potential consequences that altered levels of expression of the other human homologs in the array can have in the behavior of the GSCs. In silico analyses, taking advantage of the existence of established datasets, such as the TCGA, can help to more robustly assess, in a bigger sample size, the relevance of those human genes expression levels in GBM progression, as we observed for the gene RAP2A."

(11) The figures should be improved. Statistical significance markers (e.g., p-values) should be added to Figure 1A; timepoints/culture conditions should be clarified for Figure 6A.

Regarding the statistical significance markers, we have now included a Supplementary Figure (Fig. S1) with the statistical parameters used. Additionally, we have performed a hierarchical clustering of genes showing significant or notsignificant changes in their expression levels. 

Regarding the experimental conditions corresponding to Figure 6A, those have now been added in more detail in "Materials and Methods" ("Pair assay and Numb segregation analysis" paragraph).

(12) Redundant Drosophila background in the Discussion should be condensed; terminology should be unified (e.g., "neurosphere" vs. "cell cluster").

As we did not mention much about Drosophila ACD and NBs in the "Introduction", we needed to explain in the "Discussion" at least some very basic concepts and information about this, especially for "non-drosophilists". We have reviewed the Discussion to maintain this information to the minimum necessary.

We have also reviewed the terminology that the Reviewer mentions and have unified it.

Reviewer #1 (Recommendations for the authors):

To improve the manuscript's impact and quality, I would recommend:

(1) Expand Cell Line Validation: Include additional GBM cell lines, particularly primary patient-derived 3D cultures, to increase the robustness of the findings.

(2) Mechanistic Exploration: Further examine the conservation of the Rap2lCno/aPKC pathway in human cells using rescue experiments or protein interaction assays.

(3) Direct Evidence of ACD: Implement live imaging or lineage tracing approaches to strengthen conclusions on ACD frequency.

(4) RNAi Specificity Validation: Clarify Rap2l RNAi specificity and its expression in neuroblasts or intermediate neural progenitors.

(5) Quantitative Analysis: Improve quantification of neurosphere size, Ki-67 expression, and apoptosis to normalize findings.

(6) Figure Clarifications: Address inconsistencies in Figures 6A and 6B and refine statistical markers in Figure 1A.

(7) Alternative In Vivo Model: Consider leveraging a Drosophila glioblastoma model as a complementary in vivo validation approach.

Addressing these points will significantly enhance the manuscript's translational relevance and overall contribution to the field.

We have been able to address points 4, 5 and 6. Others are either out of the scope of this work (2) or we do not have the possibility to carry them out at this moment in the lab (1, 3 and 7). However, we will complete these requests/recommendations in other future investigations.

Reviewer #2 (Recommendations for the authors):

Major Revision /insufficient required to address methodological and mechanistic gaps.

(1) Enhance Clinical Relevance

Validate RAP2A's prognostic significance across multiple GBM subtypes (e.g., MES, DNA-methylation subtypes) using datasets like TCGA and Gravendeel to confirm specificity.

Perform multivariate survival analyses to rule out confounding factors (e.g., patient age, treatment history).

(2) Strengthen Mechanistic Insights

Investigate whether the Rap2l-Cno/aPKC pathway is conserved in human GBM through rescue experiments (e.g., RAP2A overexpression with Cno/aPKC knockdown) or interaction assays (e.g., Co-IP).

Use live-cell imaging or lineage tracing to directly validate ACD frequency instead of relying on indirect metrics (odd/even cell clusters, NUMB asymmetry).

(3) Improve Model Systems & Experimental Design

Justify the selection of GB5 cells and include additional GBM cell lines, particularly 3D primary patient-derived cell models, to enhance generalizability.

It is essential to perform xenograft or orthotopic patient-derived models to support translational relevance.

(5) Address Alternative Interpretations

Rule out other potential effects of RAP2A knockdown (e.g., differentiation or apoptosis) using GFAP/Tuj1 staining or Annexin V assays.

Explore the broader ACD regulatory network in GBM, including interactions with NUMB and TRIM3, to contextualize findings within known tumor-suppressive pathways.

(6) Improve Figures & Clarity

Add statistical significance markers (e.g., p-values) in Figure 1A and clarify timepoints/culture conditions for Figure 6A.

Condense redundant Drosophila background in the discussion and ensure consistent terminology (e.g., "neurosphere" vs. "cell cluster").

We have been able to address points 1, partially 3 and 6. Others are either out of the scope of this work or we do not have the possibility to carry them out at this moment in the lab. However, we are very interested in completing these requests/recommendations and we will approach that type of experiments in other future investigations.

Test annotation

eLife Assessment

This paper describes Unbend - a new method for measuring and correcting motions in cryo-EM images, with a particular emphasis on more challenging in situ samples such as lamella and whole cells. The method, which fits a B-spline model using cross-correlation-based local patch alignment of micrograph frames, represents a valuable tool for the cryo-EM community. The authors elegantly use 2D template matching to provide solid evidence that Unbend outperforms the previously reported method of Unblur by the same authors. The paper would benefit from the inclusion of a similar analysis for established alternative methods, such as MotionCor2.

Reviewer #1 (Public review):

Kong et al.'s work describes a new approach that does exactly what the title states: "Correction of local beam-induced sample motion in cryo-EM images using a 3D spline model." I find the method appropriate, logical, and well-explained. Additionally, the work suggests using 2DTM-related measurements to quantify the improvement of the new method compared to the old one in cisTEM, Unblur. I find this part engaging; it is straightforward, accurate, and, of course, the group has a strong command of 2DTM, presenting a thorough study.

However, everything in the paper (except some correct general references) refers to comparisons with the full-frame approach, Unblur. Still, we have known for more than a decade that local correction approaches perform better than global ones, so I do not find anything truly novel in their proposal of using local methods (the method itself- Unbend- is new, but many others have been described previously). In fact, the use of 2DTM is perhaps a more interesting novelty of the work, and here, a more systematic study comparing different methods with these proposed well-defined metrics would be very valuable. As currently presented, there is no doubt that it is better than an older, well-established approach, and the way to measure "better" is very interesting, but there is no indication of how the situation stands regarding newer methods.

Regarding practical aspects, it seems that the current implementation of the method is significantly slower than other patch-based approaches. If its results are shown to exceed those of existing local methods, then exploring the use of Unbend, possibly optimizing its code first, could be a valuable task. However, without more recent comparisons, the impact of Unbend remains unclear.

Reviewer #2 (Public review):

Summary:

The authors present a new method, Unbend, for measuring motion in cryo-EM images, with a particular emphasis on more challenging in situ samples such as lamella and whole cells<br /> (that can be more prone to overall motion and/or variability in motion across a field of view). Building on their previous approach of full-frame alignment (Unblur), they now perform full-frame alignment followed by patch alignment, and then use these outputs to generate a 3D cubic spline model of the motion. This model allows them to estimate a continuous, per-pixel shift field for each movie frame that aims to better describe complex motions and so ultimately generate improved motion-corrected micrographs. Performance of Unbend is evaluated using the 2D template matching (2DTM) method developed previously by the lab, and results are compared to using full-frame correction alone. Several different in situ samples are used for evaluation, covering a broad range that will be of interest to the rapidly growing in situ cryo-EM community.

Strengths:

The method appears to be an elegant way of describing complex motions in cryo-EM samples, and the authors present convincing data that Unbend generally improves SNR of aligned micrographs as well as increases detection of particles matching the 60S ribosome template when compared to using full-frame correction alone. The authors also give interesting insights into how different areas of a lamella behave with respect to motion by using Unbend on a montage dataset collected previously by the group. There is growing interest in imaging larger areas of in situ samples at high resolution, and these insights contribute valuable knowledge. Additionally, the availability of data collected in this study through the EMPIAR repository will be much appreciated by the field.

Weaknesses:

While the improvements with Unbend vs. Unblur appear clear, it is less obvious whether Unbend provides substantial gains over patch motion correction alone (the current norm in the field). It might be helpful for readers if this comparison were investigated for the in situ datasets. Additionally, the authors are open that in cases where full motion correction already does a good job, the extra degrees of freedom in Unbend can perhaps overfit the motions, making the corrections ultimately worse. I wonder if an adaptive approach could be explored, for example, using the readout from full-frame or patch correction to decide whether a movie should proceed to the full Unbend pipeline, or whether correction should stop at the patch estimation stage.

Reviewer #3 (Public review):

Summary

Kong and coauthors describe and implement a method to correct local deformations due to beam-induced motion in cryo-EM movie frames. This is done by fitting a 3D spline model to a stack of micrograph frames using cross-correlation-based local patch alignment to describe the deformations across the micrograph in each frame, and then computing the value of the deformed micrograph at each pixel by interpolating the undeformed micrograph at the displacement positions given by the spline model. A graphical interface in cisTEM allows the user to visualise the deformations in the sample, and the method has been proven to be successful by showing improvements in 2D template matching (2DTM) results on the corrected micrographs using five in situ samples.

Impact

This method has great potential to further streamline the cryo-EM single particle analysis pipeline by shortening the required processing time as a result of obtaining higher quality particles early in the pipeline, and is applicable to both old and new datasets, therefore being relevant to all cryo-EM users.

Strengths

(1) One key idea of the paper is that local beam induced motion affects frames continuously in space (in the image plane) as well as in time (along the frame stack), so one can obtain improvements in the image quality by correcting such deformations in a continuous way (deformations vary continuously from pixel to pixel and from frame to frame) rather than based on local discrete patches only. 3D splines are used to model the deformations: they are initialised using local patch alignments and further refined using cross-correlation between individual patch frames and the average of the other frames in the same patch stack.

(2) Another strength of the paper is using 2DTM to show that correcting such deformations continuously using the proposed method does indeed lead to improvements. This is shown using five in situ datasets, where local motion is quantified using statistics based on the estimated motions of ribosomes.

Weaknesses

(1) While very interesting, it is not clear how the proposed method using 3D splines for estimating local deformations compares with other existing methods that also aim to correct local beam-induced motion by approximating the deformations throughout the frames using other types of approximation, such as polynomials, as done, for example MotionCor2.

(2) The use of 2DTM is appropriate, and the results of the analysis are enlightening, but one shortcoming is that some relevant technical details are missing. For example, the 2DTM SNR is not defined in the article, and it is not clear how the authors ensured that no false positives were included in the particles counted before and after deformation correction. The Jupyter notebooks where this analysis was performed have not been made publicly available.

(3) It is also not clear how the proposed deformation correction method is affected by CTF defocus in the different samples (are the defocus values used in the different datasets similar or significantly different?) or if there is any effect at all.

Remove this line

eLife Assessment

This study identifies the Periportal Lamellar Complex (PLC), an important new structure revealed by a novel 3D imaging method. However, the evidence supporting its distinct cellular identity and functional role is currently incomplete, as it relies on transcriptomic re-analysis and correlation without direct experimental validation. Addressing the key issues of methodological rigor and providing functional evidence is essential to fully substantiate these significant claims.

Reviewer #1 (Public review):

Summary:

In this manuscript, Chengjian Zhao et al. focused on the interactions between vascular, biliary, and neural networks in the liver microenvironment, addressing the critical bottleneck that the lack of high-resolution 3D visualization has hindered understanding of these interactions in liver disease.

Strengths:

This study developed a high-resolution multiplex 3D imaging method that integrates multicolor metallic compound nanoparticle (MCNP) perfusion with optimized CUBIC tissue clearing. This method enables the simultaneous 3D visualization of spatial networks of the portal vein, hepatic artery, bile ducts, and central vein in the mouse liver. The authors reported a perivascular structure termed the Periportal Lamellar Complex (PLC), which is identified along the portal vein axis. This study clarifies that the PLC comprises CD34⁺Sca-1⁺ dual-positive endothelial cells with a distinct gene expression profile, and reveals its colocalization with terminal bile duct branches and sympathetic nerve fibers under physiological conditions.

Weaknesses:

This manuscript is well-written, organized, and informative. However, there are some points that need to be clarified.

(1) After MCNP-dye injection, does it remain in the blood vessels, adsorb onto the cell surface, or permeate into the cells? Does the MCNP-dye have cell selectivity?

(2) All MCNP-dyes were injected after the mice were sacrificed, and the mice's livers were fixed with PFA. After the blood flow had ceased, how did the authors ensure that the MCNP-dyes were fully and uniformly perfused into the microcirculation of the liver?

(3) It is advisable to present additional 3D perspective views in the article, as the current images exhibit very weak 3D effects. Furthermore, it would be better to supplement with some videos to demonstrate the 3D effects of the stained blood vessels.

(4) In Figure 1-I, the authors used MCNP-Black to stain the central veins; however, in addition to black, there are also yellow and red stains in the image. The authors need to explain what these stains are in the legend.

(5) There is a typo in the title of Figure 4F; it should be "stem cell".

(6) Nuclear staining is necessary in immunofluorescence staining, especially for Figure 5e. This will help readers distinguish whether the green color in the image corresponds to cells or dye deposits.

Reviewer #2 (Public review):

Summary:

The present manuscript of Xu et al. reports a novel clearing and imaging method focusing on the liver. The authors simultaneously visualized the portal vein, hepatic artery, central vein, and bile duct systems by injecting metal compound nanoparticles (MCNPs) with different colors into the portal vein, heart left ventricle, inferior vena cava, and the extrahepatic bile duct, respectively. The method involves: trans-cardiac perfusion with 4% PFA, the injection of MCNPs with different colors, clearing with the modified CUBIC method, cutting 200 micrometer thick slices by vibratome, and then microscopic imaging. The authors also perform various immunostaining (DAB or TSA signal amplification methods) on the tissue slices from MCNP-perfused tissue blocks. With the application of this methodical approach, the authors report dense and very fine vascular branches along the portal vein. The authors name them as 'periportal lamellar complex (PLC)' and report that PLC fine branches are directly connected to the sinusoids. The authors also claim that these structures co-localize with terminal bile duct branches and sympathetic nerve fibers, and contain endothelial cells with a distinct gene expression profile. Finally, the authors claim that PLC-s proliferate in liver fibrosis (CCl4 model) and act as a scaffold for proliferating bile ducts in ductular reaction and for ectopic parenchymal sympathetic nerve sprouting.

Strengths:

The simultaneous visualization of different hepatic vascular compartments and their combination with immunostaining is a potentially interesting novel methodological approach.

Weaknesses:

This reviewer has several concerns about the validity of the microscopic/morphological findings as well as the transcriptomics results. In this reviewer's opinion, the introduction contains overstatements regarding the potential of the method, there are severe caveats in the method descriptions, and several parts of the Results are not fully supported by the documentation. Thus, the conclusions of the paper may be critically viewed in their present form and may need reconsideration by the authors.

Reviewer #3 (Public review):

Summary:

In the reviewed manuscript, researchers aimed to overcome the obstacles of high-resolution imaging of intact liver tissue. They report successful modification of the existing CUBIC protocol into Liver-CUBIC, a high-resolution multiplex 3D imaging method that integrates multicolor metallic compound nanoparticle (MCNP) perfusion with optimized liver tissue clearing, significantly reducing clearing time and enabling simultaneous 3D visualization of the portal vein, hepatic artery, bile ducts, and central vein spatial networks in the mouse liver. Using this novel platform, the researchers describe a previously unrecognized perivascular structure they termed Periportal Lamellar Complex (PLC), regularly distributed along the portal vein axis. The PLC originates from the portal vein and is characterized by a unique population of CD34⁺Sca-1⁺ dual-positive endothelial cells. Using available scRNAseq data, the authors assessed the CD34⁺Sca-1⁺ cells' expression profile, highlighting the mRNA presence of genes linked to neurodevelopment, biliary function, and hematopoietic niche potential. Different aspects of this analysis were then addressed by protein staining of selected marker proteins in the mouse liver tissue. Next, the authors addressed how the PLC and biliary system react to CCL4-induced liver fibrosis, implying PLC dynamically extends, acting as a scaffold that guides the migration and expansion of terminal bile ducts and sympathetic nerve fibers into the hepatic parenchyma upon injury.

The work clearly demonstrates the usefulness of the Liver-CUBIC technique and the improvement of both resolution and complexity of the information, gained by simultaneous visualization of multiple vascular and biliary systems of the liver at the same time. The identification of PLC and the interpretation of its function represent an intriguing set of observations that will surely attract the attention of liver biologists as well as hepatologists; however, some claims need more thorough assessment by functional experimental approaches to decipher the functional molecules and the sequence of events before establishing the PLC as the key hub governing the activity of biliary, arterial, and neuronal liver systems. Similarly, the level of detail of the methods section does not appear to be sufficient to exactly recapitulate the performed experiments, which is of concern, given that the new technique is a cornerstone of the manuscript.

Nevertheless, the work does bring a clear new insight into the liver structure and functional units and greatly improves the methodological toolbox to study it even further, and thus fully deserves the attention of readers.

Strengths:

The authors clearly demonstrate an improved technique tailored to the visualization of the liver vasulo-biliary architecture in unprecedented resolution.

This work proposes a new biological framework between the portal vein, hepatic arteries, biliary tree, and intrahepatic innervation, centered at previously underappreciated protrusions of the portal veins - the Periportal Lamellar Complexes (PLCs).

Weaknesses:

Possible overinterpretation of the CD34+Sca1+ findings was built on re-analysis of one scRNAseq dataset.

Lack of detail in the materials and methods section greatly limits the usefulness of the new technique to other researchers.

Remove the text

Ignore this

eLife Assessment

This study presents valuable findings on the role of KLF6 in in vitro endothelial cells exposed to altered (high or low) shear stress with a customized microfluidic device to investigate mechanisms of atherosclerosis. The finding that altered shear stress results in endothelial cell ferroptosis through reduced expression of KLF6 is compelling and adds a new layer of complexity to the pathogenesis of atherosclerotic plaques. However, the inclusion of an arterial cell line and re-evaluation of the statistical tests used would strengthen the authors' conclusions.

Reviewer #1 (Public review):

Summary:

The authors used an in vitro microfluidic system where HUVECs are exposed to high, low, or physiologic (normal) shear stress to demonstrate that both high and low shear stress for 24 hours resulted in decreased KLF6 expression, decreased lipid peroxidation, and increased cell death, which was reversible upon treatment with Fer-1, the ferroptosis inhibitor. RNA sequencing (LSS vs normal SS) revealed decreased steroid synthesis and UPR signaling in low shear stress conditions, which they confirmed by showing reduced expression of proteins that mitigate ER stress under both LSS and HSS. Decreased KLF6 expression after exposure to HSS/LSS was associated with decreased expression of regulators of ER stress (PERK, BiP, MVD), which was restored with KLF6 overexpression. Overexpression of KLF6 also restored SLC7A11 expression, Coq10, and reduced c11 bodipy oxidation state- all markers of lipid peroxidation and ferroptosis. The authors then used vascular smooth muscle cells (atherosclerotic model) with HUVECs and monocytes to show that KLF6 overexpression reduces the adhesion of monocytes and lipid accumulation in conditions of low shear stress.

Strengths:

(1) The use of a microfluidic device to simulate shear stress while keeping the pressure constant when varying the shear stress applied is improved and more physiologic compared to traditional cone and shearing devices. Similarly, the utilization of both low and high shear stress in most experiments is a strength.

(2) This study provides a link between disturbed shear stress and ferroptosis, which is novel, and fits nicely with existing knowledge that endothelial cell ferroptosis promotes atherosclerosis. This concept was also recently reported in September 2025, when a publication also demonstrated that LSS triggers ferroptosis in vascular endothelial cells (PMID: 40939914), which partly validates these findings.

Weaknesses:

(1) While HUVECs are commonly used in endothelial in vitro studies, it would be preferable to confirm the findings using an arterial cell line, such as human coronary artery cells, when studying mechanisms of early atherosclerosis. Furthermore, physiologic arterial shear stress is higher than venous shear stress, and different vascular beds have varying responses to altered shear stress; as such, the up- and downregulated pathways in HUVECs should be confirmed in an arterial system.

(2) The authors provide convincing evidence of disturbances in shear stress inducing endothelial ferroptosis with assays for impaired lipid peroxidation and increased cell death that was reversed with a ferroptosis inhibitor. However, more detailed characterization of ferroptosis with iron accumulation assays, as well as evaluating GPX4 activity as a consequence of the impaired mevalonate pathway, and testing for concomitant apoptosis in addition to ferroptosis, would add to the data.

(3) The authors state that KLF2 and KLF4 are not amongst the differentially expressed genes downregulated by reduced shear stress, which is contrary to previous data, where both KLF2 and KLF4 are well studied to be upregulated by physiologic laminar shear stress. While this might be due to the added pressure in their microfluidic system, it also might be due to changes in gene expression over time. In this case, a time course experiment would be needed. It is possible that KLF2, KLF4 and KLF6 are all reduced in low (and high) shear stress and cooperatively regulate the endothelial cell phenotype. Both KLF2 and KLF4 have been shown to be protective against atherosclerosis.

Reviewer #2 (Public review):

Summary:

The manuscript by Cui et al. titled "abnormal shear stress induces ferroptosis in endothelial cells via KLF6 downregulation" investigated in a microfluidic device the effect of 24-hour low, medium, and high shear stress levels upon human vein endothelial cells. The authors found that KLF6 is an important regulator of endothelial cell ferroptosis through the BiP-PERK-Slc7a11 and MVD-ID11-CoQ10 axis under both low and high shear stress, postulating this may explain the spatial preference of atherosclerosis at bifurcations of the arteries.

Strengths:

The main strength of the study is the use of a microfluidic device within which the authors could vary the shear stress (low, medium, high), whilst keeping fluid pressure near the physiological range of 70 mmHg. Deciding to focus on transcription factors that respond to shear stress, the authors found KLF6 in their dataset, for which they provide compelling evidence that endothelial cell ferroptosis is triggered by both excessive and insufficient shear stress, inversely correlating with KLF6 expression. Importantly, it was demonstrated that cell death in endothelial cells during HSS and LSS was prevented through the addition of Fer-1, supporting the role of ferroptosis. Moreso, the importance of KLF6 as an essential regulator was demonstrated through KLF6 overexpression.

Weaknesses:

There are some major concerns with the results:

(1) Inappropriate statistical tests were used (i.e., an unpaired t-test cannot be used to compare more than two groups).<br /> (2) Inconsistencies in western blot normalization as different proteins seem to have been used (GAPDH and B-actin) without specifying which is used when and why this differs.<br /> (3) Absence of transcriptomic analysis on HSS-exposed endothelial cells (which is not explained).

Moreso, the conclusions are predominantly based on an in vitro microfluidic chip model seeded with HUVECs. Although providing mechanistic insight into the effects of shear stress on (venous) endothelial cells, it does not recapitulate the in vivo complexity. The absence of validation (a.o. levels of KLF6) in clinical samples and/or animal models limits the translatability of the reported findings towards atherosclerosis. Among others, assessing the spatial heterogeneity of KLF6 abundance in atherosclerotic plaques depending on its proximity to arterial bifurcations may be interesting.

Points to be addressed:

(1) As a statistical test, the authors report having used unpaired t-tests; however, often three groups are compared for which t-tests are inadequate. This is faulty as, amongst other things, it does not take multiple comparison testing into account.

(2) Both B-Actin and GAPDH seem to have been used for protein-level normalization. Why? The Figure 2HL first panel reports B-actin, whereas the other three report GAPDH. The same applies to Figures 3E-F, where both are shown, and it is not mentioned which of the two has been used. Moreso, uncropped blots seem to be unavailable as supplementary data for proper review. These should be provided as supplementary data.

(3) LSS and MSS were compared based on transcriptomic analysis. Conversely, RNA sequencing was not reported for the HSS. Why is this data missing? It would be valuable to assess transcriptomics following HSS, and also to allow transcriptomic comparison of LSS and HSS.

(4) Actual sample sizes should be reported rather than "three or more". Moreso, it would be beneficial to show individual datapoints in bar graphs rather than only mean with SD if sample sizes are below 10 (e.g., Figures 1B-H, Figure 2G, etc.).

(5) The authors claim that by modifying the thickness of the middle layer, shear stress could be modified, whilst claiming to keep on-site pressure within physiological ranges (approx. 70 mmHg) as a hallmark of their microfluidic devices. Has it been experimentally verified that pressures indeed remain around 70 mmHg?

(6) A coculture model (VSMC, EC, monocytes) is mentioned in the last part of the results section without any further information. Information on this model should be provided in the methods section (seeding, cell numbers, etc.). Moreover, comparison of LSS vs LSS+KLF6 OE and HSS vs HSS+KLF6 OE is shown. It would benefit the interpretation of the outcomes if MSS were also shown. I twould also be beneficial to demonstrate differences between LSS, MSS, and HSS in this coculture model (without KLF6 OE).

(7) The experiments were solely performed with a venous endothelial cell line (HUVECs). Was the use of an arterial endothelial cell line considered? It may translate better towards atherosclerosis, which occurs within arteries. HUVECs are not accustomed to the claimed near-physiological pressures.

ballroom, full body, solo, wrinkled_skin, mature woman, scoop_neck, (medium_dress:1.6), red_and_sleeveless_dress, bow on dress, white shirt under dress, long_sleeves_shirt, lace_trim_shirt

https://danbooru.donmai.us/posts/10170089?q=plap

https://danbooru.donmai.us/posts/9868480?q=bed_invitation+

https://danbooru.donmai.us/posts/9875047?q=bed_invitation+

意识处理能力的局限——人类在进行“智能”或“有意识”的活动（如弹钢琴或专注阅读）时，最大信息处理能力仅为 50比特每秒

CA ne mre va pas. !

DOI: 10.1038/s41598-020-68424-1

Resource: Bloomington Drosophila Stock Center (RRID:SCR_006457)

Curator: @bdscstockkeepers

SciCrunch record: RRID:SCR_006457

What is this?

DOI: 10.1016/j.celrep.2020.107866

Resource: (Abcam Cat# ab13970, RRID:AB_300798)

Curator: @scibot

SciCrunch record: RRID:AB_300798

What is this?

DOI: 10.1016/j.celrep.2020.107866

Resource: (DSHB Cat# nc82, RRID:AB_2314866)

Curator: @scibot

SciCrunch record: RRID:AB_2314866

What is this?

DOI: 10.1016/j.celrep.2020.107866

Resource: (Jackson ImmunoResearch Labs Cat# 123-605-021, RRID:AB_2338967)

Curator: @scibot

SciCrunch record: RRID:AB_2338967

What is this?

DOI: 10.1016/j.celrep.2020.107866

Resource: (DSHB Cat# 8B4D2 (MH2B), RRID:AB_528269)

Curator: @scibot

SciCrunch record: RRID:AB_528269

What is this?

DOI: 10.1016/j.celrep.2020.107866

Resource: (Millipore Cat# CP06, RRID:AB_2617116)

Curator: @scibot

SciCrunch record: RRID:AB_2617116

What is this?

Registering with l'Urssaf as an "artiste-auteur" (or micro-entrepreneur), which aligns one with a specific régime for the payment of social charges, automate registration with CPAM, the administrative arm of the French universal healthcare system.

eLife Assessment

This important study provides new insights into the synchronization of ripple oscillations in the hippocampus, both within and across hemispheres. Using carefully designed statistical methods, it presents compelling evidence that synchrony is significantly higher within a hemisphere than across. This study will be of interest to neuroscientists studying the hippocampus and memory.

Reviewer #2 (Public review):

Summary

The authors completed a statistically rigorous analysis of the synchronization of sharp-wave ripples in the hippocampal CA1 across and within hemispheres. They used a publicly available dataset (collected in the Buzsaki lab) from 4 rats (8 sessions) recorded with silicon probes in both hemispheres. Each session contained approximately 8 hours of activity recorded during rest. The authors found that the characteristics of ripples did not differ between hemispheres, and that most ripples occurred almost simultaneously on all probe shanks within a hemisphere as well as across hemispheres. The differences in amplitude and exact timing of ripples between recording sites increased slightly with distance between recording sites. However, the phase coupling of ripples (in the 100-250 Hz range), changed dramatically with distance between recording sites. Ripples in opposite hemispheres were about 90% less coupled than ripples on nearby tetrodes in the same hemisphere. Phase coupling also decreased with distance within the hemisphere. Finally, pyramidal cell and interneuron spikes were coupled to the local ripple phase and less so to ripples at distant sites or the opposite hemisphere.

The authors also analyzed the changes in ripple coupling in relation to a couple of behavioral variables. Interestingly, while exposure to a novel track increased ripple abundance by ~5%, it did not change any form of ripple coupling within or between hemispheres.

Strengths

The analysis was well-designed and rigorous. The authors used statistical tests well suited to the hypotheses being tested, and clearly explained these tests. The paper is very clearly written, making it easy to understand and reproduce the analysis. The authors included an excellent review of the literature to explain the motivation for their study.

Weaknesses

The authors have addressed all of my concerns and recommendations.

This paper presents an important and unique analysis of ripple coupling. The same method could be used in the future to analyze the effects of other behavioral variables, such as satiety versus hunger, sleep deprivation, or enrichment, to address potential functions and causes of ripple coupling.

Author response:

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

Reviewer #1 (Public review):

Summary:

In this manuscript, the authors analyze electrophysiological data recorded bilaterally from the rat hippocampus to investigate the coupling of ripple oscillations across the hemispheres. Commensurate with the majority of previous research, the authors report that ripples tend to co-occur across both hemispheres. Specifically, the amplitude of ripples across hemispheres is correlated but their phase is not. These data corroborate existing models of ripple generation suggesting that CA3 inputs (coordinated across hemispheres via the commisural fibers) drive the sharp-wave component while the individual ripple waves are the result of local interactions between pyramidal cells and interneurons in CA1.

Strengths:

The manuscript is well-written, the analyses well-executed and the claims are supported by the data.

Weaknesses:

One question left unanswered by this study is whether information encoded by the right and left hippocampi is correlated.

Thank you for raising this important point. While our study demonstrates ripple co-occurrence across hemispheres, we did not directly assess whether the information encoded in each hippocampus is correlated. Addressing this question would require analyses of coordinated activity patterns, such as neuronal assemblies formed during novelty exposure, which falls beyond the scope of the present study. However, we agree this is an important avenue for future work, and we now acknowledge this limitation and outlined it as a future direction in the Conclusion section (lines 796–802).

Reviewer #2 (Public review):

Summary:

The authors completed a statistically rigorous analysis of the synchronization of sharp-wave ripples in the hippocampal CA1 across and within hemispheres. They used a publicly available dataset (collected in the Buzsaki lab) from 4 rats (8 sessions) recorded with silicon probes in both hemispheres. Each session contained approximately 8 hours of activity recorded during rest. The authors found that the characteristics of ripples did not differ between hemispheres, and that most ripples occurred almost simultaneously on all probe shanks within a hemisphere as well as across hemispheres. The differences in amplitude and exact timing of ripples between recording sites increased slightly with the distance between recording sites. However, the phase coupling of ripples (in the 100-250 Hz range), changed dramatically with the distance between recording sites. Ripples in opposite hemispheres were about 90% less coupled than ripples on nearby tetrodes in the same hemisphere. Phase coupling also decreased with distance within the hemisphere. Finally, pyramidal cell and interneuron spikes were coupled to the local ripple phase and less so to ripples at distant sites or the opposite hemisphere.

Strengths:

The analysis was well-designed and rigorous. The authors used statistical tests well suited to the hypotheses being tested, and clearly explained these tests. The paper is very clearly written, making it easy to understand and reproduce the analysis. The authors included an excellent review of the literature to explain the motivation for their study.

Weaknesses:

The authors state that their findings (highly coincident ripples between hemispheres), contradict other findings in the literature (in particular the study by Villalobos, Maldonado, and Valdes, 2017), but fail to explain why this large difference exists. They seem to imply that the previous study was flawed, without examining the differences between the studies.

The paper fails to mention the context in which the data was collected (the behavior the animals performed before and after the analyzed data), which may in fact have a large impact on the results and explain the differences between the current study and that by Villalobos et al. The Buzsaki lab data includes mice running laps in a novel environment in the middle of two rest sessions. Given that ripple occurrence is influenced by behavior, and that the neurons spiking during ripples are highly related to the prior behavioral task, it is likely that exposure to novelty changed the statistics of ripples. Thus, the authors should analyze the pre-behavior rest and post-behavior rest sessions separately. The Villalobos et al. data, in contrast, was collected without any intervening behavioral task or novelty (to my knowledge). Therefore, I predict that the opposing results are a result of the difference in recent experiences of the studied rats, and can actually give us insight into the memory function of ripples.

We appreciate this thoughtful hypothesis and have now addressed it explicitly. Our main analysis was conducted on 1-hour concatenated SWS epochs recorded before any novel environment exposure (baseline sleep). This was not clearly stated in the original manuscript, so we have now added a clarifying paragraph (lines 131–143). The main findings therefore remain unchanged.

To directly test the reviewer’s hypothesis, we performed the suggested comparison between pre- and post-maze rest sessions, including maze-type as a factor. These new analyses are now presented in a dedicated Results subsection (lines 475 - 493) and in Supplementary Figure 5.1. While we observed a modest increase in ripple abundance after the maze sessions — consistent with known experienced-dependent changes in ripple occurrence — the key findings of interhemispheric synchrony remained unchanged. Both pre- and post-maze sleep sessions showed robust bilateral time-locking of ripple events and similar dissociations between phase and amplitude coupling across hemispheres.

In one figure (5), the authors show data separated by session, rather than pooled. They should do this for other figures as well. There is a wide spread between sessions, which further suggests that the results are not as widely applicable as the authors seem to think. Do the sessions with small differences between phase coupling and amplitude coupling have low inter-hemispheric amplitude coupling, or high phase coupling? What is the difference between the sessions with low and high differences in phase vs. amplitude coupling? I noticed that the Buzsaki dataset contains data from rats running either on linear tracks (back and forth), or on circular tracks (unidirectionally). This could create a difference in inter-hemisphere coupling, because rats running on linear tracks would have the same sensory inputs to both hemispheres (when running in opposite directions), while rats running on a circular track would have different sensory inputs coming from the right and left (one side would include stimuli in the middle of the track, and the other would include closer views of the walls of the room). The synchronization between hemispheres might be impacted by how much overlap there was in sensory stimuli processed during the behavior epoch.

Thank you for this insightful suggestion. In our new analyses comparing pre- and post-maze sessions, we have also addressed this question. Supplementary Figures 4.1 and 5.1 (E-F) present coupling metrics averaged per session and include coding for maze type. Additionally, we have incorporated the reviewer’s hypothesis regarding sensory input differences and their potential impact on inter-hemispheric synchronization into a new Results subsection (lines 475–493).

The paper would be a lot stronger if the authors analyzed some of the differences between datasets, sessions, and epochs based on the task design, and wrote more about these issues. There may be more publicly available bi-hemispheric datasets to validate their results.

To further validate our findings, we have analyzed another publicly available dataset that includes bilateral CA1 recordings (https://crcns.org/data-sets/hc/hc-18). We have added a description of this dataset and our analysis approach in the Methods section (lines 119–125 and 144-145), and present the corresponding results in a new Supplementary Figure (Supplementary Figure 4.2). These new analyses replicated our main findings, confirming robust interhemispheric time-locking of ripple events and a greater dissociation between phase and amplitude coupling in ipsilateral versus contralateral recordings.

Reviewer #1 (Recommendations for the authors):

My only suggestion is that the introduction can be shortened. The authors discuss in great length literature linking ripples and memory, although the findings in the paper are not linked to memory. In addition, ripples have been implicated in non-mnemonic functions such as sleep and metabolic homeostasis.

The reviewer`s suggestion is valid and aligns with the main message of our paper. However, we believe that the relationship between ripples and memory has been extensively discussed in the literature, sometimes overshadowing other important functional roles (based on the reviewer’s comment, we now also refer to non-mnemonic functions of ripples in the revised introduction [lines 87–89]). Thus, we find it important to retain this context because highlighting the publication bias towards mnemonic interpretations helps frame the need for studies like ours that revisit still incompletely understood basic ripple mechanisms.

We also note that, based on a suggestion from reviewer 2, we have supplemented our manuscript with a new figure demonstrating ripple abundance increases during SWS following novel environment exposure (Supplementary Figure 5.1), linking it to memory and replicating the findings of Eschenko et al. (2008), though we present this result as a covariate, aimed at controlling for potential sources of variation in ripple synchronization.

Reviewer #2 (Recommendations for the authors):

It would be useful to include more information about the analyzed dataset in the methods section, e.g. how long were the recordings, how many datasets per rat, did the authors analyze the entire recording epoch or sub-divide it in any way, how many ripples were detected per recording (approximately).

We have now included more detailed information in the Methods section (lines 104 - 145).

A few of the references to sub-figures are mislabeled (e.g. lines 327-328).

Thank you for noticing these inconsistencies. We have carefully reviewed and corrected all figure sub-panel labels and references throughout the manuscript.

In Figure 7 C&D, are the neurons on the left sorted by contralateral ripple phase? It doesn't look like it. It would be easier to compare to ipsilateral if they were.

In Figures 7C and 7D, neurons are sorted by their ipsilateral peak ripple phase, with the contralateral data plotted using the same ordering to facilitate comparison. To avoid confusion, we have clarified this explicitly in the figure legend and corresponding main text (lines 544–550).

In Figure 6, using both bin sizes 50 and 100 doesn't contribute much.

We used both 50 ms and 100 ms bin sizes to directly compare with previous studies (Villalobos et al. 2017 used 5 ms and 100 ms; Csicsvari et al. 2000 used 5–50 ms). Because the proportion of coincident ripples is a non-decreasing function of the window size, larger bins can inflate coincidence measures. Including a mid-range bin of 50 ms allowed us to show that high coincidence levels are reached well before the 100 ms upper bound, supporting that the 100 ms window is not an overshoot. We have added clarification on this point in the Methods section on ripple coincidence (lines 204–212).

eLife Assessment

This important study combines EEG, neural networks and multivariate pattern analysis to show that real-world size, retinal size and real-world depth are represented at different latencies. The evidence presented is convincing and the work will be of broader interest to the experimental and computational vision community.

Reviewer #1 (Public review):

Lu & Golomb combined EEG, artificial neural networks, and multivariate pattern analyses to examine how different visual variables are processed in the brain. The conclusions of the paper are mostly well supported.

The authors find that not only real-world size is represented in the brain (which was known), but both retinal size and real-world depth is represented, at different time points or latencies, which may reflect different stages of processing. Prior work has not been able to answer the question of real-world depth due to stimuli used. The authors made this possible by assess real-world depth and testing it with appropriate methodology, accounting for retinal and real-world size. The methodological approach combining behavior, RSA, and ANNs is creative and well thought out to appropriately assess the research questions, and the findings may be very compelling if backed up with some clarifications and further analyses.

The work will be of interest to experimental and computational vision scientists, as well as the broader computational cognitive neuroscience community as the methodology is of interest and the code is or will be made available. The work is important as it is currently not clear what the correspondence between many deep neural network models are and the brain are, and this work pushes our knowledge forward on this front. Furthermore, the availability of methods and data will be useful for the scientific community.

Reviewer #3 (Public review):

The authors used an open EEG dataset of observers viewing real-world objects. Each object had a real-world size value (from human rankings), a retinal size value (measured from each image), and a scene depth value (inferred from the above). The authors combined the EEG and object measurements with extant, pre-trained models (a deep convolutional neural network, a multimodal ANN, and Word2vec) to assess the time course of processing object size (retinal and real-world) and depth. They found that depth was processed first, followed by retinal size, and then real-world size. The depth time course roughly corresponded to the visual ANNs, while the real-world size time course roughly corresponded to the more semantic models.

The time course result for the three object attributes is very clear and a novel contribution to the literature. The authors have revised the ANN motivations to increase clarity. Additionally, the authors have appropriately toned down some of the language about novelty, and the addition of a noise ceiling has helped the robustness of the work.

While I appreciate the addition of Cornet in the Supplement, I am less compelled by the authors' argument for Word2Vec over LLMs for "pure" semantic embeddings. While I'm not digging in on this point, this choice may prematurely age this work.

Author response:

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

Reviewer #1 (Public Review):

Lu & Golomb combined EEG, artificial neural networks, and multivariate pattern analyses to examine how different visual variables are processed in the brain. The conclusions of the paper are mostly well supported, but some aspects of methods and data analysis would benefit from clarification and potential extensions.

The authors find that not only real-world size is represented in the brain (which was known), but both retinal size and real-world depth are represented, at different time points or latencies, which may reflect different stages of processing. Prior work has not been able to answer the question of real-world depth due to the stimuli used. The authors made this possible by assessing real-world depth and testing it with appropriate methodology, accounting for retinal and real-world size. The methodological approach combining behavior, RSA, and ANNs is creative and well thought out to appropriately assess the research questions, and the findings may be very compelling if backed up with some clarifications and further analyses.

The work will be of interest to experimental and computational vision scientists, as well as the broader computational cognitive neuroscience community as the methodology is of interest and the code is or will be made available. The work is important as it is currently not clear what the correspondence between many deep neural network models and the brain is, and this work pushes our knowledge forward on this front. Furthermore, the availability of methods and data will be useful for the scientific community.

Reviewer #2 (Public Review):

Summary:

This paper aims to test if neural representations of images of objects in the human brain contain a 'pure' dimension of real-world size that is independent of retinal size or perceived depth. To this end, they apply representational similarity analysis on EEG responses in 10 human subjects to a set of 200 images from a publicly available database (THINGS-EEG2), correlating pairwise distinctions in evoked activity between images with pairwise differences in human ratings of real-world size (from THINGS+). By partialling out correlations with metrics of retinal size and perceived depth from the resulting EEG correlation time courses, the paper claims to identify an independent representation of real-world size starting at 170 ms in the EEG signal. Further comparisons with artificial neural networks and language embeddings lead the authors to claim this correlation reflects a relatively 'high-level' and 'stable' neural representation.

Strengths:

The paper features insightful figures/illustrations and clear figures.

The limitations of prior work motivating the current study are clearly explained and seem reasonable (although the rationale for why using 'ecological' stimuli with backgrounds matters when studying real-world size could be made clearer; one could also argue the opposite, that to get a 'pure' representation of the real-world size of an 'object concept', one should actually show objects in isolation).

The partial correlation analysis convincingly demonstrates how correlations between feature spaces can affect their correlations with EEG responses (and how taking into account these correlations can disentangle them better).

The RSA analysis and associated statistical methods appear solid.

Weaknesses:

The claim of methodological novelty is overblown. Comparing image metrics, behavioral measurements, and ANN activations against EEG using RSA is a commonly used approach to study neural object representations. The dataset size (200 test images from THINGS) is not particularly large, and neither is comparing pre-trained DNNs and language models, or using partial correlations.

Thanks for your feedback. We agree that the methods used in our study – such as RSA, partial correlations, and the use of pretrained ANN and language models – are indeed well-established in the literature. We therefore revised the manuscript to more carefully frame our contribution: rather than emphasizing methodological novelty in isolation, we now highlight the combination of techniques, the application to human EEG data with naturalistic images, and the explicit dissociation of real-world size, retinal size, and depth representations as the primary strengths of our approach. Corresponding language in the Abstract, Introduction, and Discussion has been adjusted to reflect this more precise positioning:

(Abstract, line 34 to 37) “our study combines human EEG and representational similarity analysis to disentangle neural representations of object real-world size from retinal size and perceived depth, leveraging recent datasets and modeling approaches to address challenges not fully resolved in previous work.”

(Introduction, line 104 to 106) “we overcome these challenges by combining human EEG recordings, naturalistic stimulus images, artificial neural networks, and computational modeling approaches including representational similarity analysis (RSA) and partial correlation analysis …”

(Introduction, line 108) “We applied our integrated computational approach to an open EEG dataset…”

(Introduction, line 142 to 143) “The integrated computational approach by cross-modal representational comparisons we take with the current study…”

(Discussion, line 550 to 552) “our study goes beyond the contributions of prior studies in several key ways, offering both theoretical and methodological advances: …”

The claims also seem too broad given the fairly small set of RDMs that are used here (3 size metrics, 4 ANN layers, 1 Word2Vec RDM): there are many aspects of object processing not studied here, so it's not correct to say this study provides a 'detailed and clear characterization of the object processing process'.

Thanks for pointing this out. We softened language in our manuscript to reflect that our findings provide a temporally resolved characterization of selected object features, rather than a comprehensive account of object processing:

(line 34 to 37) “our study combines human EEG and representational similarity analysis to disentangle neural representations of object real-world size from retinal size and perceived depth, leveraging recent datasets and modeling approaches to address challenges not fully resolved in previous work.”

(line 46 to 48) “Our research provides a temporally resolved characterization of how certain key object properties – such as object real-world size, depth, and retinal size – are represented in the brain, …”

The paper lacks an analysis demonstrating the validity of the real-world depth measure, which is here computed from the other two metrics by simply dividing them. The rationale and logic of this metric is not clearly explained. Is it intended to reflect the hypothesized egocentric distance to the object in the image if the person had in fact been 'inside' the image? How do we know this is valid? It would be helpful if the authors provided a validation of this metric.

We appreciate the comment regarding the real-world depth metric. Specifically, this metric was computed as the ratio of real-world size (obtained via behavioral ratings) to measured retinal size. The rationale behind this computation is grounded in the basic principles of perspective projection: for two objects subtending the same retinal size, the physically larger object is presumed to be farther away. This ratio thus serves as a proxy for perceived egocentric depth under the simplifying assumption of consistent viewing geometry across images.

We acknowledge that this is a derived estimate and not a direct measurement of perceived depth. While it provides a useful approximation that allows us to analytically dissociate the contributions of real-world size and depth in our RSA framework, we agree that future work would benefit from independent perceptual depth ratings to validate or refine this metric. We added more discussions about this to our revised manuscript:

(line 652 to 657) “Additionally, we acknowledge that our metric for real-world depth was derived indirectly as the ratio of perceived real-world size to retinal size. While this formulation is grounded in geometric principles of perspective projection and served the purpose of analytically dissociating depth from size in our RSA framework, it remains a proxy rather than a direct measure of perceived egocentric distance. Future work incorporating behavioral or psychophysical depth ratings would be valuable for validating and refining this metric.”

Given that there is only 1 image/concept here, the factor of real-world size may be confounded with other things, such as semantic category (e.g. buildings vs. tools). While the comparison of the real-world size metric appears to be effectively disentangled from retinal size and (the author's metric of) depth here, there are still many other object properties that are likely correlated with real-world size and therefore will confound identifying a 'pure' representation of real-world size in EEG. This could be addressed by adding more hypothesis RDMs reflecting different aspects of the images that may correlate with real-world size.

We thank the reviewer for this thoughtful and important point. We agree that semantic category and real-world size may be correlated, and that semantic structure is one of the plausible sources of variance contributing to real-world size representations. However, we would like to clarify that our original goal was to isolate real-world size from two key physical image features — retinal size and inferred real-world depth — which have been major confounds in prior work on this topic. We acknowledge that although our analysis disentangled real-world size from depth and retinal size, this does not imply a fully “pure” representation; therefore, we now refer to the real-world size representations as “partially disentangled” throughout the manuscript to reflect this nuance.

Interestingly, after controlling for these physical features, we still found a robust and statistically isolated representation of real-world size in the EEG signal. This motivated the idea that realworld size may be more than a purely perceptual or image-based property — it may be at least partially semantic. Supporting this interpretation, both the late layers of ANN models and the non-visual semantic model (Word2Vec) also captured real-world size structure. Rather than treating semantic information as an unwanted confound, we propose that semantic structure may be an inherent component of how the brain encodes real-world size.

To directly address the your concern, we conducted an additional variance partitioning analysis, in which we decomposed the variance in EEG RDMs explained by four RDMs: real-world depth, retinal size, real-world size, and semantic information (from Word2Vec). Specifically, for each EEG timepoint, we quantified (1) the unique variance of real-world size, after controlling for semantic similarity, depth, and retinal size; (2) the unique variance of semantic information, after controlling for real-world size, depth, and retinal size; (3) the shared variance jointly explained by real-world size and semantic similarity, controlling for depth and retinal size. This analysis revealed that real-world size explained unique variance in EEG even after accounting for semantic similarity. And there was also a substantial shared variance, indicating partial overlap between semantic structure and size. Semantic information also contributed unique explanatory power, as expected. These results suggest that real-world size is indeed partially semantic in nature, but also has independent neural representation not fully explained by general semantic similarity. This strengthens our conclusion that real-world size functions as a meaningful, higher-level dimension in object representation space.

We now include this new analysis and a corresponding figure (Figure S8) in the revised manuscript:

(line 532 to 539) “Second, we conducted a variance partitioning analysis, in which we decomposed the variance in EEG RDMs explained by three hypothesis-based RDMs and the semantic RDM (Word2Vec RDM), and we still found that real-world size explained unique variance in EEG even after accounting for semantic similarity (Figure S9). And we also observed a substantial shared variance jointly explained by real-world size and semantic similarity and a unique variance of semantic information. These results suggest that real-world size is indeed partially semantic in nature, but also has independent neural representation not fully explained by general semantic similarity.”

The choice of ANNs lacks a clear motivation. Why these two particular networks? Why pick only 2 somewhat arbitrary layers? If the goal is to identify more semantic representations using CLIP, the comparison between CLIP and vision-only ResNet should be done with models trained on the same training datasets (to exclude the effect of training dataset size & quality; cf Wang et al., 2023). This is necessary to substantiate the claims on page 19 which attributed the differences between models in terms of their EEG correlations to one of them being a 'visual model' vs. 'visual-semantic model'.

We argee that the choice and comparison of models should be better contextualized.

First, our motivation for selecting ResNet-50 and CLIP ResNet-50 was not to make a definitive comparison between model classes, but rather to include two widely used representatives of their respective categories—one trained purely on visual information (ResNet-50 on ImageNet) and one trained with joint visual and linguistic supervision (CLIP ResNet-50 on image–text pairs). These models are both highly influential and commonly used in computational and cognitive neuroscience, allowing for relevant comparisons with existing work (line 181-187).

Second, we recognize that limiting the EEG × ANN correlation analyses to only early and late layers may be viewed as insufficiently comprehensive. To address this point, we have computed the EEG correlations with multiple layers in both ResNet and CLIP models (ResNet: ResNet.maxpool, ResNet.layer1, ResNet.layer2, ResNet.layer3, ResNet.layer4, ResNet.avgpool; CLIP: CLIP.visual.avgpool, CLIP.visual.layer1, CLIP.visual.layer2, CLIP.visual.layer3, CLIP.visual.layer4, CLIP.visual.attnpool). The results, now included in Figure S4, show a consistent trend: early layers exhibit higher similarity to early EEG time points, and deeper layers show increased similarity to later EEG stages. We chose to highlight early and late layers in the main text to simplify interpretation.

Third, we appreciate the reviewer’s point that differences in training datasets (ImageNet vs. CLIP's dataset) may confound any attribution of differences in brain alignment to the models' architectural or learning differences. We agree that the comparisons between models trained on matched datasets (e.g., vision-only vs. multimodal models trained on the same image–text corpus) would allow for more rigorous conclusions. Thus, we explicitly acknowledged this limitation in the text:

(line 443 to 445) “However, it is also possible that these differences between ResNet and CLIP reflect differences in training data scale and domain.”

The first part of the claim on page 22 based on Figure 4 'The above results reveal that realworld size emerges with later peak neural latencies and in the later layers of ANNs, regardless of image background information' is not valid since no EEG results for images without backgrounds are shown (only ANNs).

We revised the sentence to clarify that this is a hypothesis based on the ANN results, not an empirical EEG finding:

(line 491 to 495) “These results show that real-world size emerges in the later layers of ANNs regardless of image background information, and – based on our prior EEG results – although we could not test object-only images in the EEG data, we hypothesize that a similar temporal profile would be observed in the brain, even for object-only images.”

While we only had the EEG data of human subjects viewing naturalistic images, the ANN results suggest that real-world size representations may still emerge at later processing stages even in the absence of background, consistent with what we observed in EEG under with-background conditions.

The paper is likely to impact the field by showcasing how using partial correlations in RSA is useful, rather than providing conclusive evidence regarding neural representations of objects and their sizes.

Additional context important to consider when interpreting this work:

Page 20, the authors point out similarities of peak correlations between models ('Interestingly, the peaks of significant time windows for the EEG × HYP RSA also correspond with the peaks of the EEG × ANN RSA timecourse (Figure 3D,F)'. Although not explicitly stated, this seems to imply that they infer from this that the ANN-EEG correlation might be driven by their representation of the hypothesized feature spaces. However this does not follow: in EEG-image metric model comparisons it is very typical to see multiple peaks, for any type of model, this simply reflects specific time points in EEG at which visual inputs (images) yield distinctive EEG amplitudes (perhaps due to stereotypical waves of neural processing?), but one cannot infer the information being processed is the same. To investigate this, one could for example conduct variance partitioning or commonality analysis to see if there is variance at these specific timepoints that is shared by a specific combination of the hypothesis and ANN feature spaces.

Thanks for your thoughtful observation! Upon reflection, we agree that the sentence – "Interestingly, the peaks of significant time windows for the EEG × HYP RSA also correspond with the peaks of the EEG × ANN RSA timecourse" – was speculative and risked implying a causal link that our data do not warrant. As you rightly points out, observing coincident peak latencies across different models does not necessarily imply shared representational content, given the stereotypical dynamics of evoked EEG responses. And we think even variance partitioning analysis would still not suffice to infer that ANN-EEG correlations are driven specifically by hypothesized feature spaces. Accordingly, we have removed this sentence from the manuscript to avoid overinterpretation. 

Page 22 mentions 'The significant time-window (90-300ms) of similarity between Word2Vec RDM and EEG RDMs (Figure 5B) contained the significant time-window of EEG x real-world size representational similarity (Figure 3B)'. This is not particularly meaningful given that the Word2Vec correlation is significant for the entire EEG epoch (from the time-point of the signal 'arriving' in visual cortex around ~90 ms) and is thus much less temporally specific than the realworld size EEG correlation. Again a stronger test of whether Word2Vec indeed captures neural representations of real-world size could be to identify EEG time-points at which there are unique Word2Vec correlations that are not explained by either ResNet or CLIP, and see if those timepoints share variance with the real-world size hypothesized RDM.

We appreciate your insightful comment. Upon reflection, we agree that the sentence – "'The significant time-window (90-300ms) of similarity between Word2Vec RDM and EEG RDMs (Figure 5B) contained the significant time-window of EEG x real-world size representational similarity (Figure 3B)" – was speculative. And we have removed this sentence from the manuscript to avoid overinterpretation. 

Additionally, we conducted two analyses as you suggested in the supplement. First, we calculated the partial correlation between EEG RDMs and the Word2Vec RDM while controlling for four ANN RDMs (ResNet early/late and CLIP early/late) (Figure S8). Even after regressing out these ANN-derived features, we observed significant correlations between Word2Vec and EEG RDMs in the 100–190 ms and 250–300 ms time windows. This result suggests that

Word2Vec captures semantic structure in the neural signal that is not accounted for by ResNet or CLIP. Second, we conducted an additional variance partitioning analysis, in which we decomposed the variance in EEG RDMs explained by four RDMs: real-world depth, retinal size, real-world size, and semantic information (from Word2Vec) (Figure S9). And we found significant shared variance between Word2Vec and real-world size at 130–150 ms and 180–250 ms. These results indicate a partially overlapping representational structure between semantic content and real-world size in the brain.

We also added these in our revised manuscript:

(line 525 to 539) “To further probe the relationship between real-world size and semantic information, and to examine whether Word2Vec captures variances in EEG signals beyond that explained by visual models, we conducted two additional analyses. First, we performed a partial correlation between EEG RDMs and the Word2Vec RDM, while regressing out four ANN RDMs (early and late layers of both ResNet and CLIP) (Figure S8). We found that semantic similarity remained significantly correlated with EEG signals across sustained time windows (100-190ms and 250-300ms), indicating that Word2Vec captures neural variance not fully explained by visual or visual-language models. Second, we conducted a variance partitioning analysis, in which we decomposed the variance in EEG RDMs explained by three hypothesis-based RDMs and the semantic RDM (Word2Vec RDM), and we still found that real-world size explained unique variance in EEG even after accounting for semantic similarity (Figure S9). And we also observed a substantial shared variance jointly explained by realworld size and semantic similarity and a unique variance of semantic information. These results suggest that real-world size is indeed partially semantic in nature, but also has independent neural representation not fully explained by general semantic similarity.”

Reviewer #3 (Public Review):

The authors used an open EEG dataset of observers viewing real-world objects. Each object had a real-world size value (from human rankings), a retinal size value (measured from each image), and a scene depth value (inferred from the above). The authors combined the EEG and object measurements with extant, pre-trained models (a deep convolutional neural network, a multimodal ANN, and Word2vec) to assess the time course of processing object size (retinal and real-world) and depth. They found that depth was processed first, followed by retinal size, and then real-world size. The depth time course roughly corresponded to the visual ANNs, while the real-world size time course roughly corresponded to the more semantic models.

The time course result for the three object attributes is very clear and a novel contribution to the literature. However, the motivations for the ANNs could be better developed, the manuscript could better link to existing theories and literature, and the ANN analysis could be modernized. I have some suggestions for improving specific methods.

(1) Manuscript motivations

The authors motivate the paper in several places by asking " whether biological and artificial systems represent object real-world size". This seems odd for a couple of reasons. Firstly, the brain must represent real-world size somehow, given that we can reason about this question. Second, given the large behavioral and fMRI literature on the topic, combined with the growing ANN literature, this seems like a foregone conclusion and undermines the novelty of this contribution.

Thanks for your helpful comment. We agree that asking whether the brain represents real-world size is not a novel question, given the existing behavioral and neuroimaging evidence supporting this. Our intended focus was not on the existence of real-world size representations per se, but the nature of these representations, particularly the relationship between the temporal dynamics and potential mechanisms of representations of real-world size versus other related perceptual properties (e.g., retinal size and real-world depth). We revised the relevant sentence to better reflect our focue, shifting from a binary framing (“whether or not size is represented”) to a more mechanistic and time-resolved inquiry (“how and when such representations emerge”):

(line 144 to 149) “Unraveling the internal representations of object size and depth features in both human brains and ANNs enables us to investigate how distinct spatial properties—retinal size, realworld depth, and real-world size—are encoded across systems, and to uncover the representational mechanisms and temporal dynamics through which real-world size emerges as a potentially higherlevel, semantically grounded feature.”

While the introduction further promises to "also investigate possible mechanisms of object realworld size representations.", I was left wishing for more in this department. The authors report correlations between neural activity and object attributes, as well as between neural activity and ANNs. It would be nice to link the results to theories of object processing (e.g., a feedforward sweep, such as DiCarlo and colleagues have suggested, versus a reverse hierarchy, such as suggested by Hochstein, among others). What is semantic about real-world size, and where might this information come from? (Although you may have to expand beyond the posterior electrodes to do this analysis).

We thank the reviewer for this insightful comment. We agree that understanding the mechanisms underlying real-world size representations is a critical question. While our current study does not directly test specific theoretical frameworks such as the feedforward sweep model or the reverse hierarchy theory, our results do offer several relevant insights: The temporal dynamics revealed by EEG—where real-world size emerges later than retinal size and depth—suggest that such representations likely arise beyond early visual feedforward stages, potentially involving higherlevel semantic processing. This interpretation is further supported by the fact that real-world size is strongly captured by late layers of ANNs and by a purely semantic model (Word2Vec), suggesting its dependence on learned conceptual knowledge.

While we acknowledge that our analyses were limited to posterior electrodes and thus cannot directly localize the cortical sources of these effects, we view this work as a first step toward bridging low-level perceptual features and higher-level semantic representations. We hope future work combining broader spatial sampling (e.g., anterior EEG sensors or source localization) and multimodal recordings (e.g., MEG, fMRI) can build on these findings to directly test competing models of object processing and representation hierarchy.

We also added these to the Discussion section:

(line 619 to 638) “Although our study does not directly test specific models of visual object processing, the observed temporal dynamics provide important constraints for theoretical interpretations. In particular, we find that real-world size representations emerge significantly later than low-level visual features such as retinal size and depth. This temporal profile is difficult to reconcile with a purely feedforward account of visual processing (e.g., DiCarlo et al., 2012), which posits that object properties are rapidly computed in a sequential hierarchy of increasingly complex visual features. Instead, our results are more consistent with frameworks that emphasize recurrent or top-down processing, such as the reverse hierarchy theory (Hochstein & Ahissar, 2002), which suggests that high-level conceptual information may emerge later and involve feedback to earlier visual areas. This interpretation is further supported by representational similarities with late-stage artificial neural network layers and with a semantic word embedding model (Word2Vec), both of which reflect learned, abstract knowledge rather than low-level visual features. Taken together, these findings suggest that real-world size is not merely a perceptual attribute, but one that draws on conceptual or semantic-level representations acquired through experience. While our EEG analyses focused on posterior electrodes and thus cannot definitively localize cortical sources, we see this study as a step toward linking low-level visual input with higher-level semantic knowledge. Future work incorporating broader spatial coverage (e.g., anterior sensors), source localization, or complementary modalities such as MEG and fMRI will be critical to adjudicate between alternative models of object representation and to more precisely trace the origin and flow of real-world size information in the brain.”

Finally, several places in the manuscript tout the "novel computational approach". This seems odd because the computational framework and pipeline have been the most common approach in cognitive computational neuroscience in the past 5-10 years.

We have revised relevant statements throughout the manuscript to avoid overstating novelty and to better reflect the contribution of our study.

(2) Suggestion: modernize the approach

I was surprised that the computational models used in this manuscript were all 8-10 years old. Specifically, because there are now deep nets that more explicitly model the human brain (e.g., Cornet) as well as more sophisticated models of semantics (e.g., LLMs), I was left hoping that the authors had used more state-of-the-art models in the work. Moreover, the use of a single dCNN, a single multi-modal model, and a single word embedding model makes it difficult to generalize about visual, multimodal, and semantic features in general.

Thanks for your suggestion. Indeed, our choice of ResNet and CLIP was motivated by their widespread use in the cognitive and computational neuroscience area. These models have served as standard benchmarks in many studies exploring correspondence between ANNs and human brain activity. To address you concern, we have now added additional results from the more biologically inspired model, CORnet, in the supplementary (Figure S10). The results for CORnet show similar patterns to those observed for ResNet and CLIP, providing converging evidence across models.

Regarding semantic modeling, we intentionally chose Word2Vec rather than large language models (LLMs), because our goal was to examine concept-level, context-free semantic representations. Word2Vec remains the most widely adopted approach for obtaining noncontextualized embeddings that reflect core conceptual similarity, as opposed to the contextdependent embeddings produced by LLMs, which are less directly suited for capturing stable concept-level structure across stimuli.

(3) Methodological considerations

(a) Validity of the real-world size measurement

I was concerned about a few aspects of the real-world size rankings. First, I am trying to understand why the scale goes from 100-519. This seems very arbitrary; please clarify. Second, are we to assume that this scale is linear? Is this appropriate when real-world object size is best expressed on a log scale? Third, the authors provide "sand" as an example of the smallest realworld object. This is tricky because sand is more "stuff" than "thing", so I imagine it leaves observers wondering whether the experimenter intends a grain of sand or a sandy scene region. What is the variability in real-world size ratings? Might the variability also provide additional insights in this experiment?

We now clarify the origin, scaling, and interpretation of the real-world size values obtained from the THINGS+ dataset.

In their experiment, participants first rated the size of a single object concept (word shown on the screen) by clicking on a continuous slider of 520 units, which was anchored by nine familiar real-world reference objects (e.g., “grain of sand,” “microwave oven,” “aircraft carrier”) that spanned the full expected size range on a logarithmic scale. Importantly, participants were not shown any numerical values on the scale—they were guided purely by the semantic meaning and relative size of the anchor objects. After the initial response, the scale zoomed in around the selected region (covering 160 units of the 520-point scale) and presented finer anchor points between the previous reference objects. Participants then refined their rating by dragging from the lower to upper end of the typical size range for that object. If the object was standardized in size (e.g., “soccer ball”), a single click sufficed. These size judgments were collected across at least 50 participants per object, and final scores were derived from the central tendency of these responses. Although the final size values numerically range from 0 to 519 (after scaling), this range is not known to participants and is only applied post hoc to construct the size RDMs.

Regarding the term “sand”: the THINGS+ dataset distinguished between object meanings when ambiguity was present. For “sand,” participants were instructed to treat it as “a grain of sand”— consistent with the intended meaning of a discrete, minimal-size reference object. 

Finally, we acknowledge that real-world size ratings may carry some degree of variability across individuals. However, the dataset includes ratings from 2010 participants across 1854 object concepts, with each object receiving at least 50 independent ratings. Given this large and diverse sample, the mean size estimates are expected to be stable and robust across subjects. While we did not include variability metrics in our main analysis, we believe the aggregated ratings provide a reliable estimate of perceived real-world size.

We added these details in the Materials and Method section:

(line 219 to 230) “In the THINGS+ dataset, 2010 participants (different from the subjects in THINGS EEG2) did an online size rating task and completed a total of 13024 trials corresponding to 1854 object concepts using a two-step procedure. In their experiment, first, each object was rated on a 520unit continuous slider anchored by familiar reference objects (e.g., “grain of sand,” “microwave oven,” “aircraft carrier”) representing a logarithmic size range. Participants were not shown numerical values but used semantic anchors as guides. In the second step, the scale zoomed in around the selected region to allow for finer-grained refinement of the size judgment. Final size values were derived from aggregated behavioral data and rescaled to a range of 0–519 for consistency across objects, with the actual mean ratings across subjects ranging from 100.03 (‘grain of sand’) to 423.09 (‘subway’).”

(b) This work has no noise ceiling to establish how strong the model fits are, relative to the intrinsic noise of the data. I strongly suggest that these are included.

We have now computed noise ceiling estimates for the EEG RDMs across time. The noise ceiling was calculated by correlating each participant’s EEG RDM with the average EEG RDM across the remaining participants (leave-one-subject-out), at each time point. This provides an upper-bound estimate of the explainable variance, reflecting the maximum similarity that any model—no matter how complex—could potentially achieve, given the intrinsic variability in the EEG data.

Importantly, the observed EEG–model similarity values are substantially below this upper bound. This outcome is fully expected: Each of our model RDMs (e.g., real-world size, ANN layers) captures only a specific aspect of the neural representational structure, rather than attempting to account for the totality of the EEG signal. Our goal is not to optimize model performance or maximize fit, but to probe which components of object information are reflected in the spatiotemporal dynamics of the brain’s responses.

For clarity and accessibility of the main findings, we present the noise ceiling time courses separately in the supplementary materials (Figure S7). Including them directly in the EEG × HYP or EEG × ANN plots would conflate distinct interpretive goals: the model RDMs are hypothesis-driven probes of specific representational content, whereas the noise ceiling offers a normative upper bound for total explainable variance. Keeping these separate ensures each visualization remains focused and interpretable. 

Reviewer #1 (Recommendations For The Authors)::

Some analyses are incomplete, which would be improved if the authors showed analyses with other layers of the networks and various additional partial correlation analyses.

Clarity

(1) Partial correlations methods incomplete - it is not clear what is being partialled out in each analysis. It is possible to guess sometimes, but it is not entirely clear for each analysis. This is important as it is difficult to assess if the partial correlations are sensible/correct in each case. Also, the Figure 1 caption is short and unclear.

For example, ANN-EEG partial correlations - "Finally, we directly compared the timepoint-bytimepoint EEG neural RDMs and the ANN RDMs (Figure 3F). The early layer representations of both ResNet and CLIP were significantly correlated with early representations in the human brain" What is being partialled out? Figure 3F says partial correlation

We apologize for the confusion. We made several key clarifications and corrections in the revised version.

First, we identified and corrected a labeling error in both Figure 1 and Figure 3F. Specifically, our EEG × ANN analysis used Spearman correlation, not partial correlation as mistakenly indicated in the original figure label and text. We conducted parital correlations for EEG × HYP and ANN × HYP. But for EEG × ANN, we directly calculated the correlation between EEG RDMs and ANN RDM corresponding to different layers respectively. We corrected these errors: (1) In Figure 1, we removed the erroneous “partial” label from the EEG × ANN path and updated the caption to clearly outline which comparisons used partial correlation. (2) In Figure 3F, we corrected the Y-axis label to “(correlation)”.

Second, to improve clarity, we have now revised the Materials and Methods section to explicitly describe what is partialled out in each parital correlation analysis:

(line 284 to 286) “In EEG × HYP partial correlation (Figure 3D), we correlated EEG RDMs with one hypothesis-based RDM (e.g., real-world size), while controlling for the other two (retinal size and real-world depth).”

(line 303 to 305) “In ANN (or W2V) × HYP partial correlation (Figure 3E and Figure 5A), we correlated ANN (or W2V) RDMs with one hypothesis-based RDM (e.g., real-world size), while partialling out the other two.”

Finally, the caption of Figure 1 has been expanded to clarify the full analysis pipeline and explicitly specify the partial correlation or correlation in each comparison.

(line 327 to 332) “Figure 1 Overview of our analysis pipeline including constructing three types of RDMs and conducting comparisons between them. We computed RDMs from three sources: neural data (EEG), hypothesized object features (real-world size, retinal size, and real-world depth), and artificial models (ResNet, CLIP, and Word2Vec). Then we conducted cross-modal representational similarity analyses between: EEG × HYP (partial correlation, controlling for other two HYP features), ANN (or W2V) × HYP (partial correlation, controlling for other two HYP features), and EEG × ANN (correlation).”

We believe these revisions now make all analytic comparisons and correlation types full clear and interpretable.

Issues / open questions

(2) Semantic representations vs hypothesized (hyp) RDMs (real-world size, etc) - are the representations explained by variables in hyp RDMs or are there semantic representations over and above these? E.g., For ANN correlation with the brain, you could partial out hyp RDMs - and assess whether there is still semantic information left over, or is the variance explained by the hyp RDMs?

Thank for this suggestion. As you suggested, we conducted the partial correlation analysis between EEG RDMs and ANN RDMs, controlling for the three hypothesis-based RDMs. The results (Figure S6) revealed that the EEG×ANN representational similarity remained largely unchanged, indicating that ANN representations capture much more additional representational structure not accounted for by the current hypothesized features. This is also consistent with the observation that EEG×HYP partial correlations were themselves small, but EEG×ANN correlations were much greater.

We also added this statement to the main text:

(line 446 to 451) “To contextualize how much of the shared variance between EEG and ANN representations is driven by the specific visual object features we tested above, we conducted a partial correlation analysis between EEG RDMs and ANN RDMs controlling for the three hypothesis-based RDMs (Figure S6). The EEG×ANN similarity results remained largely unchanged, suggesting that ANN representations capture much more additional rich representational structure beyond these features. ”

(3) Why only early and late layers? I can see how it's clearer to present the EEG results. However, the many layers in these networks are an opportunity - we can see how simple/complex linear/non-linear the transformation is over layers in these models. It would be very interesting and informative to see if the correlations do in fact linearly increase from early to later layers, or if the story is a bit more complex. If not in the main text, then at least in the supplement.

Thank you for the thoughtful suggestion. To address this point, we have computed the EEG correlations with multiple layers in both ResNet and CLIP models (ResNet: ResNet.maxpool, ResNet.layer1, ResNet.layer2, ResNet.layer3, ResNet.layer4, ResNet.avgpool; CLIP:CLIP.visual.avgpool, CLIP.visual.layer1, CLIP.visual.layer2, CLIP.visual.layer3, CLIP.visual.layer4, CLIP.visual.attnpool). The results, now included in Figure S4 and S5, show a consistent trend: early layers exhibit higher similarity to early EEG time points, and deeper layers show increased similarity to later EEG stages. We chose to highlight early and late layers in the main text to simplify interpretation, but now provide the full layerwise profile for completeness.

(4) Peak latency analysis - Estimating peaks per ppt is presumably noisy, so it seems important to show how reliable this is. One option is to find the bootstrapped mean latencies per subject.

Thanks for your suggestion. To estimate the robustness of peak latency values, we implemented a bootstrap procedure by resampling the pairwise entries of the EEG RDM with replacement. For each bootstrap sample, we computed a new EEG RDM and recalculated the partial correlation time course with the hypothesis RDMs. We then extracted the peak latency within the predefined significant time window. Repeating this process 1000 times allowed us to get the bootstrapped mean latencies per subject as the more stable peak latency result. Notably, the bootstrapped results showed minimal deviation from the original latency estimates, confirming the robustness of our findings. Accordingly, we updated the Figure 3D and added these in the Materials and Methods section:

(line 289 to 298) “To assess the stability of peak latency estimates for each subject, we performed a bootstrap procedure across stimulus pairs. At each time point, the EEG RDM was vectorized by extracting the lower triangle (excluding the diagonal), resulting in 19,900 unique pairwise values. For each bootstrap sample, we resampled these 19,900 pairwise entries with replacement to generate a new pseudo-RDM of the same size. We then computed the partial correlation between the EEG pseudo-RDM and a given hypothesis RDM (e.g., real-world size), controlling for other feature RDMs, and obtained a time course of partial correlations. Repeating this procedure 1000 times and extracting the peak latency within the significant time window yielded a distribution of bootstrapped latencies, from which we got the bootstrapped mean latencies per subject.”

(5) "Due to our calculations being at the object level, if there were more than one of the same objects in an image, we cropped the most complete one to get a more accurate retinal size. " Did EEG experimenters make sure everyone sat the same distance from the screen? and remain the same distance? This would also affect real-world depth measures.

Yes, the EEG dataset we used (THINGS EEG2; Gifford et al., 2022) was collected under carefully controlled experimental conditions. We have confirmed that all participants were seated at a fixed distance of 0.6 meters from the screen throughout the experiment. We also added this information in the method (line 156 to 157).

Minor issues/questions - note that these are not raised in the Public Review

(6) Title - less about rigor/quality of the work but I feel like the title could be improved/extended. The work tells us not only about real object size, but also retinal size and depth. In fact, isn't the most novel part of this the real-world depth aspect? Furthermore, it feels like the current title restricts its relevance and impact... Also doesn't touch on the temporal aspect, or processing stages, which is also very interesting. There may be something better, but simply adding something like"...disentangled features of real-world size, depth, and retinal size over time OR processing stages".

Thanks for your suggestion! We changed our title – “Human EEG and artificial neural networks reveal disentangled representations and processing timelines of object real-world size and depth in natural images”.

(7) "Each subject viewed 16740 images of objects on a natural background for 1854 object concepts from the THINGS dataset (Hebart et al., 2019). For the current study, we used the 'test' dataset portion, which includes 16000 trials per subject corresponding to 200 images." Why test images? Worth explaining.

We chose to use the “test set” of the THINGS EEG2 dataset for the following two reasons:

(1) Higher trial count per condition: In the test set, each of the 200 object images was presented 80 times per subject, whereas in the training set, each image was shown only 4 times. This much higher trial count per condition in the test set allows for substantially higher signal-tonoise ratio in the EEG data.

(2) Improved decoding reliability: Our analysis relies on constructing EEG RDMs based on pairwise decoding accuracy using linear SVM classifiers. Reliable decoding estimates require a sufficient number of trials per condition. The test set design is thus better suited to support high-fidelity decoding and robust representational similarity analysis.

We also added these explainations to our revised manuscript (line 161 to 164).

(8) "For Real-World Size RDM, we obtained human behavioral real-world size ratings of each object concept from the THINGS+ dataset (Stoinski et al., 2022).... The range of possible size ratings was from 0 to 519 in their online size rating task..." How were the ratings made? What is this scale - do people know the numbers? Was it on a continuous slider?

We should clarify how the real-world size values were obtained from the THINGS+ dataset.

In their experiment, participants first rated the size of a single object concept (word shown on the screen) by clicking on a continuous slider of 520 units, which was anchored by nine familiar real-world reference objects (e.g., “grain of sand,” “microwave oven,” “aircraft carrier”) that spanned the full expected size range on a logarithmic scale. Importantly, participants were not shown any numerical values on the scale—they were guided purely by the semantic meaning and relative size of the anchor objects. After the initial response, the scale zoomed in around the selected region (covering 160 units of the 520-point scale) and presented finer anchor points between the previous reference objects. Participants then refined their rating by dragging from the lower to upper end of the typical size range for that object. If the object was standardized in size (e.g., “soccer ball”), a single click sufficed. These size judgments were collected across at least 50 participants per object, and final scores were derived from the central tendency of these responses. Although the final size values numerically range from 0 to 519 (after scaling), this range is not known to participants and is only applied post hoc to construct the size RDMs.

We added these details in the Materials and Method section:

(line 219 to 230) “In the THINGS+ dataset, 2010 participants (different from the subjects in THINGS EEG2) did an online size rating task and completed a total of 13024 trials corresponding to 1854 object concepts using a two-step procedure. In their experiment, first, each object was rated on a 520unit continuous slider anchored by familiar reference objects (e.g., “grain of sand,” “microwave oven,” “aircraft carrier”) representing a logarithmic size range. Participants were not shown numerical values but used semantic anchors as guides. In the second step, the scale zoomed in around the selected region to allow for finer-grained refinement of the size judgment. Final size values were derived from aggregated behavioral data and rescaled to a range of 0–519 for consistency across objects, with the actual mean ratings across subjects ranging from 100.03 (‘grain of sand’) to 423.09 (‘subway’).”

(9) "For Retinal Size RDM, we applied Adobe Photoshop (Adobe Inc., 2019) to crop objects corresponding to object labels from images manually... " Was this by one person? Worth noting, and worth sharing these values per image if not already for other researchers as it could be a valuable resource (and increase citations).

Yes, all object cropping were performed consistently by one of the authors to ensure uniformity across images. We agree that this dataset could be a useful resource to the community. We have now made the cropped object images publicly available https://github.com/ZitongLu1996/RWsize.

We also updated the manuscript accordingly to note this (line 236 to 239).

(10) "Neural RDMs. From the EEG signal, we constructed timepoint-by-timepoint neural RDMs for each subject with decoding accuracy as the dissimilarity index " Decoding accuracy is presumably a similarity index. Maybe 1-accuracy (proportion correct) for dissimilarity?

Decoding accuracy is a dissimilarity index instead of a similarity index, as higher decoding accuracy between two conditions indicates that they are more distinguishable – i.e., less similar – in the neural response space. This approach aligns with prior work using classification-based representational dissimilarity measures (Grootswagers et al., 2017; Xie et al., 2020), where better decoding implies greater dissimilarity between conditions. Therefore, there is no need to invert the decoding accuracy values (e.g., using 1 - accuracy).

Grootswagers, T., Wardle, S. G., & Carlson, T. A. (2017). Decoding dynamic brain patterns from evoked responses: A tutorial on multivariate pattern analysis applied to time series neuroimaging data. Journal of Cognitive Neuroscience, 29(4), 677-697.

Xie, S., Kaiser, D., & Cichy, R. M. (2020). Visual imagery and perception share neural representations in the alpha frequency band. Current Biology, 30(13), 2621-2627.

(11) Figure 1 caption is very short - Could do with a more complete caption. Unclear what the partial correlations are (what is being partialled out in each case), what are the comparisons "between them" - both in the figure and the caption. Details should at least be in the main text.

Related to your comment (1). We revised the caption and the corresponding text.

Reviewer #2 (Recommendations For The Authors):

(1) Intro:

Quek et al., (2023) is referred to as a behavioral study, but it has EEG analyses.

We corrected this – “…, one recent study (Quek et al., 2023) …”

The phrase 'high temporal resolution EEG' is a bit strange - isn't all EEG high temporal resolution? Especially when down-sampling to 100 Hz (40 time points/epoch) this does not qualify as particularly high-res.

We removed this phrasing in our manuscript.

(2) Methods:

It would be good to provide more details on the EEG preprocessing. Were the data low-pass filtered, for example?

We added more details to the manuscript:

(line 167 to 174) “The EEG data were originally sampled at 1000Hz and online-filtered between 0.1 Hz and 100 Hz during acquisition, with recordings referenced to the Fz electrode. For preprocessing, no additional filtering was applied. Baseline correction was performed by subtracting the mean signal during the 100 ms pre-stimulus interval from each trial and channel separately. We used already preprocessed data from 17 channels with labels beginning with “O” or “P” (O1, Oz, O2, PO7, PO3, POz, PO4, PO8, P7, P5, P3, P1, Pz, P2) ensuring full coverage of posterior regions typically involved in visual object processing. The epoched data were then down-sampled to 100 Hz.”

It is important to provide more motivation about the specific ANN layers chosen. Were these layers cherry-picked, or did they truly represent a gradual shift over the course of layers?

We appreciate the reviewer’s concern and fully agree that it is important to ensure transparency in how ANN layers were selected. The early and late layers reported in the main text were not cherry-picked to maximize effects, but rather intended to serve as illustrative examples representing the lower and higher ends of the network hierarchy. To address this point directly, we have computed the EEG correlations with multiple layers in both ResNet and CLIP models (ResNet: ResNet.maxpool, ResNet.layer1, ResNet.layer2, ResNet.layer3, ResNet.layer4, ResNet.avgpool; CLIP: CLIP.visual.avgpool, CLIP.visual.layer1, CLIP.visual.layer2, CLIP.visual.layer3, CLIP.visual.layer4, CLIP.visual.attnpool). The results, now included in Figure S4, show a consistent trend: early layers exhibit higher similarity to early EEG time points, and deeper layers show increased similarity to later EEG stages.

It is important to provide more specific information about the specific ANN layers chosen. 'Second convolutional layer': is this block 2, the ReLu layer, the maxpool layer? What is the 'last visual layer'?

Apologize for the confusing! We added more details about the layer chosen:

(line 255 to 257) “The early layer in ResNet refers to ResNet.maxpool layer, and the late layer in ResNet refers to ResNet.avgpool layer. The early layer in CLIP refers to CLIP.visual.avgpool layer, and the late layer in CLIP refers to CLIP.visual.attnpool layer.”

Again the claim 'novel' is a bit overblown here since the real-world size ratings were also already collected as part of THINGS+, so all data used here is available.

We removed this phrasing in our manuscript.

Real-world size ratings ranged 'from 0 - 519'; it seems unlikely this was the actual scale presented to subjects, I assume it was some sort of slider?

You are correct. We should clarify how the real-world size values were obtained from the THINGS+ dataset.

In their experiment, participants first rated the size of a single object concept (word shown on the screen) by clicking on a continuous slider of 520 units, which was anchored by nine familiar real-world reference objects (e.g., “grain of sand,” “microwave oven,” “aircraft carrier”) that spanned the full expected size range on a logarithmic scale. Importantly, participants were not shown any numerical values on the scale—they were guided purely by the semantic meaning and relative size of the anchor objects. After the initial response, the scale zoomed in around the selected region (covering 160 units of the 520-point scale) and presented finer anchor points between the previous reference objects. Participants then refined their rating by dragging from the lower to upper end of the typical size range for that object. If the object was standardized in size (e.g., “soccer ball”), a single click sufficed. These size judgments were collected across at least 50 participants per object, and final scores were derived from the central tendency of these responses. Although the final size values numerically range from 0 to 519 (after scaling), this range is not known to participants and is only applied post hoc to construct the size RDMs.

We added these details in the Materials and Method section:

(line 219 to 230) “In the THINGS+ dataset, 2010 participants (different from the subjects in THINGS EEG2) did an online size rating task and completed a total of 13024 trials corresponding to 1854 object concepts using a two-step procedure. In their experiment, first, each object was rated on a 520unit continuous slider anchored by familiar reference objects (e.g., “grain of sand,” “microwave oven,” “aircraft carrier”) representing a logarithmic size range. Participants were not shown numerical values but used semantic anchors as guides. In the second step, the scale zoomed in around the selected region to allow for finer-grained refinement of the size judgment. Final size values were derived from aggregated behavioral data and rescaled to a range of 0–519 for consistency across objects, with the actual mean ratings across subjects ranging from 100.03 (‘grain of sand’) to 423.09 (‘subway’).”

Why is conducting a one-tailed (p<0.05) test valid for EEG-ANN comparisons? Shouldn't this be two-tailed?

Our use of one-tailed tests was based on the directional hypothesis that representational similarity between EEG and ANN RDMs would be positive, as supported by prior literature showing correspondence between hierarchical neural networks and human brain representations (e.g., Cichy et al., 2016; Kuzovkin et al., 2014). This is consistent with a large number of RSA studies which conduct one-tailed tests (i.e., testing the hypothesis that coefficients were greater than zero: e.g., Kuzovkin et al., 2018; Nili et al., 2014; Hebart et al., 2018; Kaiser et al., 2019; Kaiser et al., 2020; Kaiser et al., 2022). Thus, we specifically tested whether the similarity was significantly greater than zero.

Cichy, R. M., Khosla, A., Pantazis, D., Torralba, A., & Oliva, A. (2016). Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence. Scientific reports, 6(1), 27755.

Kuzovkin, I., Vicente, R., Petton, M., Lachaux, J. P., Baciu, M., Kahane, P., ... & Aru, J. (2018). Activations of deep convolutional neural networks are aligned with gamma band activity of human visual cortex. Communications biology, 1(1), 107.

Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., & Kriegeskorte, N. (2014). A toolbox for representational similarity analysis. PLoS computational biology, 10(4), e1003553.

Hebart, M. N., Bankson, B. B., Harel, A., Baker, C. I., & Cichy, R. M. (2018). The representational dynamics of task and object processing in humans. Elife, 7, e32816.

Kaiser, D., Turini, J., & Cichy, R. M. (2019). A neural mechanism for contextualizing fragmented inputs during naturalistic vision. elife, 8, e48182.

Kaiser, D., Inciuraite, G., & Cichy, R. M. (2020). Rapid contextualization of fragmented scene information in the human visual system. Neuroimage, 219, 117045.

Kaiser, D., Jacobs, A. M., & Cichy, R. M. (2022). Modelling brain representations of abstract concepts. PLoS Computational Biology, 18(2), e1009837.

Importantly, we note that using a two-tailed test instead would not change the significance of our results. However, we believe the one-tailed test remains more appropriate given our theoretical prediction of positive similarity between ANN and brain representations.

The sentence on the partial correlation description (page 11 'we calculated partial correlations with one-tailed test against the alternative hypothesis that the partial correlation was positive (greater than zero)') didn't make sense to me; are you referring to the null hypothesis here?

We revised this sentence to clarify that we tested against the null hypothesis that the partial correlation was less than or equal to zero, using a one-tailed test to assess whether the correlation was significantly greater than zero.

(line 281 to 284) “…, we calculated partial correlations and used a one-tailed test against the null hypothesis that the partial correlation was less than or equal to zero, testing whether the partial correlation was significantly greater than zero.”

(3) Results:

I would prevent the use of the word 'pure', your measurement is one specific operationalization of this concept of real-world size that is not guaranteed to result in unconfounded representations. This is in fact impossible whenever one is using a finite set of natural stimuli and calculating metrics on those - there can always be a factor or metric that was not considered that could explain some of the variance in your measurement. It is overconfident to claim to have achieved some form of Platonic ideal here and to have taken into account all confounds.

Your point is well taken. Our original use of the term “pure” was intended to reflect statistical control for known confounding factors, but we recognize that this wording may imply a stronger claim than warranted. In response, we revised all relevant language in the manuscript to instead describe the statistically isolated or relatively unconfounded representation of real-world size, clarifying that our findings pertain to the unique contribution of real-world size after accounting for retinal size and real-world depth.

Figure 2C: It's not clear why peak latencies are computed on the 'full' correlations rather than the partial ones.

No. The peak latency results in Figure 2C were computed on the partial correlation results – we mentioned this in the figure caption – “Temporal latencies for peak similarity (partial Spearman correlations) between EEG and the 3 types of object information.”

SEM = SEM across the 10 subjects?

Yes. We added this in the figure caption.

Figure 3F y-axis says it's partial correlations but not clear what is partialled out here.

We identified and corrected a labeling error in both Figure 1 and Figure 3F. Specifically, our EEG × ANN analysis used Spearman correlation, not partial correlation as mistakenly indicated in the original figure label and text. We conducted parital correlations for EEG × HYP and ANN × HYP. But for EEG × ANN, we directly calculated the correlation between EEG RDMs and ANN RDM corresponding to different layers respectively. We corrected these errors: (1) In Figure 1, we removed the erroneous “partial” label from the EEG × ANN path and updated the caption to clearly outline which comparisons used partial correlation. (2) In Figure 3F, we corrected the Y-axis label to “(correlation)”.

Reviewer #3 (Recommendations For The Authors):

(1) Several methodologies should be clarified:

(a) It's stated that EEG was sampled at 100 Hz. I assume this was downsampled? From what original frequency?

Yes. We added more detailed about EEG data:

(line 167 to 174) “The EEG data were originally sampled at 1000Hz and online-filtered between 0.1 Hz and 100 Hz during acquisition, with recordings referenced to the Fz electrode. For preprocessing, no additional filtering was applied. Baseline correction was performed by subtracting the mean signal during the 100 ms pre-stimulus interval from each trial and channel separately. We used already preprocessed data from 17 channels with labels beginning with “O” or “P” (O1, Oz, O2, PO7, PO3, POz, PO4, PO8, P7, P5, P3, P1, Pz, P2) ensuring full coverage of posterior regions typically involved in visual object processing. The epoched data were then down-sampled to 100 Hz.”

(b) Why was decoding accuracy used as the human RDM method rather than the EEG data themselves?

Thanks for your question! We would like to address why we used decoding accuracy for EEG RDMs rather than correlation. While fMRI RDMs are typically calculated using 1 minus correlation coefficient, decoding accuracy is more commonly used for EEG RDMs (Grootswager et al., 2017; Xie et al., 2020). The primary reason is that EEG signals are more susceptible to noise than fMRI data. Correlation-based methods are particularly sensitive to noise and may not reliably capture the functional differences between EEG patterns for different conditions. Decoding accuracy, by training classifiers to focus on task-relevant features, can effectively mitigate the impact of noisy signals and capture the representational difference between two conditions.

Grootswagers, T., Wardle, S. G., & Carlson, T. A. (2017). Decoding dynamic brain patterns from evoked responses: A tutorial on multivariate pattern analysis applied to time series neuroimaging data. Journal of Cognitive Neuroscience, 29(4), 677-697.

Xie, S., Kaiser, D., & Cichy, R. M. (2020). Visual imagery and perception share neural representations in the alpha frequency band. Current Biology, 30(13), 2621-2627.

We added this explanation to the manuscript:

(line 204 to 209) “Since EEG has a low SNR and includes rapid transient artifacts, Pearson correlations computed over very short time windows yield unstable dissimilarity estimates (Kappenman & Luck, 2010; Luck, 2014) and may thus fail to reliably detect differences between images. In contrast, decoding accuracy - by training classifiers to focus on task-relevant features - better mitigates noise and highlights representational differences.”

(c) How were the specific posterior electrodes selected?

The 17 posterior electrodes used in our analyses were pre-selected and provided in the THINGS EEG2 dataset, and corresponding to standard occipital and parietal sites based on the 10-10 EEG system. Specifically, we included all 17 electrodes with labels beginning with “O” or “P”, ensuring full coverage of posterior regions typically involved in visual object processing (Page 7).

(d) The specific layers should be named rather than the vague ("last visual")

Apologize for the confusing! We added more details about the layer information:

(line 255 to 257) “The early layer in ResNet refers to ResNet.maxpool layer, and the late layer in ResNet refers to ResNet.avgpool layer. The early layer in CLIP refers to CLIP.visual.avgpool layer, and the late layer in CLIP refers to CLIP.visual.attnpool layer.”

(line 420 to 434) “As shown in Figure 3F, the early layer representations of both ResNet and CLIP (ResNet.maxpool layer and CLIP.visual.avgpool) showed significant correlations with early EEG time windows (early layer of ResNet: 40-280ms, early layer of CLIP: 50-130ms and 160-260ms), while the late layers (ResNet.avgpool layer and CLIP.visual.attnpool layer) showed correlations extending into later time windows (late layer of ResNet: 80-300ms, late layer of CLIP: 70-300ms). Although there is substantial temporal overlap between early and late model layers, the overall pattern suggests a rough correspondence between model hierarchy and neural processing stages.

We further extended this analysis across intermediate layers of both ResNet and CLIP models (from early to late, ResNet: ResNet.maxpool, ResNet.layer1, ResNet.layer2, ResNet.layer3, ResNet.layer4, ResNet.avgpool; from early to late, CLIP: CLIP.visual.avgpool, CLIP.visual.layer1, CLIP.visual.layer2, CLIP.visual.layer3, CLIP.visual.layer4, CLIP.visual.attnpool).”

(e) p19: please change the reporting of t-statistics to standard APA format.

Thanks for the suggestion. We changed the reporting format accordingly:

(line 392 to 394) “The representation of real-word size had a significantly later peak latency than that of both retinal size, t(9)=4.30, p=.002, and real-world depth, t(9)=18.58, p<.001. And retinal size representation had a significantly later peak latency than real-world depth, t(9)=3.72, p=.005.”

(2) "early layer of CLIP: 50-130ms and 160-260ms), while the late layer representations of twoANNs were significantly correlated with later representations in the human brain (late layer of ResNet: 80-300ms, late layer of CLIP: 70-300ms)."

This seems a little strong, given the large amount of overlap between these models.

We agree that our original wording may have overstated the distinction between early and late layers, given the substantial temporal overlap in their EEG correlations. We revised this sentence to soften the language to reflect the graded nature of the correspondence, and now describe the pattern as a general trend rather than a strict dissociation:

(line 420 to 427) “As shown in Figure 3F, the early layer representations of both ResNet and CLIP (ResNet.maxpool layer and CLIP.visual.avgpool) showed significant correlations with early EEG time windows (early layer of ResNet: 40-280ms, early layer of CLIP: 50-130ms and 160-260ms), while the late layers (ResNet.avgpool layer and CLIP.visual.attnpool layer) showed correlations extending into later time windows (late layer of ResNet: 80-300ms, late layer of CLIP: 70-300ms). Although there is substantial temporal overlap between early and late model layers, the overall pattern suggests a rough correspondence between model hierarchy and neural processing stages.”

(3) "Also, human brain representations showed a higher similarity to the early layer representation of the visual model (ResNet) than to the visual-semantic model (CLIP) at an early stage. "

This has been previously reported by Greene & Hansen, 2020 J Neuro.

Thanks! We added this reference.

(4) "ANN (and Word2Vec) model RDMs"

Why not just "model RDMs"? Might provide more clarity.

We chose to use the phrasing “ANN (and Word2Vec) model RDMs” to maintain clarity and avoid ambiguity. In the literature, the term “model RDMs” is sometimes used more broadly to include hypothesis-based feature spaces or conceptual models, and we wanted to clearly distinguish our use of RDMs derived from artificial neural networks and language models. Additionally, explicitly referring to ANN or Word2Vec RDMs improves clarity by specifying the model source of each RDM. We hope this clarification justifies our choice to retain the original phrasing for clarity.

FIT/FIP認定

/hyperpost/🌐/🎭/gyuri/=/

A subcommand is missing from the manual: nssm edit <servicename>, which allow us to show service editing GUI.

Ev88 san choi ca cuoc chuyen nghiep ngay cang khang dinh vi the cua minh tren thi truong. Nen tang mang den cho anh em trai nghiem tot nhat.

Dia chi: 214 D. Le Duc Tho, Phuong 6, Go Vap, Ho Chi Minh, Viet Nam

Email: thriftdarianh50648@gmail.com

Website: https://ev88.miami/

Dien thoai: (+84) 886602928

ev88 #ev88com #casinoev88 #thethaoev88 #bancaev88 #nhacaiev88 #trangchuev88 #bacaratev88

Social Links:

https://ev88.miami/

https://www.youtube.com/channel/UCmc7c2XgimpDY6V6QFTzIEw

https://twitter.com/ev88miami

https://www.reddit.com/user/ev88miami/

https://www.pinterest.com/ev88miami/

https://rant.li/ev88miami/ev88miami

https://gravatar.com/ev88miami

https://band.us/@ev88miami

https://www.blogger.com/profile/11156575866609057322

https://thriftdarianh50648.wixsite.com/ev88miami

https://www.tumblr.com/ev88miami

https://ev88miami.wordpress.com/

https://www.twitch.tv/ev88miami/about

https://sites.google.com/view/ev88miami/home

https://ev88miami.webflow.io/

https://bookmarksclub.com/backlink/ev88miami/

https://ev88miami.mystrikingly.com/

https://ev88miami.amebaownd.com/

https://telegra.ph/ev88miami-07-04

https://mk.gta5-mods.com/users/ev88miami

https://686750e952e4f.site123.me/

https://myspace.com/ev88miami

https://scholar.google.com/citations?hl=vi&user=2EVhH7AAAAAJ

https://www.pearltrees.com/ev88miami/item724998896

https://ev88miami.localinfo.jp/

https://ev88miami.shopinfo.jp/

https://ev88miami.hashnode.space/default-guide/ev88miami

https://ev88miami.themedia.jp/

https://rapidapi.com/user/ev88miami

https://729433.8b.io/

https://ev88miami.theblog.me/

https://fliphtml5.com/homepage/tbhcq/ev88miami/

https://ev88miami.therestaurant.jp/

https://ask.mallaky.com/?qa=user/ev88miami

https://ev88miami.website3.me/

https://www.quora.com/profile/Ev88miami

https://ev88miami.pixieset.com/

https://ev88miami.gumroad.com/

https://flipboard.com/@ev88miami/ev88miami-5298g17jy

https://www.threadless.com/@ev88miami/activity

https://wakelet.com/@ev88miami

https://www.magcloud.com/user/ev88miami

https://hackmd.io/@ev88miami/ev88miami

https://ev88miami.blogspot.com/2025/07/ev88miami.html

https://ev88miami.doorkeeper.jp/

https://ev88miami.storeinfo.jp/

https://velog.io/@ev88miami/about

https://bato.to/u/2796582-ev88miami

https://zb3.org/ev88miami/ev88miami

https://github.com/ev88miami

https://community.fabric.microsoft.com/t5/user/viewprofilepage/user-id/1302474

https://bit.ly/ev88miami

https://tinyurl.com/ev88miami

https://tawk.to/ev88miami

https://gitlab.com/ev88miami

https://rebrand.ly/ev88miami

https://www.question-ksa.com/user/ev88miami

https://bulkwp.com/support-forums/users/ev88miami/

https://orcid.org/0009-0001-5206-0502

https://community.cisco.com/t5/user/viewprofilepage/user-id/1894315

https://linktr.ee/ev88miami

https://archive.org/details/@ev88miami/web-archive

https://wpfr.net/support/utilisateurs/ev88miami

https://youbiz.com/profile/ev88miami/

https://plaza.rakuten.co.jp/ev88miami/diary/202507040000/

https://pad.darmstadt.social/s/8SuC_5Ur-

https://pixabay.com/users/51169801/

https://disqus.com/by/ev88miami/about/

https://www.reverbnation.com/artist/ev88miami

https://es.gta5-mods.com/users/ev88miami

https://www.gamblingtherapy.org/forum/users/ev88miami/

https://heylink.me/ev88miami/

https://forum.m5stack.com/user/ev88miami/

https://app.readthedocs.org/profiles/ev88miami/

https://gitee.com/ev88miami

https://public.tableau.com/app/profile/ev88.miami/viz/ev88miami/Sheet1#1

https://connect.garmin.com/modern/profile/8c107765-b180-48ad-8f3c-ff50448b4ce5

https://www.pixiv.net/en/users/117693818

https://community.amd.com/t5/user/viewprofilepage/user-id/511056

https://readtoto.com/u/2796582-ev88miami

https://s.id/wlmHq

https://qna.habr.com/user/ev88miami

https://linkr.bio/ev88miami

https://www.bark.com/en/gb/company/ev88miami/3Oj1Rg/

https://pastebin.com/u/ev88miami

https://www.storeboard.com/ev88miami

https://etextpad.com/phjlllk9dj

https://md.darmstadt.ccc.de/s/zosh_lgJ_

https://vc.ru/id5073070

https://qiita.com/ev88miami

https://comicvine.gamespot.com/profile/ev88miami/

https://padlet.com/ev88miami/ev88miami

https://3dwarehouse.sketchup.com/by/ev88miami

https://muckrack.com/ev88-miami/bio

https://hedgedoc.k8s.eonerc.rwth-aachen.de/s/yQb4mFtzl

https://connect.informs.org/network/speakerdirectory/speaker?UserKey=7f46d71a-bc0a-4a17-9ae4-0197d412e9ea

https://nl.gta5-mods.com/users/ev88miami

https://openlibrary.org/people/ev88miami

https://anyflip.com/homepage/womca#About

https://lu.ma/user/ev88miami

https://pbase.com/ev88miami/ev88miami

Learn practical strategies to design AI systems that avoid bias and ensure fairness. Discover techniques like diverse data, transparent algorithms, and robust evaluation pipelines to build ethical AI.

Compare AWS Transcribe, Deepgram, and Whisper for speech-to-text accuracy, pricing, integrations, and use cases. Find the best AI transcription service for your business.

Long Tail Keyword Generator, Learn how to generate useful keywords to start exploring useful different topic.

why do we need different formulas for different waves

I agree with this. Teachers today need to do more than just teach from books. Students are already learning via memes, influencers, and YouTube. Schools should teach kids how to think critically about what they read and see online. Media literacy lessons can help teens tell the difference between facts and lies and be better digital citizens.

These ideas are strong because they show that media is never neutral. Everything, from news broadcasts to TikTok videos, has a reason for being, a target audience, and a bias. I enjoy how this relates to how marketing and algorithms today affect what we see online. It makes me think about how media literacy is about figuring out what those hidden motives are and how to spot persuasion and manipulation in ordinary media.

This is a good reminder that being "media literate" isn't simply understanding how to use media; it's also about questioning and examining it. A lot of individuals look at social media or watch the news without thinking about bias or purpose. Being media literate entails asking why something was made, who profits from it, and what it says. It's not just about consuming; it's an active and thoughtful process.

for - Social change media - Participant Media - question - Did the success of Participant Media lead to its own obsolescence?

Marwick fights against the idea that technology is the only thing that changes civilization. She says that politics, society, and economics all affect how tools are used. This relates to everything we've read about the history of media. Just like the printing press didn't start the Reformation, social media didn't start democracy. Her criticism reminds us that we should be careful of Silicon Valley's assertion that apps and platforms automatically make the world "better."

This reflects Silicon Valley's long-held view that technology always leads to development. Marwick says that this optimism covers problems, including mixing libertarian capitalism with virtues from the counterculture, like creativity and disobedience. It's amazing how this way of thinking still affects modern tech culture, where businesses say they want to "change the world" but instead make things worse by supporting systems that make money and inequity (like AI and gig platforms today).

Marwick explains that Web 2.0 started out with the hope that technology could give regular people more influence and take the place of "Big Media." This idealism said that everyone would be equal if they took part, but in reality, platforms like Facebook and YouTube were taken over by ads and big companies. Her point makes me think about how current social media still offers the idea of community while making money off of users' attention and data.

Connects research to practical classroom applications, advocating for integrated language learning strategies.

Positions the study as contributing to broader understanding of how language changes in response to technology and globalization.

Suggests that code-switching and translanguaging are not just descriptive phenomena.

Reaffirms that multilingual practices online are intentional, flexible, and resourceful strategies.

Highlights implications for education, suggesting translanguaging could inform more integrated, flexible teaching methods.

Suggests that these practices are contributing to the creation of hybrid, evolving forms of language.

Emphasizes translanguaging’s role in creating contextually rich, audience-focused content.

Platform-specific evidence that Twitter encourages code-switching to manage space and maintain clarity.

Shows the practical function of code-switching for concise, effective communication in character-limited environments.

Indicates that code-switching is highly influenced by social context and platform-specific communication rules.

Introduces the discussion and emphasizes that the study provides insights into digital multilingual practices.

Compares platforms, showing how design and technical constraints influence the type of multilingual practices users employ.

Reinforces the idea that multilingual strategies improve clarity, precision, and expressiveness in online communication.

Shows translanguaging’s role in enhancing the narrative and expressive potential of digital content, particularly when visuals or audio are involved.

Suggests that users strategically switch languages to connect with multilingual audiences and fit platform-specific communication conventions.

Shows that code-switching is the most common multilingual strategy online and highlights how social context and platform rules shape language use.

Thematic coding allows exploration of motivations, challenges, and identity construction in online multilingual practices.

Explains the criteria used to interpret online multilingual behavior, emphasizing both linguistic and social dimensions.

Surveys capture both quantitative and qualitative data, providing insights into behaviors, contexts, and motivations.

Highlights the study’s focus on diversity in language use, allowing analysis of patterns across multiple languages.

This establishes the sample size and multilingual criteria, showing the study focuses on active multilingual individuals.

Digital communication shapes linguistic practices, blending languages, and creating both challenges and opportunities for minority languages and global interaction.

How can platforms be designed to reduce linguistic inequities?

Could translanguaging reshape how language learning is approached in digital education?

Translanguaging is the integrated use of multiple linguistic resources beyond simple alternation.

Switching languages online can signal shifts in topic or emotion; hashtags serve as cross-lingual bridges

How might character limits on platforms like Twitter influence the frequency and style of code-switching compared to face-to-face communication?

Important definition of code-switching

The study frames digital multilingual communication as a key site for exploring language use, identity, and cultural expression.

There’s a lack of research on how these multilingual practices operate in digital contexts, which this study aims to address.

Translanguaging differs from code-switching by blending linguistic resources rather than alternating distinct languages, showing fluidity in multilingual expression.

Could code-switching online lead to new language norms or hybrid forms unique to digital spaces?

Code-switching is an established multilingual practice, now amplified by online platforms that make switching easier and more frequent.

The author sets the stage by linking digital technologies to multilingual communication, emphasizing code-switching and translanguaging as central phenomena.

This comment highlights how journalism lost control over dissemination and revenues as it became dependent on social media infrastructure. Ytreberg says that the same platforms that say they "connect" people also hurt professional journalism and public trust. It relates to thoughts on the fall of gatekeeping and the growth of "gatewatching" (Bruns). The end effect is a broken public sphere where emotive and viral content often takes the place of accurate reporting.

Ytreberg shows here how algorithms have replaced editors and curators, but they don't have to be open or responsible. People assume they are choose what to see and share, but algorithms discreetly pick what gets seen. This connects to what we've been talking about with disinformation and media literacy. Just like in Cat Park, people often don't know how digital networks change what they believe or value. The fact that they are not visible gives internet corporations a lot of power in politics and culture.

This statement sums up how social media changed the way people talk to each other throughout the world. Ytreberg says that the early days of the internet focused on individual freedom, but the 2000s saw a move toward social connectedness, which was paradoxically caused by companies like Facebook and Twitter. This "return" wasn't only social; it was also commercial, since sharing and talking were ways to make money. It makes me think about how a lot of what feels like "community" online is really built into systems that are meant to collect data and change behavior.

for - SRG Corporation2CO-OPeration program - worker-owned cooperatives - Apis & Heritage - inequality reduction - via worker-owned cooperatives

summary - Apis & Heritage is a unique US private equity firm that has established an investment fund called "The Legacy Fund" which is used to facilitate Employee-Led BuyOut (ELBO). Studies show the enormous potential for reducing inequality and it is an issue that receives rare bipartisan political support in the US. The "Silver Tsunami" describes 3 million small business owners likely to retire in 2035. Together, their businesses account for $10 trillion in assets. Apis & Heritage helps faciliate a smooth transition for owners to sell to their employees, increasing their net worth by as much as 10x by the time they retire.

.> for - quote - worker-owned cooperatives - Michael Brownrigg

for - Apis & Heritage - champions of worker-owned cooperative movement

for - worker-owned cooperatives - rare bipartisan support

for - stats - 2022 - US worker-owned cooperative potential - about 140,000 firms - employing around 33 million workers - would have been suitable candidates for ESOP employee buyouts, - nearly 1.1 million firms - employing over 25 million workers - [are] suitable candidates for cooperative employee buyouts. - Collectively, these firms accounted for roughly $25 trillion in total revenues. - Aspen Institute

for - worker-owned cooperative - opportunities

for - worker-owned cooperative - attitude shift - from paycheck - to investment in future

for - worker-owned cooperatives - attitude shift - more responsibility

for - Apis & Heritage Legacy Fund employee buyout process - Apis & Heritage values the enterprise and offers seller fair price for their life work - Once purchased, they transfer the company's assets to a trust - Using private debt capital, they finance a portion of that transaction. - The trust administers the ESOP - The seller has full liquidity upfront and can retire immediately, The Legacy Fund saves seller from having to manage the complex process of selling to employees. - ESOP is a retirement account for the new employee-owhers. - After 5 years, each employee become vested, with new share allocations made each year.based on wages as a percentage of total payroll - If value of business grows, so do employee share value. - When employee-owner is ready to retire, they sell back the shares based on current valuation - new employee-owners receive training from Democracy at Work Institute (DAWI) - The trust repays debt from initial transaction on behalf of the business to Apis & Heritage and its investors who make an attractive return -

for - stats - comparison - savings - worker-owned cooperative employee vs non - Rutgers - average median-earnings household - 17K - worker-owned cooperative - 165K

for - definition - Legacy Fund - Apis & Heritage fund that converts small businesses to worker-owned cooperatives - identify well run businesses that can deliver financial returns via interest and principal repayment. - target businesses with low- and middle-income hourly workers in industries: - construction, - manufacturing, - in-home care - uplifting everyday, hardworking Americans. - Deliver - competitive, - risk-adjusted returns - with rates in the low- to mid-teens - that are comparable to traditional investments for this asset class.

for - definition - sliver tsunami

for - stats - small businesses - USA - 3 million - 10 trillion in assets - 11 million baby boomers retiring by 2035 - US - worker-owned cooperatives - potential

for - definition - Employee Led Buyout (ELBO)

= local-first-reflections

auto-nomous/poeietic permanent evergreen interpersonal social networked intentionally transparent conversational structures and colaboration and ownership for users

Study uses machine learning to analyze speech from 106 languages; compares relationships using embeddings instead of traditional historical/typological methods.

Determine partnerships with schools, regional festivals, and cultural tourism programs, these could be examples of opportunities to eliminate potential threats such as new entrants into the market.

Offering different entertainment services to consumers (cinema, dance competitions, art gallery, music) highlights The Capitol Centre's efforts to innovate and diversify its offerings.

It is important because suppliers play an important role within The Capitol Centre, such as local artists, stage technicians and service providers, and those in charge of the halls within the venue.

Allows better understanding of the competitive environment, identifying opportunities and threats, developing new strategies to improve their market position, increasing their competitiveness and differentiate themselves from other places in and around the North Bay.

Similarly, the social aspect, which reflects the cultural values of the community, is where The Capitol Centre can provide free educational programs, promote inclusion and strengthen the sense of local identity.

I would like to highlight the technological aspect as a great strategic tool for the Capitol Centre, since nowadays adapting to technology is one of the new ways to encourage community engagement and innovation, in terms of promotion, sales, attractiveness, etc.

It can help to better understand the external environment and develop strategies according to each of these six aspects, in order to take advantage of opportunities and mitigate threats, incentivizing the community and beyond outside of North Bay.

According to the following research on the analysis of the Business Analysis Environment, The Capitol Centre can benefit significantly from the strategic approach that this framework allows in understanding both the external environment and the competitive dynamics that affect its operation as a cultural institution.

eLife Assessment

In this important study, the authors set out to determine the molecular interactions between the AQP2 from Trypanosoma brucei (TbAQP2) and the trypanocidal drugs pentamidine and melarsoprol to understand how TbAQP2 mutations lead to drug resistance. Using cryo-EM, molecular dynamics simulations, and lysis assays the authors present convincing evidence that mutations in TbAQP2 make permeation of trypanocidal drugs energetically less favourable, and that this impacts the ability of drugs to achieve a therapeutic dose. Overall, this data will be of interest for those working on aquaporins, and development of trypanosomiasis drugs as well as drugs targeting aquaporins in general.

Reviewer #1 (Public review):

This study presents cryoEM-derived structures of the Trypanosome aquaporin AQP2, in complex with its natural ligand, glycerol, as well as two trypanocidal drugs, pentamidine and melarsoprol, which use AQP2 as an uptake route. The structures are high quality and the density for the drug molecules is convincing, showing a binding site in the centre of the AQP2 pore.

The authors then continue to study this system using molecular dynamics simulations. Their simulations indicate that the drugs can pass through the pore and identify a weak binding site in the centre of the pore, which corresponds with that identified through cryoEM analysis. They also simulate the effect of drug resistance mutations which suggests that the mutations reduce the affinity for drugs and therefore might reduce the likelihood that the drugs enter into the centre of the pore, reducing the likelihood that they progress through into the cell.

While the cryoEM and MD studies are well conducted, it is a shame that the drug transport hypothesis was not tested experimentally. For example, did they do cryoEM with AQP2 with drug resistance mutations and see if they could see the drugs in these maps? They might not bind, but another possibility is that the binding site shifts, as seen in Chen et al? Do they have an assay for measuring drug binding? I think that some experimental validation of the drug binding hypothesis would strengthen this paper. The authors describe in their response why these experiments are challenging.

Reviewer #2 (Public review):

Summary:

The authors present 3.2-3.7 Å cryo-EM structures of Trypanosoma brucei aquaglyceroporin-2 (TbAQP2) bound to glycerol, pentamidine or melarsoprol and combine them with extensive all-atom MD simulations to explain drug recognition and resistance mutations. The work provides a persuasive structural rationale for (i) why positively selected pore substitutions enable diamidine uptake, and (ii) how clinical resistance mutations weaken the high-affinity energy minimum that drives permeation. These insights are valuable for chemotherapeutic re-engineering of diamidines and aquaglyceroporin-mediated drug delivery.

My comments are on the MD part

Strengths:

The study

Integrates complementary cryo-EM, equilibrium and applied voltage MD simulations, and umbrella-sampling PMFs, yielding a coherent molecular-level picture of drug permeation.

Offers direct structural rationalisation of long-standing resistance mutations in trypanosomes, addressing an important medical problem.

Comments on revisions:

Most of the weaknesses have been resolved during the revision process.

Reviewer #3 (Public review):

Summary:

Recent studies have established that trypanocidal drugs, including pentamidine and melarsoprol, enter the trypanosomes via the glyceroaquaporin AQP2 (TbAQP2). Interestingly, drug resistance in trypanosomes is, at least in part, caused by recombination with the neighbouring gene, AQP3, which is unable to permeate pentamidine or melarsoprol. The effect of the drugs on cells expressing chimeric proteins is significantly reduced. In addition, controversy exists regarding whether TbAQP2 permeates the drugs like an ion channel, or whether it serves as a receptor that triggers downstream processes upon drug binding. In this study the authors set out to achieve these objectives: 1) to understand the molecular interactions between TbAQP2 and glycerol, pentamidine, and melarsoprol, and 2) to determine the mechanism by which mutations that arise from recombination with TbAQP3 result in reduced drug permeation.

The cryo-EM structures provide details of glycerol and drug binding, and show that glycerol and the drugs occupy the same space within the pore. Finally, MD simulations and lysis assays are employed to determine how mutations in TbAQP2 result in reduced permeation of drugs by making entry and exit of the drug relatively more energy-expensive. Overall, the strength of evidence used to support the author's claims is solid.

Strengths:

The cryo-EM portion of the study is strong, and while the overall resolution of the structures is in the 3.5Å range, the local resolution within the core of the protein and the drug binding sites is considerably higher (~2.5Å).<br /> I also appreciated the MD simulations on the TbAQP2 mutants and the mechanistic insights that resulted from this data.

Weaknesses:

(1) The authors do not provide any experimental validation the drug binding sites in TbAQP2 due to lacking resources. However, the claims have been softened in the revised paper.

Author response:

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

Reviewer #1 (Public review): 

This study presents cryoEM-derived structures of the Trypanosome aquaporin AQP2, in complex with its natural ligand, glycerol, as well as two trypanocidal drugs, pentamidine and melarsoprol, which use AQP2 as an uptake route. The structures are high quality, and the density for the drug molecules is convincing, showing a binding site in the centre of the AQP2 pore. 

The authors then continue to study this system using molecular dynamics simulations. Their simulations indicate that the drugs can pass through the pore and identify a weak binding site in the centre of the pore, which corresponds with that identified through cryoEM analysis. They also simulate the effect of drug resistance mutations, which suggests that the mutations reduce the affinity for drugs and therefore might reduce the likelihood that the drugs enter into the centre of the pore, reducing the likelihood that they progress through into the cell. 

While the cryoEM and MD studies are well conducted, it is a shame that the drug transport hypothesis was not tested experimentally. For example, did they do cryoEM with AQP2 with drug resistance mutations and see if they could see the drugs in these maps? They might not bind, but another possibility is that the binding site shifts, as seen in Chen et al. 

TbAQP2 from the drug-resistant mutants does not transport either melarsoprol or pentamidine and there was thus no evidence to suggest that the mutant TbAQP2 channels could bind either drug. Moreover, there is not a single mutation that is characteristic for drug resistance in TbAQP2: references 12–15 show a plethora of chimeric AQP2/3 constructs in addition to various point mutations in laboratory strains and field isolates. In reference 17 we describe a substantial number of SNPs that reduced pentamidine and melarsoprol efficacy to levels that would constitute clinical resistance to acceptable dosage regimen. It thus appears that there are many and diverse mutations that are able to modify the protein sufficiently to induce resistance, and likely in multiple different ways, including the narrowing of the pore, changes to interacting amino acids, access to the pore etc. We therefore did not attempt to determine the structures of the mutant channels because we did not think that in most cases we would see any density for the drugs in the channel, and we would be unable to define ‘the’ resistance mechanism if we did in the case of one individual mutant TbAQP2. Our MD data suggests that pentamidine binding affinity is in the range of 50-300 µM for the mutant TbAQP2s selected for that test (I110W and L258Y/L264R), i.e. >1000-fold higher than TbAQP2WT. Thus these structures will be exceedingly challenging to determine with pentamidine in the pore but, of course, until the experiment has been tried we will not know for sure.

Do they have an assay for measuring drug binding? 

We tried many years ago to develop a ³H-pentamidine binding assay to purified wild type TbAQP2 but we never got satisfactory results even though the binding should be in the doubledigit nanomolar range. This may be for any number of technical reasons and could also be partly because flexible di-benzamidines bind non-specifically to proteins at µM concentrations giving rise to high background. Measuring binding to the mutants was not tested given that they would be binding pentamidine in the µM range. If we were to pursue this further, then isothermal titration calorimetry (ITC) may be one way forward as this can measure µM affinity binding using unlabelled compounds, although it uses a lot of protein and background binding would need to be carefully assessed; see for example our work on measuring tetracycline binding to the tetracycline antiporter TetAB (https://doi.org/10.1016/j.bbamem.2015.06.026 ). Membrane proteins are also particularly tricky for this technique as the chemical activity of the protein solution must be identical to the chemical activity of the substrate solution which titrates in the molecule binding to the protein; this can be exceedingly problematic if any free detergent remains in the purified membrane protein. Another possibility may be fluorescence polarisation spectroscopy, although this would require fluorescently labelling the drugs which would very likely affect their affinity for TbAQP2 and how they interact with the wild type and mutant proteins – see the detailed SAR analysis in Alghamdi et al. 2020 (ref. 17). As you will appreciate, it would take considerable time and effort to set up an assay for measuring drug binding to mutants and is beyond the current scope of the current work.

I think that some experimental validation of the drug binding hypothesis would strengthen this paper. Without this, I would recommend the authors to soften the statement of their hypothesis (i.e, lines 65-68) as this has not been experimentally validated.

We agree with the referee that direct binding of drugs to the mutants would be very nice to have, but we have neither the time nor resources to do this. We have therefore softened the statement on lines 65-68 to read ‘Drug-resistant TbAQP2 mutants are still predicted to bind pentamidine, but the much weaker binding in the centre of the channel observed in the MD simulations would be insufficient to compensate for the high energy processes of ingress and egress, hence impairing transport at pharmacologically relevant concentrations.’ 

Reviewer #2 (Public review): 

Summary: 

The authors present 3.2-3.7 Å cryo-EM structures of Trypanosoma brucei aquaglyceroporin-2 (TbAQP2) bound to glycerol, pentamidine, or melarsoprol and combine them with extensive allatom MD simulations to explain drug recognition and resistance mutations. The work provides a persuasive structural rationale for (i) why positively selected pore substitutions enable diamidine uptake, and (ii) how clinical resistance mutations weaken the high-affinity energy minimum that drives permeation. These insights are valuable for chemotherapeutic re-engineering of diamidines and aquaglyceroporin-mediated drug delivery. 

My comments are on the MD part. 

Strengths: 

The study 

(1) Integrates complementary cryo-EM, equilibrium, applied voltage MD simulations, and umbrella-sampling PMFs, yielding a coherent molecular-level picture of drug permeation. 

(2) Offers direct structural rationalisation of long-standing resistance mutations in trypanosomes, addressing an important medical problem. 

Weaknesses: 

Unphysiological membrane potential. A field of 0.1 V nm ¹ (~1 V across the bilayer) was applied to accelerate translocation. From the traces (Figure 1c), it can be seen that the translocation occurred really quickly through the channel, suggesting that the field might have introduced some large changes in the protein. The authors state that they checked visually for this, but some additional analysis, especially of the residues next to the drug, would be welcome. 

This is a good point from the referee, and we thank them for raising it. It is common to use membrane potentials in simulations that are higher than the physiological value, although these are typically lower than used here. The reason we used the higher value was to speed sampling and it still took 1,400 ns for transport in the physiologically correct direction, and even then, only in 1/3 repeats. Hence this choice of voltage was probably necessary to see the effect. The exceedingly slow rate of pentamidine permeation seen in the MD simulation was consistent with the experimental observations, as discussed in Alghamdi et al (2020) [ref. 17] where we estimated that TbAQP2-mediated pentamidine uptake in T. brucei bloodstream forms proceeds at just 9.5×10⁵ molecules/cell/h; the number of functional TbAQP2 units in the plasma membrane is not known but their location is limited to the small flagellar pocket (Quintana et al. PLoS Negl Trop Dis 14, e0008458 (2020)). 

The referee is correct that it is important to make sure that the applied voltage is not causing issues for the protein, especially for residues in contact with the drug. We have carried out RMSF analysis to better test this. The data show that comparing our simulations with the voltage applied to the monomeric MD simulations + PNTM with no voltage reveals little difference in the dynamics of the drug-contacting residues. 

We have added these new data as Supplementary Fig12b with a new legend (lines1134-1138) 

‘b, RMSF calculations were run on monomeric TbAQP2 with either no membrane voltage or a 0.1V nm^-1 voltage applied (in the physiological direction). Shown are residues in contact with the pentamidine molecule, coloured by RMSF value. RMSF values are shown for residues Leu122, Phe226, Ile241, and Leu264. The data suggest the voltage has little impact on the flexibility or stability of the pore lining residues.’

We have also added the following text to the manuscript (lines 524-530):

‘Membrane potential simulations were run using the computational electrophysiology protocol. An electric field of 0.1 V/nm was applied in the z-axis dimension only, to create a membrane potential of about 1 V (see Fig. S10a). Note that this is higher than the physiological value of 87.1 ± 2.1 mV at pH 7.3 in bloodstream T. brucei, and was chosen to improve the sampling efficiency of the simulations. The protein and lipid molecules were visually confirmed to be unaffected by this voltage, which we quantify using RMSF analysis on pentamidine-contacting residues (Fig. S12b).’ 

Based on applied voltage simulations, the authors argue that the membrane potential would help get the drug into the cell, and that a high value of the potential was applied merely to speed up the simulation. At the same time, the barrier for translocation from PMF calculations is ~40 kJ/mol for WT. Is the physiological membrane voltage enough to overcome this barrier in a realistic time? In this context, I do not see how much value the applied voltage simulations have, as one can estimate the work needed to translocate the substrate on PMF profiles alone. The authors might want to tone down their conclusions about the role of membrane voltage in the drug translocation.

We agree that the PMF barriers are considerable, however we highlight that other studies have seen similar landscapes, e.g. PMID 38734677 which saw a barrier of ca. 10-15 kcal/mol (ca. 4060 kJ/mol) for PNTM transversing the channel. This was reduced by ca. 4 kcal/mol when a 0.4 V nm ¹ membrane potential was applied, so we expect a similar effect to be seen here. 

We have updated the Results to more clearly highlight this point and added the following text (lines 274-275):

We note that previous studies using these approaches saw energy barriers of a similar size, and that these are reduced in the presence of a membrane voltage[17,31].’ 

Pentamidine charge state and protonation. The ligand was modeled as +2, yet pKa values might change with the micro-environment. Some justification of this choice would be welcome. 

Pentamidine contains two diamidine groups and each are expected to have a pKa above 10 in solution (PMID: 20368397), suggesting that the molecule will carry a +2 charge. Using the +2 charge is also in line with previous MD studies (PMID: 32762841). We have added the following text to the Methods (lines 506-509):

‘The pentamidine molecule used existing parameters available in the CHARMM36 database under the name PNTM with a charge state of +2 to reflect the predicted pKas of >10 for these groups [73] and in line with previous MD studies[17].’

We note that accounting for the impact of the microenvironment is an excellent point – future studies might employ constant pH calculations to address this.

The authors state that this RMSD is small for the substrate and show plots in Figure S7a, with the bottom plot being presumably done for the substrate (the legends are misleading, though), levelling off at ~0.15 nm RMSD. However, in Figure S7a, we see one trace (light blue) deviating from the initial position by more than 0.2 nm - that would surely result in an RMSD larger than 0.15, but this is somewhat not reflected in the RMSD plots. 

The bottom plot of Fig. S9a (previously Fig. S7a) is indeed the RMSD of the drug (in relation to the protein). We have clarified the legend with the following text (lines 1037-1038): ‘… or for the pentamidine molecule itself, i.e. in relation to the Cα of the channel (bottom).’ 

With regards the second comment, we assume the referee is referring to the light blue trace from Fig S9c. These data are actually for the monomeric channel rather than the tetramer. We apologise for not making this clearer in the legend. We have added the word ‘monomeric’ (line 1041).

Reviewer #3 (Public review): 

Summary: 

Recent studies have established that trypanocidal drugs, including pentamidine and melarsoprol, enter the trypanosomes via the glyceroaquaporin AQP2 (TbAQP2). Interestingly, drug resistance in trypanosomes is, at least in part, caused by recombination with the neighbouring gene, AQP3, which is unable to permeate pentamidine or melarsoprol. The effect of the drugs on cells expressing chimeric proteins is significantly reduced. In addition, controversy exists regarding whether TbAQP2 permeates drugs like an ion channel, or whether it serves as a receptor that triggers downstream processes upon drug binding. In this study the authors set out to achieve three objectives: 

(1) to determine if TbAQP2 acts as a channel or a receptor,

We should clarify here that this was not an objective of the current manuscript as the transport activity has already been extensively characterised in the literature, as described in the introduction.

(2) to understand the molecular interactions between TbAQP2 and glycerol, pentamidine, and melarsoprol, and 

(3) to determine the mechanism by which mutations that arise from recombination with TbAQP3 result in reduced drug permeation. 

Indeed, all three objectives are achieved in this paper. Using MD simulations and cryo-EM, the authors determine that TbAQP2 likely permeates drugs like an ion channel. The cryo-EM structures provide details of glycerol and drug binding, and show that glycerol and the drugs occupy the same space within the pore. Finally, MD simulations and lysis assays are employed to determine how mutations in TbAQP2 result in reduced permeation of drugs by making entry and exit of the drug relatively more energy-expensive. Overall, the strength of evidence used to support the author's claims is solid. 

Strengths: 

The cryo-EM portion of the study is strong, and while the overall resolution of the structures is in the 3.5Å range, the local resolution within the core of the protein and the drug binding sites is considerably higher (~2.5Å). 

I also appreciated the MD simulations on the TbAQP2 mutants and the mechanistic insights that resulted from this data. 

Weaknesses: 

(1) The authors do not provide any empirical validation of the drug binding sites in TbAQP2. While the discussion mentions that the binding site should not be thought of as a classical fixed site, the MD simulations show that there's an energetically preferred slot (i.e., high occupancy interactions) within the pore for the drugs. For example, mutagenesis and a lysis assay could provide us with some idea of the contribution/importance of the various residues identified in the structures to drug permeation. This data would also likely be very valuable in learning about selectivity for drugs in different AQP proteins.

On a philosophical level, we disagree with the requirement for ‘validation’ of a structure by mutagenesis. It is unclear what such mutagenesis would tell us beyond what was already shown experimentally through ³H-pentamidine transport, drug sensitivity and lysis assays i.e. a given mutation will impact permeation to a certain extent. But on the structural level, what does mutagenesis tell us? If a bulky aromatic residue that makes many van der Waals interactions with the substrate is changed to an alanine residue and transport is reduced, what does this mean? It would confirm that the phenylalanine residue is very likely indeed making van der Waals contacts to the substrate, but we knew that already from the WT structure. And if it doesn’t have any effect? Well, it could mean that the van der Waals interactions with that particular residue are not that important or it could be that the substrate has changed its positions slightly in the channel and the new pose has similar energy of interactions to that observed in the wild type channel. Regardless of the result, any data from mutagenesis would be open to interpretation and therefore would not impact on the conclusions drawn in this manuscript. We might not learn anything new unless all residues interacting with the substrate are mutated, the structure of each mutant was determined and MD simulations were performed for all, which is beyond the scope of this work. Even then, the value for understanding clinical drug resistance would be limited, as this phenomenon has been linked to various chimeric rearrangements with adjacent TbAQP3 (references 12–15), each with a structure distinct from TbAQP2 with a single SNP. We also note that the recent paper by Chen et al. did not include any mutagenesis of the drug binding sites in TbAQP2 in their analysis of TbAQP2, presumably for similar reasons as discussed above.

(2) Given the importance of AQP3 in the shaping of AQP2-mediated drug resistance, I think a figure showing a comparison between the two protein structures/AlphaFold structures would be beneficial and appropriate

We agree that the comparison is of considerably interest and would contribute further to our understanding of the unique permeation capacities of TbAQP2. As such, we followed the reviewer’s suggestion and made an AlphaFold model of TbAQP3 and compared it to our structures of TbAQP2. The RMSD is 0.6 Å to the pentamidine-bound TbAQP2, suggesting that the fold of TbAQP3 has been predicted well, although the side chain rotamers cannot be assessed for their accuracy. Previous work has defined the selectivity filter of TbAQP3 to be formed by W102, R256, Y250. The superposition of the TbAQP3 model and the TbAQP2 pentamidine-bound structure shows that one of the amine groups is level with R256 and that there is a clash with Y250 and the backbone carbonyl of Y250, which deviates in position from the backbone of TbAQP2 in this region. There is also a clash with Ile252. 

Although these observations are indeed interesting, on their own they are highly preliminary and extensive further work would be necessary to draw any convincing conclusions regarding these residues in preventing uptake of pentamidine and melarsoprol. The TbAQP3 AlphaFold model would need to be verified by MD simulations and then we would want to look at how pentamidine would interact with the channel under different experimental conditions like we have done with TbAQP2. We would then want to mutate to Ala each of the residues singly and in combination and assess them in uptake assays to verify data from the MD simulations. This is a whole new study and, given the uncertainties surrounding the observations of just superimposing TbAQP2 structure and the TbAQP3 model, we feel that, regrettably, this is just too speculative to add to our manuscript. 

(3) A few additional figures showing cryo-EM density, from both full maps and half maps, would help validate the data. 

Two new Supplementary Figures have been made, on showing the densities for each of the secondary structure elements (the new Figure S5) and one for the half maps showing the ligands (the new Figure S6). All the remaining supplementary figures have been renamed accordingly.

(4) Finally, this paper might benefit from including more comparisons with and analysis of data published in Chen et al (doi.org/10.1038/s41467-024-48445-4), which focus on similar objectives. Looking at all the data in aggregate might reveal insights that are not obvious from either paper on their own. For example, melarsoprol binds differently in structures reported in the two respective papers, and this may tell us something about the energy of drug-protein interactions within the pore. 

We already made the comparisons that we felt were most pertinent and included a figure (Fig. 5) to show the difference in orientation of melarsoprol in the two structures. We do not feel that any additional comparison is sufficiently interesting to be included. As we point out, the structures are virtually identical (RMSD 0.6 Å) and therefore there are no further mechanistic insights we would like to make beyond the thorough discussion in the Chen et al paper.

Reviewer #1 (Recommendations for the authors): 

(1) Line 65 - I don't think that the authors have tested binding experimentally, and so rather than 'still bind', I think that 'are still predicted to bind' is more appropriate. 

Changed as suggested

(2) Line 69 - remove 'and' 

Changed as suggested

(3) Line 111 - clarify that it is the protein chain which is 'identical'. Ligands not. 

Changed to read ‘The cryo-EM structures of TbAQP2 (excluding the drugs/substrates) were virtually identical…

(4) Line 186 - make the heading of this section more descriptive of the conclusion than the technique? 

We have changed the heading to read: ‘Molecular dynamics simulations show impaired pentamidine transport in mutants’

Reviewer #2 (Recommendations for the authors): 

(1) Methods - a rate of 1 nm per ns is mentioned for pulling simulations, is that right? 

Yes, for the generation of the initial frames for the umbrella sampling a pull rate of 1 nm/ns was used in either an upwards or downwards z-dimension

(2) Figure S9 and S10 have their captions swapped. 

The captions have been swapped to their proper positions.

(3) Methods state "40 ns per window" yet also that "the first 50 ns of each window was discarded as equilibration". 

Well spotted - this line should have read “the first 5 ns of each window was discarded as equilibration”. This has been corrected (line 541).

Reviewer #3 (Recommendations for the authors): 

(1) Abstract, line 68-70: incomplete sentence.

The sentence has been re-written: ‘The structures of drug-bound TbAQP2 represent a novel paradigm for drug-transporter interactions and are a new mechanism for targeting drugs in pathogens and human cells.

(2) Line 312-313: The paper you mention here came out in May 2024 - a year ago. I appreciate that they reported similar structural data, but for the benefit of the readers and the field, I would recommend a more thorough account of the points by which the two pieces of work differ. Is there some knowledge that can be gleaned by looking at all the data in the two papers together? For example, you report a glycerol-bound structure while the other group provides an apo one. Are there any mechanistic insights that can be gained from a comparison?

We already made the comparisons that we felt were most pertinent and included a figure (Fig. 5) to show the difference in orientation of melarsoprol in the two structures. We do not feel that any additional comparison is sufficiently interesting to be included. As we point out, the structures are virtually identical (RMSD 0.6 Å) and therefore there are no further mechanistic insights we would like to make beyond the thorough discussion in the Chen et al paper.

(3) Similarly, you can highlight the findings from your MD simulations on the TbAQP2 drug resistance mutants, which are unique to your study. How can this data help with solving the drug resistance problem?

New drugs will need to be developed that can be transported by the mutant chimera AQP2s and the models from the MD simulations will provide a starting point for molecular docking studies. Further work will then be required in transport assays to optimise transport rather than merely binding. However, the fact that drug resistance can also arise through deletion of the AQP2 gene highlights the need for developing new drugs that target other proteins.

(4) A glaring question that one has as a reader is why you have not attempted to solve the structures of the drug resistance mutants, either in complex with the two compounds or in their apo/glycerol-bound form? To be clear, I am not requesting this data, but it might be a good idea to bring this up in the discussion.

TbAQP2 containing the drug-resistant mutants does not transport either melarsoprol or pentamidine (Munday et al., 2014; Alghamdi et al., 2020); there was thus no evidence to suggest that the mutant TbAQP2 channels could bind either drug. We therefore did not attempt to determine the structures of the mutant channels because we did not think that we would see any density for the drugs in the channel. Our MD data suggests that pentamidine binding affinity is in the range of 50-300 µM for the mutant TbAQP2, supporting the view that getting these structures would be highly challenging, but of course until the experiment is tried we will not know for sure.

We also do not think we would learn anything new about doing structures of the drug-free structures of the transport-negative mutants of TbAQP2. The MD simulations have given novel insights into why the drugs are not transported and we would rather expand effort in this direction and look at other mutants rather than expend further effort in determining new structures.

(5) Line 152-156: Is there a molecular explanation for why the TbAQP2 has 2 glycerol molecules captured in the selectivity filter while the PfAQP2 and the human AQP7 and AQP10 have 3?

The presence of glycerol molecules represents local energy minima for binding, which will depend on the local disposition of appropriate hydrogen bonding atoms and hydrophobic regions, in conjunction with the narrowness of the channel to effectively bind glycerol from all sides. It is noticeable that the extracellular region of the channel is wider in TbAQP2 than in AQP7 and AQP10, so this may be one reason why additional ordered glycerol molecules are absent, and only two are observed. Note also that the other structures were determined by X-ray crystallography, and the environment of the crystal lattice may have significantly decreased the rate of diffusion of glycerol, increasing the likelihood of observing their electron densities.

(6) I would also think about including the 8JY7 (TbAQP2 apo) structure in your analysis.

We included 8JY7 in our original analyses, but the results were identical to 8JY6 and 8JY8 in terms of the protein structure, and, in the absence of any modelled substrates in 8JY7 (the interesting part for our manuscript), we therefore have not included the comparison.

(7) I also think, given the importance of AQP3 in this context, it would be really useful to have a comparison with the AQP3 AlphaFold structure in order to examine why it does not permeate drugs.

We made an AlphaFold model of TbAQP3 and compared it to our structures of TbAQP2. The RMSD is 0.6 Å to the pentamidine-bound TbAQP2, suggesting that the fold of TbAQP3 has been predicted well, although the side chain rotamers cannot be assessed for their accuracy. Previous work has defined the selectivity filter of TbAQP3 to be formed by W102, R256, Y250. The superposition of the TbAQP3 model and the TbAQP2 pentamidine-bound structure shows that one of the amine groups is level with R256 and that there is a clash with Y250 and the backbone carbonyl of Y250, which deviates in position from the backbone of TbAQP2 in this region. There is also a clash with Ile252. 

Although these observations are interesting, on their own they are preliminary in the extreme and extensive further work will be necessary to draw any convincing conclusions regarding these residues in preventing uptake of pentamidine and melarsoprol. The TbAQP3 AlphaFold model would need to be verified by MD simulations and then we would want to look at how pentamidine would interact with the channel under different experimental conditions like we have done with TbAQP2. We would then want to mutate to Ala each of the residues singly and in combination and assess them in uptake assays to verify data from the MD simulations. This is a whole new study and, given the uncertainties surrounding the observations of just superimposing TbAQP2 structure and the TbAQP3 model, we feel this is just too speculative to add to our manuscript. 

(8) To validate the densities representing glycerol and the compounds, you should show halfmap densities for these. 

A new figure, Fig S6 has been made to show the half-map densities for the glycerol and drugs.

(9) I would also like to see the density coverage of the individual helices/structural elements. 

A new figure, Fig S5 has been made to show the densities for the structural elements.

(10) While the LigPlot figure is nice, I think showing the data (including the cryo-EM density) is necessary validation.

The LigPlot figure is a diagram (an interpretation of data) and does not need the densities as these have already been shown in Fig. 1c (the data).

(11) I would recommend including a figure that illustrates the points described in lines 123-134.

All of the points raised in this section are already shown in Fig. 2a, which was referred to twice in this section. We have added another reference to Fig.2a on lines 134-135 for completeness.

(12) Line 202: I would suggest using "membrane potential/voltage" to avoid confusion with mitochondrial membrane potential. 

We have changed this to ‘plasma membrane potential’ to differentiate it from mitochondrial membrane potential.

(13) Figure 4: Label C.O.M. in the panels so that the figure corresponds to the legend. 

We have altered the figure and added and explanation in the figure legend (lines 716-717):

‘Cyan mesh shows the density of the molecule across the MD simulation. and the asterisk shows the position of the centre of mass (COM).’

(14) Figure S2: Panels d and e appear too similar, and it is difficult to see the stick representation of the compound. I would recommend either using different colours or showing a close-up of the site.

We have clarified the figure by including two close-up views of the hot-spot region, one with melarsoprol overlaid and one with pentamidine overlaid

(15) Figure S2: Typo in legend: 8YJ7 should be 8JY7.

Changed as suggested  

(16) Figure S3 and Figure S4: Please clarify which parts of the process were performed in cryoSPARC and which in Relion. 

Figure S3 gives an overview of the processing and has been simplified to give the overall picture of the procedures. All of the details were included in the Methods section as other programmes are used, not just cryoSPARC and Relion. Given the complexities of the processing, we have referred the readers to the Methods section rather than giving confusing information in Fig. S3.

We have updated the figure legend to Fig. S4 as requested.

(17) Figure S9 and Figure S10: The legends are swapped in these two figures.

The captions have been swapped to their proper positions.

(18) For ease of orientation and viewing, I would recommend showing a vertical HOLE plot aligned with an image of the AQP2 pore. 

The HOLE plot has been re-drawn as suggest (Fig. S2)

These theories bridge the gap between philosophical ethics and real-world business situations. Their main goal is to guide ethical decision-making in a way that is practical for the business environment.

you dont have to agree with everything in a discourse community to be apart of it

they share common goals

may work closer in a job setting

discourse communities vary, but everyone belongs to something

difference between linguistics and discourse, dont have to speak the same language

number 5: certain abbreviations are specific to a discourse community

the discourse community is what is alike, not everything else

more skilled people in a discourse community help newer members become more acclimated to the group