Reviewer #1 (Public review):

Summary:

This is an interesting study that addresses whether mitochondrial DNA (mitoDNA) variants impact telomere length (TL), which may be relevant to potential maternal inheritance of TL in offspring. The study addresses this question using an innovative cybrid model approach in which mitochondria from donor platelets from 7 individuals that vary in TL and differ in mitoDNA variants are fused with 143B cells that lack mitochondria. MitoDNA variants that exhibited reduced complex I activity showed telomere shortening in cybrids, and increased telomere dysfunction. Interestingly, these phenotypes could be reduced with NAC antioxidant and NAD+ supplementation, suggesting that ROS and oxidative DNA damage at telomeres contributed to the telomere shortening. They further showed that cybrids with lower levels of ROS correlated with longer TL in the lymphocytes of the mitochondrial donors.

This study has many strengths, including the use of cybrid technology, which allows for analysis of how mitochondria and mtDNA variants from various human donors influence telomere length when introduced into the same genetic background of 143B mitochondria deficient cells. Other strengths include the use of human donors that vary in leukocyte telomere length, and analysis of multiple endpoints of telomere integrity and mitochondrial function in the resulting cybrids. This study provides convincing evidence that mtDNA variants influence TL, through a mechanism involving mitochondrial derived ROS, potentially causing telomeric oxidative damage. These results may help explain prior work demonstrating a link between mitochondrial function and telomere length regulation during embryogenesis and development. The data are robust, and the manuscript is well written. Overall, the results support the conclusions and will likely have an impact on the field of mitochondrial function and telomere length regulation, with important implications for development, aging, and aging-related diseases.

Given the heterogeneity in TL and mtDNA variants in the human population, the conclusion could be further strengthened by increasing the number of donors and cybrids analyzed. Other minor weaknesses stem from the limited number of samples. For example, the study presented limited telomere FISH metaphase analysis of chromosome end fusions, telomere damage, and telomere attrition rates due to low sample numbers. Furthermore, the proposal that acute oxidative stress was induced by a metabolic shift from glycolysis to OXPHOS in vitro is based primarily on assumptions from prior work, rather than direct demonstration in the current study. However, results with the NAC antioxidant combined with NR strongly support this assumption. Overall, the findings from this study are convincing and provide a solid foundation for expanding this analysis in the future to include a greater number of samples from donors for cybrid analysis.

Reviewer #2 (Public review):

Summary:

The authors aim to determine whether mitochondrial genotype influence telomere length. By generating cybrids harboring different mitochondrial backgrounds, the authors seek to establish a mechanistic link between mitochondrial status and telomere biology.

Strengths:

A major strength of the study is the use of cybrid technology, which provides a great approach to investigate the role of mitochondrial DNA independently of the nuclear genome. The authors also employ multiple complementary assays to assess telomere-related phenotypes associated with mitochondrial dysfunction. Together, these experiments generate an interesting dataset that will be of value to researchers interested in the intersection between mitochondrial biology, genome stability, aging and development. These results also build on previous work supporting roles for ROS/mitochondria in driving telomere shortening.

Overall, this study provides interesting evidence linking mitochondrial background to telomere biology. The cybrid models represent a useful resource for the field and the work raises important questions regarding mitochondria-telomere communication.

Comments on revised version.

In my view, authors have answered satisfactorily my concerns.

Reviewer #3 (Public review):

Strengths:

Mahieu and colleagues ask whether mitochondrial genetic variation helps explain differences in human telomere length. They selected seven donors from a Flow-FISH reference cohort of 491 individuals and compared their mitochondrial backgrounds in cybrids sharing the same nuclear background.

The combination of telomere and mitochondrial assays is a strength. The association between low complex I activity and telomere shortening, together with the NAC/NR rescue in Cyb2, provides useful evidence linking mitochondrial function to telomere maintenance. The correlation between donor telomere length and cybrid ROS is also interesting, although the cybrid results are more persuasive than the claims about inheritance.

Weaknesses:

(1) The authors have clarified that the seven donors include intermediate telomere lengths. Their analyses excluding donors #1, #3 and #6 are reassuring, since the correlation persists in each case. However, these results appear only in the response, and the full leave-one-out analysis is not reported. The small, selected sample still limits how broadly the correlation can be interpreted.

(2) The revised manuscript makes the distinction between the cybrid and donor findings clearer. However, CI activity shows no significant association with donor telomere length. The proposed PARP1 mechanism also remains untested: the authors acknowledge that they did not establish whether PARP1 is required for the NAC/NR rescue, but still describe its role too strongly.

(3) The authors appropriately acknowledge that the ATP6 A177T association is inconclusive (p = 0.153). Some inheritance claims remain stronger than this warrants. The family observations are too limited to separate mitochondrial from nuclear or shared environmental effects, and the cybrid experiments do not isolate ATP6 A177T from the other variants transferred with it.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

This is an interesting study that addresses whether mitochondrial DNA (mitoDNA) variants impact telomere length (TL), which may be relevant to potential maternal inheritance of TL in offspring. The study addresses this question using a cybrid model approach in which mitochondria from donor platelets from 7 individuals that vary in TL and differ in mitoDNA variants are introduced into 143B cells that lack mitochondria. MitoDNA variants that exhibited reduced complex I activity showed telomere shortening in cybrids and increased telomere dysfunction. Interestingly, these phenotypes could be reduced with NAC antioxidant and NAD+ supplementation, suggesting that ROS and oxidative DNA damage at telomeres contributed to the telomere shortening. They further showed that cybrids with lower levels of ROS correlated with longer TL in the lymphocytes of the mitochondrial donors.

Strengths:

This study provides compelling evidence that mtDNA variants influence TL through a mechanism involving mitochondrial-derived ROS, potentially causing telomeric oxidative damage. The data are robust, and the manuscript is well written. However, the study could be strengthened by addressing the following questions and minor weaknesses below.

We thank the reviewer for this very positive evaluation of our work.

Weaknesses:

(1) Introduction. Line 81, the relationship between TL and the risk of lymphoid and myeloid leukemia is not straightforward. POT1 variants associated with long TL increase the risk for lymphoid and myeloproliferative neoplasms (see PMID: 41564438 for example).

We appreciate the reviewer raising the important link between pathogenic POT1 variants and lymphoid malignancies driven by elongated telomeres. We would like to clarify, however, that our introduction focused not on the pathological telomere attrition characteristic of telomere biology disorders, but rather on the natural, non-pathological variations observed in individuals with baseline telomere lengths on the shorter end of the spectrum.

Nevertheless, the two following publications by the Armanios lab have been added in the introduction to underscore the complex relationship between telomere length and tumorigenesis:

DeBoy EA et al. Familial clonal hematopoiesis in a long telomere syndrome. 2023. N Engl J Med 388, 2422-2433.

Davidson-Swinton HR et al. Lymphoid malignancy and clonality in the POT1-mediated long telomere syndrome. 2026. Blood 147, 2226-2237.

Here is the corresponding sentence in the revised manuscript:

“However, some pathogenic mutations in POT1 gene, linked to very long telomeres, were also previously associated with increased risk for lymphoid and myeloproliferative neoplasms (5,6), underscoring the complex relationship between TL and cancer predisposition.”

(2) Figure 1. Since sex also influences TL, it would be good to know the sex of the selected individuals or explain why this is not necessary.

Because this study investigated the potential maternal inheritance of TL via the mitochondrial genome, our cohort consisted predominantly of female donors (6/7). Donor #5, the husband of Donor #7, was the only male included. We have added the information in the text:

“All were of European origin, thus excluding a possible bias linked to African ethnic groups (7), and the seven donors belonged to distinct mitochondrial subhaplogroups (Fig 1D). All but one (#5) were female donors.”

Notably, our analysis revealed minimal influence of sex on TL, which cannot account for the observed differences between the extreme groups.

(3) Please include a description of the 143B cells that were used for cybrid formation in the Results section when introducing the cybrids.

We did it:

“To evaluate the impact of distinct mitochondrial genomes on TL in human cells, we used a transmitochondrial cybrid strategy in which mtDNA-depleted Rho⁰ (ρ⁰) cells, derived from 143B human osteosarcoma cancer cells, are fused with donor platelets, thereby restoring functional mitochondria carrying donor-specific mtDNA variants (21).”

(4) Lines 155-156. The authors note that cybrids from donors 1 and 2 show "pronounced" telomere damage. This result indicates an increase in 53BP1-positive telomeres, which could be indicative of telomere dysfunction or damage. Quantification of the increased chromosome end fusions for cybrids 1 and 2 would strengthen the result.

We thank the reviewer for this suggestion. Following their advice, we used our metaphase spread FISH analyses to quantify chromosome end fusions in the parental 143B Rho0, Cyb1 and Cyb6 cells. However, because Cyb2 metaphase spreads were of insufficient quality for adequate chromosome analysis, we restricted our telomere fusion comments exclusively to Cyb1 and included the quantifications in our revised manuscript (Fig 2E).

Do the increased fusions correlate with an increase in telomere signal-free ends? These should be apparent in the telomere FISH images of metaphase chromosomes.

While this is a strong argument, the telomeres within these cybrid models are critically short. Consequently, the FISH signal intensity falls below the threshold required for reliable quantification of telomere-free ends.

(5) Lines 168-169. What is the evidence that the "in vitro metabolic shift" causes acute oxidative stress?

The reviewer is correct that we have not formally demonstrated this in our experimental system. Instead, our assumption was based on the fact that the initial phase of Rho0 cell repopulation involves a temporary ROS burst, partly driven by incompletely assembled ETC supercomplexes that are known to elevate ROS levels (Maranzana et al., 2013). This is supported by our experimental observation that the NAC antioxidant, combined with NR, potently inhibits telomere shortening in cybrids with low CI activity. We added the following sentence in the revised manuscript:

“When mitochondria are replenished in ρ⁰ cells during the formation of transmitochondrial cybrids, a shift from glycolysis to OXPHOS occurs (Fig S4A-B) and cells need to adapt to this metabolic change. Notably, the initial phase of ρ⁰ cell repopulation involves a temporary ROS burst, partly driven by incompletely assembled ETC supercomplexes that are known to elevate ROS levels (22).”

(6) Why did the elevated ROS in cybrid #3 (Figure 4C) not translate to shorter telomeres in the cybrid (Figure 2A)? Perhaps there is a difference between factors that determine TL in the cybrid vs the donor's lymphocytes?

We cannot fully explain the discrepancy, but we indeed suspect in vivo oxidative stress differs significantly from cell cultures (21% O₂). Additionally, early cybrid replenishment involves an unknown telomere elongation step that may offset mitochondrial ROS-induced shortening.

To further investigate this question, we measured telomere length across varying population doublings (PDs) and observed the following shortening after 66-67 PDs:

- Cyb3: about 1.6 kb reduction

- Cyb6: about 1.2 kb reduction

- Cyb7: about 0.8 kb reduction

These results suggest that the rate of telomere shortening in culture may be higher in Cyb3 cells, possibly due to increased ROS levels.

We included these results as Supplementary Figure 7 and have discussed these data in the revised manuscript:

“Consistently, tracking TL over sequential population doublings revealed that Cyb3 cells are more vulnerable to progressive telomere shortening than Cyb6 or Cyb7 cells (Fig S7). Together, these findings support a role for mitochondrial ROS in human TL regulation.”

In Figure 4B, it appears that the statistical comparisons for mitochondrial superoxide are all relative to Cyb3. If so, why are the comparisons not with the parental 143B rho0 cell line? Please clarify.

We excluded Rho0 cells from our superoxide measurements because these cells were grown in a different culture medium. Instead, we focused on comparing mitochondrial ROS across cybrids to accurately correlate these values with the telomere length of the corresponding donors’ lymphocytes. Consequently, Rho0 cell measurements would not have contributed to this correlation analysis.

We have discussed this in the revised manuscript:

“Mitochondrial superoxide production was measured by electron paramagnetic resonance (27) in clones #1 of all cybrids and in 143B cells (Fig 4B). We excluded 143B ρ⁰ cells from our superoxide measurements because these cells were grown in a different culture medium.”

(7) Given the heterogeneity in TL and mtDNA variants in the human population, the conclusions could be further strengthened by increasing the number of donors and cybrids analyzed. However, there are admittedly practical factors. Overall, these findings are compelling and provide a solid foundation for expanding this analysis in the future. This is more of a comment than a weakness.

We thank the reviewer for this constructive feedback and entirely agree that including more donors would have added depth to our findings. While we acknowledge this limitation, pursuing this further is currently impossible without acquiring new ethical approvals and establishing fresh collaborations with clinicians.

Reviewer #2 (Public review):

Summary:

The authors aim to determine whether mitochondrial genotype influences telomere length. By generating cybrids harboring different mitochondrial backgrounds, the authors seek to establish a mechanistic link between mitochondrial status and telomere biology.

Strengths:

A major strength of the study is the use of cybrid technology, which provides a great approach to investigate the role of mitochondrial DNA independently of the nuclear genome. The authors also employ multiple complementary assays to assess telomere-related phenotypes associated with mitochondrial dysfunction. Together, these experiments generate an interesting dataset that will be of value to researchers interested in the intersection between mitochondrial biology, genome stability, aging, and development. These results also build on previous work supporting roles for ROS/mitochondria in driving telomere shortening.

We thank the reviewer for this very positive evaluation of our work.

Weaknesses:

The data support the conclusion that mitochondrial background is associated with differences in telomere length and telomere-related phenotypes. However, some of the mechanistic interpretations would benefit from additional evidence. In particular, the manuscript discusses mitochondrial influences on telomere shortening, yet telomere length in some experiments is assessed at a single time point. Consequently, the current data do not directly address the rate of telomere attrition. Differences observed between cybrid lines could potentially arise from events occurring during cybrid formation, clonal selection, or subsequent cell expansion. Longitudinal analyses across multiple passages, ideally beginning immediately after cybrid generation and controlling for population doublings, would help establish whether mitochondrial function directly affects telomere shortening dynamics. Some experimental results would also benefit from additional quantification, clarification, and some biological replicates are missing.

We limited our TL measurements to the earliest viable time point after cybrid formation to avoid the confounding effects of cellular adaptation in culture. For instance, the early telomere shortening observed in Cyb1 and Cyb2 cells was later alleviated—likely due to the upregulation of the NAD+ salvage pathway genes NAMPT and NAPRT1. To accurately capture the effects specific to cybrid formation, we isolated 4 independent clones per donor, all of which showed highly consistent TL values, as shown in Figure 2A and S3D.

While we acknowledge the reviewer's point about multi-passage longitudinal analyses, we did, in fact, measure telomere length across varying population doublings (PDs) in Cyb3 (high mito ROS levels), Cyb6 and Cyb7 (moderate to low mito ROS levels) and observed the following shortening after 66-67 PDs:

- Cybrid 3: about 1.6 kb reduction

- Cybrid 6: about 1.2 kb reduction

- Cybrid 7: about 0.8 kb reduction

These results suggest that the rate of telomere shortening in culture may be higher in Cyb3 cells, possibly due to increased ROS levels.

We included these results as Supplementary Figure 7 and have discussed these data in the revised manuscript:

“Consistently, tracking TL over sequential population doublings revealed that Cyb3 cells are more vulnerable to progressive telomere shortening than Cyb6 or Cyb7 cells (Fig S7). Together, these findings support a role for mitochondrial ROS in human TL regulation.”

We have also carefully edited the manuscript so as to clarify the text and, whenever possible, have added quantifications. Among others, as suggested by Reviewer #1, we added the quantification of chromosome end fusions in the parental 143B Rho0, Cyb1 and Cyb6 cells in our revised manuscript:

Author response image 1.

Overall, this study provides interesting evidence linking mitochondrial background to telomere biology. The cybrid models represent a useful resource for the field, and the work raises important questions regarding mitochondria-telomere communication.

Reviewer #3 (Public review):

Strengths:

Mahieu and colleagues address an interesting and underexplored question: whether non-pathogenic variation in the mitochondrial genome contributes to the inter-individual variability of human telomere length (TL). Using a Belgian Flow-FISH reference cohort (n=491) to identify donors at TL extremes, they generate trans mitochondrial cybrids from platelets of seven donors of distinct mtDNA subhaplogroups and characterize the resulting cells with a broad and well-executed toolkit (TRF, TeSLA, ddTRAP, EPR-based mitoROS, Seahorse with permeabilized-cell ETC dissection, LC-MS metabolomics, telomeric PARFISH). The most compelling finding is that cybrids derived from donors with low complex I (CI) activity undergo rapid telomere shortening during the glycolysis-to-OXPHOS transition of cybrid formation, and that this is largely prevented by co-treatment with NAC and the NAD⁺ precursor nicotinamide riboside, supporting a model in which CI sustains the NAD⁺ pool required for PARP1-mediated repair of oxidative damage at telomeres. The authors further report an inverse correlation between donor lymphocyte TL and mitoROS in the corresponding cybrids, and provide preliminary evidence that the K1a-defining ATP6 A177T variant (m.G9055>A) may be enriched in long-telomere individuals.

We thank the reviewer for this very positive evaluation of our work.

Weaknesses:

(1) Statistical support and donor sampling for the central in vivo correlation (Figure 4C).

The inverse correlation between donor lymphocyte TL and cybrid mitoROS (R²=0.794, p=0.007) is the principal in vivo claim of the paper, but it is built on seven donors deliberately selected from the extremes of the Flow-FISH distribution. Sampling at the tails of the outcome variable can substantially inflate apparent correlation strength and significance. I would encourage the authors to (i) explicitly state this sampling structure where the correlation is introduced, (ii) report a leave-one-out sensitivity analysis to confirm the relationship is not driven by one or two donors (Cyb3 and Cyb6 appear to anchor the line), and (iii) where feasible, extend the analysis to additional donors with intermediate TL to test whether the relationship holds across the full distribution. Even a modest expansion (e.g., 4 to 5 additional donors at P25 to P75) would substantially strengthen this central claim.

We thank the reviewer for this insightful comment. We have conducted the requested leave-one-out analysis:

- removing Cyb3: R²=0.730; p=0.0302

- removing Cyb6 : R²=0.782 ; p=0.0194

- removing Cyb1 : R²=0.740; p=0.0278

While we agree that additional donors would enhance the study, further experiments are however currently impossible without new ethical clearances and additional clinical partnerships.

(2) Reconciling the cybrid CI / TL relationship (Fig 3B) with the absence of a CI / TL relationship in donor lymphocytes (Figure 4A).

Figure 3B shows a strong correlation between CI activity and TL in cybrids (R²=0.87), while Figure 4A shows no correlation between donor CI activity (measured in the same cybrids) and donor lymphocyte TL. The authors acknowledge this, but the manuscript subsequently builds toward a CI-centric model of in vivo TL regulation, which seems to outrun the data. The most internally consistent interpretation is that the cybrid CI phenotype reports a sensitized in vitro response to the acute oxidative stress of the metabolic shift, rather than a steady-state determinant of leukocyte TL. I would suggest reframing the abstract, significance statement, and Discussion to make this distinction clearer. The in vitro CI / NAD⁺ / PARP1 axis is a strong finding on its own, while the in vivo role of CI activity (as opposed to ROS more broadly) is not yet established here. Donor #1's profile (very long lymphocyte TL, low CI activity, severe shortening in cybrids, no telomere inheritance in offspring) is informative in this regard and could be discussed more directly as a case that helps delineate where the cybrid model does and does not recapitulate in vivo biology.

We acknowledge that our study does not establish the in vivo role of CI activity in TL regulation. Our abstract specifically highlights an in vitro phenomenon: “Under the specific conditions of cybrid formation, which involve a metabolic shift from glycolysis to oxidative phosphorylation, mtDNA variants associated with reduced CI activity induced rapid telomere shortening, …”.

We nevertheless extensively revised the Abstract/Significance Statement to clearly separate our in vitro results from in vivo biology as requested:

“Abstract

Telomere shortening is a hallmark of aging, yet telomere length (TL) varies considerably among individuals and is strongly influenced by inheritance. In mice, efficient mitochondrial function, characterised by low reactive oxygen species (ROS) production, is critical for telomere elongation during early embryogenesis. Since mitochondrial DNA (mtDNA) encodes several subunits of the electron transport chain, it may influence TL at birth by regulating mitochondrial function in utero. To explore the relationship between mtDNA and TL in human, we used a transmitochondrial cybrid approach, introducing mitochondria from donor platelets with varying telomere lengths into mtDNA-depleted cells. This revealed an inverse correlation between donor blood TL and mitochondrial ROS levels measured in the resulting cybrids, suggesting that specific mtDNA variants may contribute to the maintenance of long telomeres in humans by enhancing mitochondrial fitness. During in vitro cybrid formation, a transient phase of oxidative stress precedes cellular adaptation. In this specific window, mtDNA variants associated with reduced complex I (CI) activity induced rapid telomere shortening—an effect rescued by antioxidant and NAD⁺ precursor supplementation. While these variants occur naturally in certain individuals with long telomeres, our data suggest that, at least under in vitro conditions of acute oxidative stress, CI is critical to support PARP1 activity by maintaining the NAD⁺/NADH balance, thereby preserving telomere integrity. Collectively, these findings solidify the link between mtDNA variants and human TL regulation, highlighting potential therapeutic opportunities for mitochondrial replacement strategies.”

“Significance Statement

“Telomere length at birth influences aging trajectories and disease risk later in life, yet the mechanisms governing this trait remain incompletely understood. Using a trans mitochondrial cybrid approach, we show that single-nucleotide variants in the mitochondrial genome of healthy donors directly affect mitochondrial metabolism and reactive oxygen species production. In addition, mitochondrial ROS levels measured in cybrids inversely correlate with blood cell telomere length in donors. During cybrid formation, mitochondrial DNA variants associated with reduced CI activity promote telomere shortening. Attrition was reversed by antioxidant and NAD⁺ precursor supplementation, pointing to an essential role for robust CI function in sustaining telomere length during acute oxidative stress, at least under in vitro conditions. Together, these findings establish a direct link between mitochondrial genetics, redox homeostasis, and telomere maintenance in human cells.”

Overall, the manuscript has also been carefully edited to account for this comment.

(3) The K1a / ATP6 A177T inheritance claim.

The proposal that K1a (and specifically ATP6 A177T) contributes to maternal inheritance of long telomeres is intriguing but currently rests on three pedigrees (one of which, donor #1, does not support the hypothesis) and a chi-square test that does not reach significance (p=0.153, Figure 4F). The supporting evidence is also limited by the fact that platelet-mediated mitochondrial transfer delivers donor mitochondrial proteins, lipids, and residual mtRNA in addition to mtDNA, making it difficult to attribute the cybrid phenotype of donor #6 specifically to the ATP6 A177T variant. I would recommend either: (a) extending the genotyping screen to additional unrelated donors and, if feasible, confirming the effect of ATP6 A177T through an isogenic approach (e.g., mtDNA base editing in a clean background), or (b) softening the relevant statements to "suggestive trend warranting larger studies," and presenting the K1a observation as hypothesis-generating rather than supportive. The Ashkenazi-centenarian connection raised in the Discussion is an excellent direction for follow-up and could be framed accordingly.

We agree that the evidence for the AT6 A177T inheritance claim remains inconclusive. To clarify, we do not argue that this mitochondrial variant is solely responsible for longer telomeres; indeed, the mtDNA genome of donor #1 suggests otherwise. Furthermore, the phenotypic impact of such mtDNA variants likely depends on nuclear variants in other telomere-related genes (e.g., hTERT or hTR), meaning AT6 A177T may not consistently result in elongated telomeres. Unfortunately, our ethical protocol precludes screening additional unrelated donors. We have revised the text to soften our statements accordingly.

New references:

DeBoy EA et al. Familial clonal hematopoiesis in a long telomere syndrome. 2023. N Engl J Med 388, 2422-2433.

Davidson-Swinton HR et al. Lymphoid malignancy and clonality in the POT1-mediated long telomere syndrome. 2026. Blood 147, 2226-2237.

Maranzana E et al. Mitochondrial respiratory supercomplex association limits production of reactive oxygen species from complex I. 2013. Antioxid Redox Signal 19, 1469-1480.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Figure 3C. Please clarify. Do the first two bars represent NAMPT, and bars 3-4 represent NAPRT1?

We thank the reviewer for pointing this out, and we have fixed the problem.

Reviewer #2 (Recommendations for the authors):

The authors discuss telomere shortening throughout the manuscript, but telomere length is measured only at a single time point. It would be more informative to pass the cells over several population doublings and determine the rate of telomere shortening. Longitudinal analysis across multiple passages, ideally starting immediately after cybrid generation and controlling for population doublings, would be required to determine whether mitochondria directly influence telomere attrition rates. Overall, I like the manuscript- I think it is interesting and impactful, but some controls are missing.

We thank the reviewer for this positive evaluation of our work and have inserted the TRF results of our longitudinal analysis of telomere length in Cybrids #3 (high ROS), #6 (low ROS) and #7 (very low ROS) cultured until PD 66-67 after cloning. Data are shown as Fig S7.

We observed the following shortening after 66-67 PDs:

- Cyb3: about 1.6 kb reduction

- Cyb6: about 1.2 kb reduction

- Cyb7: about 0.8 kb reduction

These results suggest that the rate of telomere shortening in culture may be higher in Cyb3 cells, possibly due to increased ROS levels.

(1) Why would rho (0) have shorter telomeres than parental cells? Or decreased telomerase activity?

We will look at this soon. Our current hypothesis, based on the recent work by the Agarwal’s lab, is that the short telomeres in 143B Rho0 cells happen because these cells lack the nuclear-encoded TK1 enzyme and cannot make pyrimidines through the salvage pathway or de novo (since dihydroorotate dehydrogenase needs working mitochondria). When mitochondria return, thymidine levels may rise and help telomerase work again, potentially through an allosteric mechanism as previosuly suggested (Mannherz and Agarwal, Nat Genet 2023).

This was already discussed in our manuscript (Discussion):

“Our study also unexpectedly revealed a reduction in telomere length in 143B cells following mitochondrial depletion. This telomere shortening was associated with decreased telomerase activity, as measured by ddTRAP, but not with reduced expression of the telomerase subunits hTERT or hTR. 143B osteosarcoma cells are deficient in thymidine kinase 1 and therefore lack an effective pyrimidine salvage pathway. In the absence of oxidative phosphorylation, the de novo thymidine synthesis pathway is also compromised, rendering 143B r⁰ cells dependent on uridine supplementation for survival. Although this hypothesis requires further investigation, we propose that the shorter telomeres observed in 143B r⁰ cells may be linked to the emerging role of thymidine nucleotide metabolism in the regulation of human telomere length, potentially through an allosteric effect of dTTP on telomerase activity (4143).”

(2) Why does 2C data show no variation? How many biological replicates?

We showed hTR and hTERT expression levels in one biological replicate of two independent cybrid clones from each donor, with technical replicates for each measurement. Measurements of hTERT and hTR expression levels over successive PDs were performed for some clones and showed good consistency.

(3) Why did the TIF analysis not compare parental and Rho (0) cells?

We have now added the data in Fig 2D.

Figure 2E is only a representative image; no quantification is provided, so no conclusions about fusion can be made.

We have now quantified chromosome end fusions in the parental 143B Rho0, Cyb1 and Cyb6 cells and have included the quantifications in our revised manuscript (Fig 2E):

“E. Left: Visualisation of telomere fusions (white arrowheads) in Cyb1 cells at PD7. 143B r⁰ and Cyb6 cells are shown as controls. Telomeres are detected by FISH (red) and DAPI stains DNA (blue). Scale bar: 5 µm. Right: Quantification of telomere fusions. The total number of analyzed chromosomes is shown for each cell line. Χ² test. ns: p>0.05.”

(4) In Figure 3, I found it difficult to interpret the data shown in D (what is each lane?).

Each lane shows TL of an independent hybrid clone. We have clarified the legend accordingly.

Why is the data in E shown this way? How many biological replicates, and why no experimental validation?

Panel E quantifies the TRF analyses conducted on independent cybrid clones. These clones were isolated from new cybridizations of donor 2’s platelets with 143B Rho0 cells under the specified conditions. The specific number of independent cybrid clones (which we consider as biological replicates) for each condition is shown in the graph: NAA: 5, NAC: 7, NAC + NR: 8, NAA + NR: 8. We have clarified the legend as follows. We hope this is clearer now.

“E. Quantification of the TRF analyses conducted on independent cybrid clones isolated in the conditions detailed in D. The percentage of clones with average TL>2.5 kb was calculated. The specific number of independent cybrid clones analyzed for each condition is shown in the graph: NAA: 5, NAC: 7, NAC + NR: 8, NAA + NR: 8.”

(5) Did NAC and NR also reduce TIF or telomere fusions?

We did not test this. We expect NAC/NR treatment to reduce TIF and telomere fusions indeed.

For the Parylation experiments, how many independent experiments were performed?

Eight independent cybrid clones obtained from 4 independent cybridisations were analyzed. This has now been detailed in the legend as follows :

“F. Representative images and quantification of telomeric parylation in 8 independent Cyb2 clones obtained from 4 independent cybridisations, comparing cells that did or did not exhibit telomere shortening during the initial PDs following cybridisation (from panel D-E).”

(6) Figure 4. Why were mitochondrial superoxide levels not measured in parental cells?

We excluded Rho0 cells from our superoxide measurements because these cells were grown in a different culture medium. Instead, we focused on comparing mitochondrial ROS across cybrids to accurately correlate these values with the telomere length of the corresponding donors’ lymphocytes. Consequently, Rho0 cell measurements would not have contributed to this correlation analysis.

We have discussed this in the revised manuscript:

“Mitochondrial superoxide production was measured by electron paramagnetic resonance (24) in clones #1 of all cybrids and in 143B cells (Fig 4B). We excluded 143B r⁰ cells from our superoxide measurements because these cells were grown in a different culture medium.”

Reviewer #3 (Recommendations for the authors):

Additional analyses for Figure 4C. Please report a leave-one-out sensitivity analysis for the lymphocyte TL versus cybrid mitoROS correlation,

We have conducted the requested leave-one-out analysis:

- removing Cyb3: R²=0.730; p=0.0302

- removing Cyb6: R²=0.782; p=0.0194

- removing Cyb1: R²=0.740; p=0.0278

We chose not to include this analysis in the revised manuscript.

and state in the legend that donors were selected at the extremes of the Flow-FISH distribution.

This is not exactly the case, but we nevertheless added this sentence in the legend:

“TL in the donor lymphocytes was distributed as follows: >P90 (#1, #6 and #7), around P50 (#2 and #4), between P10 and P50 (#5) and at P10 (#3).”

If feasible, extend the analysis to 4 or 5 additional donors with intermediate TL (P25 to P75) to test whether the relationship holds across the full distribution.

While we agree that additional donors would enhance the study, further experiments are however currently impossible without new ethical clearances and additional clinical partnerships.

Reframing of the in vivo claims. Please calibrate the language in the abstract, significance statement, and Discussion so that the in vitro complex I / NAD⁺ / PARP1 mechanism and the in vivo role of the mitochondrial genome are presented at the appropriate levels of evidence. A brief paragraph in the Discussion explicitly addressing donor #1 (long lymphocyte telomeres, low complex I activity, telomere shortening in cybrids, no inheritance of long telomeres in offspring) would be helpful in delineating where the cybrid model does and does not recapitulate in vivo biology.

The text has been thoroughly revised in accordance with this suggestion. We kindly invite you to review the updated manuscript for full details.

Strengthening the K1a / ATP6 A177T claim. Two options would address the concern raised in the public review. First, extending the genotyping screen to additional unrelated donors until the chi-square test reaches the level of evidence consistent with the wording in the abstract.

Unfortunately, further experiments are currently impossible without new ethical clearances.

Second, if a collaboration is available, an isogenic introduction of the m. G9055>A variant in a clean mitochondrial background through mitochondrial base editing or mitoTALENs would directly test causality.

This would certainly be an excellent approach to test causality indeed and we fully agree that future collaborations with specialized labs would be highly relevant in this context.

If neither is feasible, please soften the statements to a suggestive trend warranting larger studies.

We chose to soften our claims instead; please see the revised manuscript for details.

Direct test of PARP1 involvement. A PARP1 knockdown or a PARP inhibitor co-treatment in Cyb2 during fusion would directly test whether the NAC + nicotinamide riboside rescue depends on PARP1 activity, as the model proposes. If this experiment is not feasible, please state so in the Discussion.

Again, redoing cybridisations is currently impossible without new ethical clearances. We nevertheless acknowledge that the suggested experiment is interesting and we discussed it in the revised manuscript as follows:

“Mitochondrial platelets with low CI activity triggered significant telomere shortening in recipient cells during early cybrid formation -a loss that was prevented by co-treatment with an antioxidant and an NAD⁺ precursor. Although we did not determine whether the rescue was dependent on PARP1, these results support the essential role of CI in maintaining PARP1 activity via regulation of the NAD⁺/NADH balance.”

Figure 1C/D legend says "eight selected donors" twice (lines 635, 636), but the text says "seven donors" (lines 132, 258), and donors are numbered #1-#7 throughout. This is a real internal inconsistency in the manuscript that the authors should fix.

We have fixed this mistake. Thanks for bringing this to our attention.

eLife Assessment

This study presents an important investigation of how inhibition of the WNK-SPAK/OSR1 pathway influences neuronal chloride homeostasis and seizure-like activity in organotypic hippocampal slice cultures. The evidence supporting the principal mechanistic conclusions is solid, although the proposed regulation of chloride transporters is inferred rather than directly demonstrated in this model. The work will be of interest to researchers studying chloride homeostasis, inhibitory neurotransmission, and epilepsy.

Reviewer #1 (Public review):

Summary:

Epileptogenesis involves deficits in GABAergic inhibition mediated by chloride-permeable GABA(A) channels, which are caused by elevated intracellular chloride. These changes in chloride homeostasis result from increased activity of the NKCC1 chloride importer and loss of function of the KCC2 chloride extruder. The authors used organotypic hippocampal slice preparations from mice expressing the genetically encoded chloride indicator Clomeleon as a model of posttraumatic epileptogenesis to show that these changes are mediated by the WNK-SPAK/OSR1 pathway. Blockade of the WNK-SPAK/OSR1 pathway with WNK463 alleviated epileptic activity. These findings provide important insights into the mechanisms of epileptogenesis and open new avenues for treating epilepsy.

Strengths:

The conclusions are well supported by the detailed pharmacological analysis.

Comments on revised version.

The authors addressed all the recommendations in their revised manuscript.

Reviewer #2 (Public review):

Summary:

The authors investigate whether inhibition of the WNK-SPAK/OSR1 pathway using the allosteric inhibitor WNK463 improves neuronal chloride homeostasis and suppresses epileptiform activity in organotypic hippocampal slice cultures. Using SuperClomeleon imaging combined with extracellular field recordings, they demonstrate that WNK463 accelerates recovery of intracellular chloride following chloride loading, reduces interictal chloride accumulation, and progressively suppresses recurrent ictal-like discharges. Pharmacological inhibition and siRNA-mediated knockdown of NKCC1 and KCC2 are then used to investigate the contribution of these transporters to the anti-ictal effects of WNK463.

Strengths:

The study addresses an important question in the field of chloride homeostasis and epilepsy and combines complementary experimental approaches. In particular, the distinction between baseline chloride measured in the presence of TTX and activity-dependent interictal chloride accumulation provides a useful conceptual framework for interpreting previous studies of WNK-SPAK inhibition. The imaging, electrophysiological, and pharmacological data are internally consistent and support the conclusion that WNK463 alters chloride dynamics and substantially suppresses ictal-like activity in this model.

Weaknesses:

The revised manuscript addresses several concerns raised in my initial review. The authors have broadened the discussion of KCC2 dysfunction and added further context on therapeutic targeting, as well as a useful discussion of the limitations of the organotypic hippocampal slice model.

My main remaining concern is the mechanistic interpretation of the data. The authors have appropriately softened several statements and now acknowledge that membrane potential and the GABA reversal potential were not directly measured. However, strong mechanistic conclusions remain in the revised manuscript, including statements that the anti-ictal effects of WNK463 are "primarily mediated" by reduced NKCC1 and enhanced KCC2 transport activity and are "mainly due to" NKCC1 inhibition and KCC2 activation. These conclusions go beyond what is directly demonstrated, as WNK463-induced regulation of NKCC1 and KCC2 was not directly measured in this preparation. Lee et al. (2022) provides important support for this mechanism, demonstrating reduced KCC2-T1007 phosphorylation and enhanced KCC2-dependent chloride extrusion following WNK463 treatment. However, these findings were obtained using different experimental preparations, including cultured cortical neurons and an acute 4-AP seizure model. The current experiments therefore support a contribution of NKCC1 and KCC2, but do not directly establish coordinated regulation of these transporters as the primary mechanism underlying seizure suppression. This is also relevant to the simultaneous NKCC1/KCC2 inhibition experiments, where the overall anti-ictal effect of WNK463 is reduced but not completely abolished.

The statistical reporting has also been improved, but the experimental unit remains unclear for analyses involving multiple cells recorded from individual slices. The authors clarify the use of repeated-measures analyses, but it remains unclear whether individual cells or slices were treated as independent observations and how multiple cells from the same slice were accounted for statistically.

Author response:

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

eLife Assessment:

This study presents a valuable investigation of how inhibition of the WNK-SPAK/OSR1 pathway influences neuronal chloride homeostasis and seizure-like activity in organotypic hippocampal slice cultures. The evidence supporting the principal mechanistic conclusions is incomplete because several proposed mechanisms are inferred from changes in chloride dynamics rather than directly demonstrated. The work will be of interest to researchers studying chloride homeostasis, inhibitory neurotransmission, and epilepsy.

We appreciate the thoughtful reviews and assessment. Additional data analysis has been performed, and the paper has been carefully modified to address each concern, as described below.

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors used extracellular field potential recordings and two-photon imaging to monitor neuronal network activity and intracellular chloride concentration ([Cl-]i) in organotypic hippocampal slices from mice expressing the genetically encoded chloride fluorophore Clomeleon. These slices were used as a model of acute traumatic brain injury and epileptogenesis in vitro. The study provides evidence that blocking the WNK-SPAK/OSR1 pathway with WNK463 alleviates epileptic activity, and that this anticonvulsant effect involves suppression of the chloride loader NKCC1 and enhancement of chloride extrusion via KCC2. Overall, this is a solid study with a comprehensive pharmacological analysis.

Strengths:

The conclusions are well supported by the detailed pharmacological analysis.

Weaknesses:

The only weakness I see is the absence of cellular-level electrophysiology, which precludes interpretation of the imaging data in the context of GABA action polarity.

The referee correctly points out that we did not perform cellular-level electrophysiological measurements. Before the advent of transgenic ratiometric chloride-sensitive fluorophores, we conducted detailed studies of neuronal cation-chloride transport using electrophysiological measures (Staley and Proctor, 1999). We did not measure membrane potential in the current study because this data is not needed for characterization of cation-chloride membrane cotransport, which is electroneutral. Electroneutral processes are not affected by membrane potential, and electroneutral processes in turn do not alter the membrane potential. We did not measure membrane GABA currents because the polarity of GABA currents was clearly demonstrated by our fluorometric measures of the corresponding increases or decreases in cytoplasmic chloride concentration (Figure 2). We have added the following sentence to the Discussion (Page 16): “In this fluorometric study, we did not measure RMP and E_GABA directly, but rather inferred the impact of electroneutral cation-chloride transport on these parameters from the effects on [Cl^-]_i (Figure 2) and seizure activity (Figures 1, 3-9)”.

Reviewer #2 (Public review):

Summary:

The authors investigate whether inhibition of the WNK-SPAK/OSR1 pathway using the allosteric inhibitor WNK463 improves neuronal chloride homeostasis and suppresses epileptiform activity in organotypic hippocampal slice cultures. Using Super Clomeleon imaging combined with extracellular field recordings, they demonstrate that WNK463 accelerates recovery of intracellular chloride following chloride loading, reduces interictal chloride accumulation, and progressively suppresses recurrent ictal-like discharges. Pharmacological inhibition and siRNA-mediated knockdown of NKCC1 and KCC2 are then used to investigate the contribution of these transporters to the anti-ictal effects of WNK463.

Strengths:

The study addresses an important question in the field of chloride homeostasis and epilepsy and combines complementary experimental approaches. In particular, the distinction between baseline chloride measured in the presence of TTX and activity-dependent interictal chloride accumulation provides a useful conceptual framework for interpreting previous studies of WNK-SPAK inhibition. The imaging, electrophysiological, and pharmacological data are internally consistent and support the conclusion that WNK463 alters chloride dynamics and substantially suppresses ictal-like activity in this model.

Weaknesses:

The principal limitation of the manuscript is that several mechanistic conclusions extend beyond the experimental observations. Throughout the results and discussion, the authors interpret the observed changes in intracellular chloride dynamics as evidence of enhanced CCC-mediated chloride extrusion, while the pharmacological and siRNA-mediated experiments are interpreted as supporting coordinated NKCC1 inhibition and KCC2 activation, ultimately leading to restoration of GABAergic inhibition and negative shifts in EGABA. While these interpretations are plausible and consistent with the data, they remain inferential because transporter phosphorylation or activity, EGABA, and inhibitory synaptic function were not directly assessed in the current study. Moreover, although the pharmacological and knockdown experiments support a contribution of NKCC1 and KCC2 to the actions of WNK463, they do not definitively establish coordinated modulation of both transporters as the primary mechanism underlying seizure suppression. These mechanistic conclusions should therefore be presented more cautiously.

The referee correctly points that we did not definitively establish the link between coordinated modulation of NKCC1 and KCC2 transporter activity, changes in EGABA and GABA function to the anti-ictal action of WNK463. Indeed we have an entire section of results dedicated to the anti-ictal effects of WNK463 that persist after block of NKCC1 and KCC2. The primarily biochemical study by Lee et al (2022) demonstrated that the WNK-SPAK/OSR1 inhibitor WNK463 reduced KCC2-T1007 phosphorylation and EGABA in cultured cortical neurons challenged with high 30 mM [Cl^-]_i in vitro, and limited KA-induced status epilepticus in vivo. We are studying the same system with different pharmacological, electrophysiological, optogenetic and morphological tools during spontaneous ictal-like epileptiform discharges in organotypic hippocampal slices in vitro. We prominently cite the previous study and clearly report new findings and demonstrate underlying mechanisms. We have also modified the language surrounding our conclusions (e.g. last sentence of abstract, and last sentence of first paragraph of Discussion).

The manuscript would also benefit from broader contextualization within the current literature. The introduction largely focuses on previous work from the authors' group and provides a relatively narrow overview of chloride homeostasis in epilepsy. In particular, the discussion would benefit from broader consideration of studies examining KCC2 dysfunction in human epilepsy and experimental models, alternative mechanisms regulating KCC2 activity following seizures, and recent therapeutic strategies targeting KCC2.

We agree. We have added the following paragraph to the Discussion (p. 16-17): “The KCC2 protein is a potassium-chloride co-transporter that maintains low [Cl^-]_i (Rivera et al., 1999). KCC2 dysfunction has been implicated in epilepsy across human and experimental studies. Reduced KCC2 expression and impaired chloride extrusion have been demonstrated in resected tissue from patients with temporal lobe epilepsy, and pathogenic variants in SLC12A5 further support a causal role for impaired KCC2 function in human epilepsy (Huberfeld et al., 2007; Puskarjov et al., 2014). Experimental seizures can suppress KCC2 through several mechanisms in addition to WNK-SPAK/OSR1 signaling, including BDNF-TrkB-dependent downregulation, dephosphorylation of the stabilizing Ser940 site with subsequent internalization, and calpain-mediated proteolysis (Rivera et al., 2004; Silayeva et al., 2015; Puskarjov et al., 2012). These convergent mechanisms suggest that loss of KCC2 function may both result from and amplify seizure activity by weakening GABAergic inhibition. Accordingly, KCC2 has emerged as a therapeutic target: strategies include modulation of regulatory phosphorylation, enhancement of KCC2 expression, and direct pharmacological potentiation. Notably, the KCC2 activator OV350 enhances chloride extrusion and suppresses experimental seizures, including benzodiazepine-refractory status epilepticus, providing proof-of-concept for direct KCC2 activation as an antiseizure strategy (Jarvis et al., 2023).”

Finally, although the authors appropriately acknowledge that the experiments were performed exclusively in vitro, the discussion could more explicitly address the limitations of the organotypic hippocampal slice model, including how culture-induced network reorganization and spontaneous epileptiform activity may influence chloride homeostasis and the extent to which these findings generalize to traumatic brain injury and chronic epilepsy in vivo. In addition, the statistical analysis would benefit from clarification regarding the experimental unit and the treatment of repeated measurements.

Thank you. We addressed the limitation of organotypic hippocampal slices in vitro in the Limitation and Future Directions paragraph (p. 18-19): “Our experiments were conducted in organotypic hippocampal slices in vitro. Advantages of this preparation include homogenous injury and consistent epileptogenesis. This provides excellent control of the extracellular environment and imaging but does not replicate the in vivo vascular supply and maintenance of ionic homeostasis. While we find that increased WNK-SPAK/OSR1-mediated KCC2 activity via enhanced Cl^- extrusion rates robustly reduces seizure activity in vitro, it is possible that in vivo, increased KCC2 activity could lead to extracellular potassium accumulations (Viitanen et al., 2010) that would limit Cl^- extrusion independently of the velocity of KCC2 transport (Thompson et al., 1988; Staley and Proctor, 1999). Thus, it will be important to assess these findings in vivo.”

We provided better clarification regarding the treatment of repeated measurements (p. 22): “Experimental data were analyzed using a repeated measures ANOVA for two types of study design. Studies that investigated either (1) changes in mean (median) parameters over three or more time points or consecutive time intervals; or (2) differences in mean (median) parameters under three or more different conditions”.

Reviewer #3 (Public review):

Summary:

Dzhala and colleagues present findings from organotypic slice cultures suggesting that simultaneous modulation of the complementary cation-chloride cotransporters NKCC1 and KCC2 through inhibition of the WNK-SPAK/OSR1 pathway reduces seizure-like activity. The manuscript is generally well written, and the data support the conclusion that WNK463 exerts robust anti-ictal effects in this model. However, several issues should be addressed to strengthen the mechanistic interpretation and statistical rigor of the study, and improve confidence in the conclusions.

Strengths:

(1) The study addresses an important mechanistic question by investigating how inhibition of the WNK-SPAK/OSR1 pathway with WNK463 influences seizure activity and neuronal chloride homeostasis.

(2) The experimental design is logical and comprehensive, progressing from characterization of chloride dynamics to pharmacological and genetic interrogation of the underlying mechanism using multiple complementary approaches, including pharmacological inhibition, siRNA-mediated knockdown, electrophysiology, and chloride imaging.

(3) The combination of simultaneous extracellular electrophysiology and two-photon chloride imaging provides complementary functional and mechanistic information and represents a major technical strength of the study.

(4) The TTX experiments elegantly distinguish activity-dependent chloride accumulation from resting intracellular chloride concentration, substantially strengthening the central mechanistic conclusions.

Weaknesses:

(1) The mechanistic conclusions regarding KCC2 activation and NKCC1 inhibition are stronger than the data directly support. Throughout the manuscript, the authors conclude that WNK463 activates KCC2 and inhibits NKCC1. Although this interpretation is consistent with the established biology of the WNK-SPAK/OSR1 pathway, the evidence presented here is indirect. Specifically, the authors infer KCC2 activation and NKCC1 inhibition from the observation that pharmacological inhibition or siRNA-mediated knockdown of these transporters alters the effects of WNK463, together with measurements of chloride dynamics. While these findings are compatible with a KCC2- and NKCC1-dependent mechanism, they do not directly establish that WNK463 regulates either transporter. Direct evidence would require measurements of transporter activity, phosphorylation state, membrane expression, or other biochemical indices of transporter regulation. I therefore recommend that the authors both temper the mechanistic language throughout the manuscript and explicitly acknowledge in the Discussion that the proposed regulation of KCC2 and NKCC1 is inferred from indirect evidence rather than directly demonstrated in the present study.

We agree. We modified the key conclusory sentences in the Abstract and Discussion as described in the response to referee 2. The referee correctly points that we did not directly establish the link between the action of WNK463 and coordinated activation of KCC2 and inhibition NKCC1. We highlighted the established link in the discussion (p. 17-18): “NKCC1 and KCC2 are rapidly stimulated and inhibited, respectively, by direct phosphorylation mediated by the Cl^--sensitive WNK-SPAK/OSR1 complex (Alessi et al., 2014). The WNK family senses changes in the [Cl^-]_i, cell volume, extracellular osmolarity and transduces this information to CCCs (Shekarabi et al., 2017). WNKs stimulate the kinases SPAK and OSR1, which directly phosphorylate and stimulate NKCC1 or inhibit KCC2 (Kahle et al., 2013), providing compensatory and reciprocal regulation between transporters activity to maintain chloride homeostasis and neuronal excitability (Hartmann and Nothwang, 2022). Conversely, inhibiting the WNK-SPAK/OSR1 pathway might be an especially potent strategy to reduce neuronal Cl^- elevation and facilitate neuronal Cl^- extrusion by coincident NKCC1 inhibition and KCC2 activation…” Consistently with the established biology of the WNK-SPAK/OSR1 pathway we found that (p.15): “…WNK463 improved [Cl^-]_i homeostasis, progressively enhanced postictal [Cl^-]_i extrusion rates, reduced ictal duration and abolished recurrent ILDs. Our pharmacological investigations suggest that the powerful anti-ictal effects of WNK463 are mediated by reduced NKCC1 and enhanced KCC2 transport activity...”

(2) The conclusions drawn from the siRNA-mediated knockdown experiments should be interpreted more cautiously. First, it is unclear whether silencing NKCC1 or KCC2 induced compensatory changes in the expression or function of the complementary cotransporter. Given the well-established interplay between NKCC1 and KCC2 in regulating intracellular chloride homeostasis, compensatory adaptations could influence the interpretation of these experiments and should be addressed or acknowledged as a limitation. Second, the sample size for the siRNA experiments appears relatively small. It is unclear how many independent animals contributed slices to each experimental group, making it difficult to assess the degree of biological replication. In addition, effect sizes are not reported. Clarifying the number of biological replicates and reporting effect sizes would improve the rigor of the statistical analysis and increase confidence in these findings.

We agree and now address this in the discussion (p. 17-18): “WNKs stimulate the kinases SPAK and OSR1, which directly phosphorylate and stimulate NKCC1 or inhibit KCC2 (Kahle et al., 2013), providing compensatory and reciprocal regulation between transporters activity to maintain chloride homeostasis and neuronal excitability (Hartmann and Nothwang, 2022). ”

We agree that the sample size for the siRNA electrophysiological experiments is relatively small (n=3-4 slices). We now clarify the number of mice were used per each experimental group (p.19). We provide the n (number of slices) and the number of paired cells in the Results and for every experimental conditions.

(3) The final pharmacological experiments require clarification, as the conclusions appear internally inconsistent. Earlier experiments suggest that the anticonvulsant effects of WNK463 depend on coordinated regulation of both NKCC1 and KCC2. However, in the final experiment, the authors state that simultaneous pharmacological inhibition of NKCC1 and KCC2 does not prevent the anticonvulsant effects of WNK463. In contrast, the accompanying statistical analysis indicates that combined transporter inhibition significantly reduces the effect of WNK463 relative to control conditions. These interpretations appear inconsistent and make it difficult to determine the extent to which the anticonvulsant action of WNK463 depends on NKCC1 and KCC2. The authors should clarify whether simultaneous inhibition of both transporters completely abolishes, partially attenuates, or merely reduces the magnitude of the WNK463 response, and revise the text accordingly. If the effect is only partially attenuated, alternative mechanisms contributing to the anticonvulsant actions of WNK463 should also be considered and discussed.

We agree. Our data demonstrated that WNK463 had no effect in the presence of high concentration of bumetanide (500 µM) that nonspecifically blocks CCCs (NKCC1 and KCC2-3) (Figure 6). In contrast (p. 15): ”Simultaneous inhibition of NKCC1 and KCC2 by the combination of bumetanide (10 µM) and VU0463271 (1 µM) significantly reduced the anti-ictal effects of WNK463 (Figure 9).” The conclusory sentence in the Results indicates (p.15): “These results suggest that the anti-ictal effects of WNK-SPAK/OSR1 inhibitor WNK463 are mainly due to downstream interactions with NKCC1 inhibition and KCC2 activation, with additional effects due to interactions with an unidentified Cl^- transporter (e.g. KCC3) or exchanger, or channel.”

(4) The statistical analysis and reporting require further attention. First, median values should not be reported with standard deviations, as standard deviation describes variability around the mean rather than the median. For non-normally distributed data, the authors should report median values together with an appropriate measure of variability, such as the interquartile range (25th-75th percentile) or another suitable summary. Second, in several instances, ANOVA results are reported using only a single degree of freedom value (e.g., page 6, DF = 53). This is incomplete, as an F statistic is defined by two degrees of freedom: the numerator degrees of freedom (between-group variability) and the denominator degrees of freedom (within-group variability). Reporting statistical results using standard notation (F(df_between, df_within) = F statistic, p = value) would improve clarity and allow proper interpretation of the analyses.

Thank you. Group measures are expressed now as mean ± standard deviation (SD) or median (25%-75%) as indicated. In a one-way RM ANOVA, (DF) represents the between-groups degree of freedom, and (F) represents the F-statistic and reported as F(DF_treatment, DF_residual)=value. In the Friedman RM ANOVA on Ranks, the chi-square statistic was used to assess the differences between the groups. The Tukey Test or Holm-Sidak Test was used for all pairwise comparisons of the responses to the different treatment groups. Please see Statistical Analysis in the Methods section (p.23-24).

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

I have only minor comments/suggestions.

(1) While [Cl^-]i and ECl are informative parameters, interpreting the GABA driving force and determining whether depolarizing/excitatory GABA actions during seizures or interictal states are alleviated by treatment requires knowledge of the membrane potential and its dynamics during epileptic events. I would recommend including this information if available. Gramicidin perforated-patch recordings (ideally) would substantially strengthen the conclusions, at least for a critical subset of experiments.

Please see our response above. We have added the following sentence to the discussion (page 16): “In this fluorometric study, we did not measure RMP and EGABA directly, but rather inferred the impact of electroneutral cation-chloride transport on these parameters from the effects on [Cl^-]_i (Figure 2) and seizure activity (Figures 1, 3-9)”.

(2) Did the authors observe any signs of cell shrinkage in response to WNK463, or transient cellular swelling during interictal-like discharges (ILDs) and its prevention by WNK463? Given that cytotoxic edema is typically associated with increased tissue light transparency and shrinkage of the extracellular space (e.g., as measured with TMA⁺), this reviewer wonders whether such approaches could be used as a proxy for cell swelling. Tissue edema, which manifests as increased tissue volume, could also be assessed by measuring slice thickness.

These are excellent ideas and cell volume is an area of interest for us, but we would argue that assessment of volume effects are outside the scope of the current 9-figure study. We observed cell volume (transverse area) increases during prolonged ictal-like discharges. This effect was prevented by WNK463 during suppression of ictal-like discharges. However, we did not provide the volume analysis in this paper.

(3) The paper should be corrected for spelling and grammar (e.g., Figure 2 legend: "(b) Linier regression analysis revealed that... " - probably should be "Linear... "?).

Thank you. Corrected.

Reviewer #2 (Recommendations for the authors):

(1) Temper mechanistic statements throughout the manuscript by clearly distinguishing experimental observations from inferred mechanisms and by explicitly framing conclusions regarding coordinated NKCC1/KCC2 modulation, chloride extrusion, EGABA, and restoration of GABAergic inhibition as interpretations supported by, rather than directly demonstrated by, the current data.

We have softened the key conclusory sentences in the Abstract and Discussion as described above. In the introduction, we highlighted the potential of both NKCC1 and KCC2 targeting for control of seizures (p. 2): “Previous studies have demonstrated that inhibition of NKCC1 reduced [Cl^-]_i in injured neurons, enhanced GABAergic inhibition and the efficacy of anticonvulsants (Pond et al., 2006; Dzhala et al., 2008; Dzhala et al., 2024; Nardou et al., 2011; Cleary et al., 2013; Sivakumaran and Maguire, 2016; Soul et al., 2021). Conversely, enhancement of KCC2 activity improved [Cl^-]_i extrusion and control of seizures (Gagnon et al., 2013; Moore et al., 2018; Dzhala and Staley, 2021).”

In the discussion we highlighted the potential of coordinated NKCC1 and KCC2 modulation (p. 17-18): “NKCC1 and KCC2 are rapidly stimulated and inhibited, respectively, by direct phosphorylation mediated by the Cl^--sensitive WNK-SPAK/OSR1 complex (Alessi et al., 2014). The WNK family senses changes in the [Cl^-]_i, cell volume, extracellular osmolarity and transduces this information to CCCs (Shekarabi et al., 2017). WNKs stimulate the kinases SPAK and OSR1, which directly phosphorylate and stimulate NKCC1 or inhibit KCC2 (Kahle et al., 2013). Conversely, inhibiting the WNK-SPAK/OSR1 pathway might be an especially potent strategy to reduce neuronal Cl^- elevation and facilitate neuronal Cl^- extrusion by coincident NKCC1 inhibition and KCC2 activation…”

(2) Expand the introduction and discussion to provide a more balanced overview of the current literature on chloride homeostasis in epilepsy, including alternative mechanisms regulating KCC2 dysfunction and recent therapeutic strategies targeting KCC2.

In the introduction, we highlighted the potential of both NKCC1 and KCC2 targeting for control of seizures (p. 2): “Previous studies have demonstrated that inhibition of NKCC1 reduced [Cl^-]_i in injured neurons, enhanced GABAergic inhibition and the efficacy of anticonvulsants (Pond et al., 2006; Dzhala et al., 2008; Dzhala et al., 2024; Nardou et al., 2011; Cleary et al., 2013; Sivakumaran and Maguire, 2016; Soul et al., 2021). Conversely, enhancement of KCC2 activity improved [Cl^-]_i extrusion and control of seizures (Gagnon et al., 2013; Moore et al., 2018; Dzhala and Staley, 2021).” We did not provide additional references to comply with the word limit. We have added a paragraph and citations (pp. 16-17) to the Discussion to address this issue.

(3) Discuss more explicitly the limitations of the organotypic hippocampal slice model when extrapolating the findings to traumatic brain injury and chronic epilepsy in vivo.

Please see our response above. We discussed the limitation of the organotypic hippocampal slice model in the Limitations and future directions section (p. 18-19): “Our experiments were conducted in organotypic hippocampal slices in vitro. Advantages of this preparation include homogenous injury and consistent epileptogenesis (Berdichevsky et al., 2012). This provides excellent control of the extracellular environment and imaging but does not replicate the in vivo vascular supply and maintenance of ionic homeostasis. While we find that increased WNK-SPAK/OSR1-mediated KCC2 activity via enhanced Cl^- extrusion rates robustly reduces seizure activity in vitro, it is possible that in vivo, increased KCC2 activity could lead to extracellular potassium accumulations (Viitanen et al., 2010) that would limit Cl^- extrusion independently of the velocity of KCC2 transport (Thompson et al., 1988; Staley and Proctor, 1999). Thus, it will be important to assess these findings in vivo.”

(4) Clarify how repeated measurements from individual cells and slices were handled in the statistical analyses, including the experimental unit for each analysis.

We clarified in the methods (p.22): “Experimental data were analyzed using a repeated measures ANOVA for two types of study design. Studies that investigated either (1) changes in mean (or median) parameters over three or more time points or consecutive time intervals); or (2) differences in mean (or median) parameters under three or more different conditions”.

Reviewer #3 (Recommendations for the authors):

(1) The manuscript would also benefit from including statements regarding randomization, blinding (beyond the immunohistochemistry analysis), and any inclusion or exclusion criteria, as these details are increasingly expected to improve transparency and reproducibility.

We agree. Images were blinded for the immunohistochemistry analysis (p. 21) and quantitative analysis of intracellular chloride changes (p. 23). Cells that did not respond to GABA/L-glutamate puff applications (Figure 2) and chloride transients that failed to fit an exponential decay function (p. 5) were excluded from the analysis. All other data generated or analyzed during this study were included in the manuscript.

(2) The duration of baseline recordings prior to drug application should be specified in the Methods.

Thank you. This information included in the Method section (p.19): “Slices were stabilized for 10-20 min in ACSF prior recordings and/or imaging”.

(3) Some wording could be refined for readability. For example, "insignificantly reduced " on pages 8 and 14 would be more appropriately stated as "did not significantly reduce. "

Thank you. Corrected.

(4) The first sentence of the Introduction refers to "anticonvulsant (AED)-resistant seizures. " I suspect the authors intended "antiepileptic drug (AED)-resistant seizures, " as the current wording does not match the acronym.

Thank you. Corrected.

eLife Assessment

This is an important study establishing the mechanistic principles on how NUP98-KDM5A phase separates together with H3K4me3 chromatin in leukemia. Methods are sound and results are convincing, providing a framework to understand gene expression changes in leukemia patients.

Reviewer #1 (Public review):

[Editor's note: this version has been assessed by the Reviewing Editor without further input from the original reviewers.]

Leukemia-driving NUP98 oncofusion proteins form chromatin-associated biomolecular condensates in the nucleus, and these structures are important for oncogenic transformation. Most NUP98 fusions do not contain domains that mediate the recognition of specific DNA elements. Instead, they entail domains that are important for chromatin regulation. For instance, the NUP98::KDM5A fusion features a fusion of the NUP98 N-terminus with the third PHD domain of the histone demethylase KDM5A. As PHD domains are critical for the recognition of methylated histones without any sequence specificity, it is not clear what controls the condensation and chromatin binding of NUP98::KDM5A, leading to the induction of oncogenic transcriptional programs.

In this work, the authors use a combination of cellular and in vitro studies to show that biomolecular condensation of NUP98::KDM5A is dependent on H3K4me3 binding. Their model proposes that concentration-dependent chromatin-associated condensation of NUP98::KDM5A depends on local densities of H3K4me3 on chromatin and the levels of the fusion oncoprotein. In line with this, the analysis of gene expression data from NUP98::KDM5A-positive AML cells shows a positive correlation between differentially expressed genes and H3K4me3 levels.

This is an interesting manuscript that aims to dissect the molecular mechanisms underlying biomolecular condensation of the NUP98::KDM5A oncoprotein. The work is solid, and the results are well explained and presented in a logical order.

Reviewer #2 (Public review):

In this manuscript, the authors investigate how the oncogenic fusion protein NUP98-KDM5A alters gene expression in leukemia, using a combination of cellular experiments with model and patient cell lines, as well as in vitro studies. Upon transfection of U2OS cells with mEGFP-tagged NUP98-KDM5A, the authors show that the fusion proteins form sub-micrometer puncta, whereas KDM5A alone does not. These foci are also observed at expected native expression levels (using OpenCell data). The tag has an effect here, as switching to an mCherry tag raises the apparent saturation concentration for phase separation. Finally, the authors show via super-resolution imaging that the foci correlate with H3K4me3 distribution.

In vitro, the fusion protein forms amorphous, gel-like condensates at double-digit nanomolar concentrations. Truncation analysis identifies PHD3 of KDM5A as required for maximal phase separation, consistent with the ability of the protein to bind H3K4me3 peptides. Addition of polynucleosomes increases the amount of fusion protein partitioning into the condensate in an H3K4me3-binding-dependent manner. Condensates are gel-like with slow internal dynamics in vitro; in cells, however, the dynamics depend on the position of the EGFP tag (no corresponding experiments with mCherry are shown). Reconstitution with H3K4me3- and H3K4me0-modified arrays shows colocalization with both wild-type NUP98-KDM5A and the binding mutant. Here, H3K4me3 arrays recruit ~20% more protein and yield gel-like structures in a manner dependent on the PTM and on the PHD finger.

In cells, the fusion protein colocalizes with H3K4me3-marked loci, including the HOX clusters, as confirmed by FISH. Finally, re-analysis of published expression datasets from patient cells shows that genes are predominantly upregulated and that the upregulated genes are H3K4me3-marked.

This is a well-executed mechanistic study. The data convincingly establish that NUP98-KDM5A forms sub-micrometer foci at realistic expression levels, that these foci correlate with H3K4me3-marked sites, that the PHD3-H3K4me3 interaction mediates chromatin binding while the NUP98 moiety drives phase separation in vitro, that foci in cells overlap genes heavily decorated with H3K4me3, and that H3K4me3-marked genes are those found to be upregulated in patient datasets. These are important mechanistic findings and of interest to the community.

Author response:

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

Public Review:

Reviewer #1 (Public review):

Leukemia-driving NUP98 oncofusion proteins form chromatin-associated biomolecular condensates in the nucleus, and these structures are important for oncogenic transformation. Most NUP98 fusions do not contain domains that mediate the recognition of specific DNA elements. Instead, they entail domains that are important for chromatin regulation. For instance, the NUP98::KDM5A fusion features a fusion of the NUP98 N-terminus with the third PHD domain of the histone demethylase KDM5A. As PHD domains are critical for the recognition of methylated histones without any sequence specificity, it is not clear what controls the condensation and chromatin binding of NUP98::KDM5A, leading to the induction of oncogenic transcriptional programs.

In this work, the authors use a combination of cellular and in vitro studies to show that biomolecular condensation of NUP98::KDM5A is dependent on H3K4me3 binding. Their model proposes that concentration-dependent chromatin-associated condensation of NUP98::KDM5A depends on local densities of H3K4me3 on chromatin and the levels of the fusion oncoprotein. In line with this, the analysis of gene expression data from NUP98::KDM5A-positive AML cells shows a positive correlation between differentially expressed genes and H3K4me3 levels.

This is an interesting manuscript that aims to dissect the molecular mechanisms underlying biomolecular condensation of the NUP98::KDM5A oncoprotein. The work is solid, and the results are well explained and presented in a logical order. However, the study suffers from several weaknesses that if addressed would improve the study.

Major points:

(1) All cellular experiments are performed in settings of transient transfection of NUP98::KDM5A in non-hematopoietic cell types. These conditions are not physiologically relevant, as these cells do not depend on the fusion oncogene. Therefore, any claims about concentration-dependent effects on condensation need to be validated in AML cells that are driven by NUP98::KDM5A. While this may not be possible in primary patient-derived cells, several groups have published AML models of NUP98::KDM5A-driven AML that could be used.

We appreciate the reviewer’s point regarding physiological relevance. Our goal in these cellular microscopy experiments was to dissect the biophysical principles of NUP98::KDM5A (including its mutants) condensate formation under controlled expression levels (concentration). While AML model systems driven by NUP98::KDM5A are available, they do not offer this possibility because of pre-existing NUP98::KDM5A expression. Suspension culture of hematopoietic cells also poses practical challenges for correlative FISH+IF and high-resolution microscopy analysis. We agree that validating our observed behaviors in AML models would be valuable, e.g. by creating HSPC lines with inducible expression of tagged NUP98::KDM5A and its mutants, but such experiments fall outside the scope of the current study. We will add text acknowledging this limitation and clarifying that our mechanistic conclusions are grounded in biophysical principles that should generalize across cell types.

In the revised manuscript, we have added a section on study limitations at the end of the discussion section:

Study limitations: Cellular experiments in this study were performed using transient transfection of WT NUP98::KDM5A and its mutants in HEK293 and U2OS cells. While these non-hematopoietic cell models limit physiological relevance, they allowed for controlled expression levels in the absence of endogenous oncofusion and the ease of imaging, enabling investigation of the biophysical principles of condensate formation, the main objective of our study.

(2) The results presented in Figure 4 are not entirely supportive of the mechanism. It is known that active gene expression correlates with high H3K4me3 levels; therefore, the correlations shown by the authors are expected. Yet, the authors do not discuss the fact that many H3K4me3-positive genomic regions do not show NUP98::KDM5A binding. This should be elaborated on in the discussion section.

We agree that active transcription correlates with H3K4me3, and that this baseline relationship must be considered. Our analysis explicitly uses fold-change between patient cells and healthy controls as the readout. This comparison inherently accounts for the activating effect of H3K4me3 itself. Regarding the reviewer’s comment that “many H3K4me3-positive genomic regions do not show NUP98::KDM5Abinding”, we would like to clarify that this point is exactly what our manuscript aims to explain. Our cell line studies show that, at patient-relevant expression levels, NUP98::KDM5A condensates form preferentially at H3K4me3 loci with high local mark density. Considering that NUP98::KDM5A concentration in the nucleus is lower than the KD between KDM5A PHD3 and H3K4me3, this means that H3K4me3 loci with lower mark density will not see NUP98::KDM5A binding without the high local concentration of the fusion protein (as a result of condensate formation). This is consistent with our observation that genes with the highest local density show disproportionately stronger upregulation.

We have modified the final paragraph of the section describing our analysis of gene expression in patient samples to add:

Importantly, to account for the activating effect of H3K4me3 itself, our analysis explicitly uses fold-change between patient cells and healthy controls as the readout.

(3) While the focus of the manuscript is on NUP98::KDM5A, this oncofusion is part of a family of >30 fusions that join the NUP98 N-terminus to a variety of factors with roles in epigenetic control and transcription. While the repertoire of NUP98 fusion partners is diverse with regard to functional domains, they all induce a conserved set of target genes that is characteristic of this leukemia subtype. How can this be achieved in the context of NUP98 fusion proteins that do not contain a PHD domain, such as NUP98::NSD1 or NUP98::HOXA9? Please discuss this.

We appreciate this important conceptual question. Many NUP98 fusions, despite diverse partner domains, produce similar transcriptional programs. Our current hypothesis is that the highly active status of the HOX cluster genes in HSPC attracts NUP98 oncofusions targeting H3K4me3 (e.g., KDM5A and PHF23), while NUP98::HOXA9 and other transcription factor fusions directly target the HOX cluster via DNA sequence recognition. This convergence and the broad targets of the dysregulated HOX transcription factors explain the transcriptional program similarity, sustained by the previously described enrichment of transcriptional co-activators by NUP98 oncofusion condensates. Of note, while several fusion partners of NUP98 lack PHD domains, the NUP98::NSD1 oncofusion includes several PHD domains of NSD1, which have previously been suggested to be important for leukemogenesis (PMID: 17589499). The discussion section has been modified to add:

We envision that a similar mechanism targets other NUP98 fusions containing H3K4me3-recognizing reader domains (e.g., NUP98::PHF23) to leukemogenic loci, complementing direct targeting of the HOX genes through DNA sequence recognition in fusions containing DNA-binding domains (e.g., NUP98::HOXA9).

Reviewer #2 (Public review):

In this manuscript, the authors investigate how the oncogenic fusion protein NUP98KDM5Aalters gene expression in leukemia, using a combination of cellular experiments with model and patient cell lines, as well as in vitro studies. Upon transfection of U2OS cells with mEGFP-tagged NUP98-KDM5A, the authors show that the fusion proteins form sub-micrometre puncta, whereas KDM5A alone does not. These foci are also observed at expected native expression levels (using OpenCell data). The tag has an effect here, as switching to an mCherry tag raises the apparent saturation concentration for phase separation. Finally, the authors show via super-resolution imaging that the foci correlate with H3K4me3 distribution.

In vitro, the fusion protein forms amorphous, gel-like condensates at double-digit nanomolar concentrations. Truncation analysis identifies PHD3 of KDM5A as required for maximal phase separation, consistent with the ability of the protein to bind H3K4me3 peptides. Addition of polynucleosomes increases the amount of fusion protein partitioning into the condensate in anH3K4me3-binding-dependent manner. Condensates are gel-like with slow internal dynamics in vitro; in cells, however, the dynamics depend on the position of the EGFP tag (no corresponding experiments with mCherry are shown). Reconstitution with H3K4me3- and H3K4me0-modifiedarrays shows colocalization with both wild-type NUP98-KDM5A and the binding mutant. Here, H3K4me3 arrays recruit ~20% more protein and yield gel-like structures in a manner dependent on the PTM and on the PHD finger.

In cells, the fusion protein colocalizes with H3K4me3-marked loci, including the HOX clusters, as confirmed by FISH. Finally, re-analysis of published expression datasets from patient cells shows that genes are predominantly upregulated and that the upregulated genes are H3K4me3-marked.

This is a well-executed mechanistic study. The data convincingly establish that NUP98KDM5A forms sub-micrometre foci at realistic expression levels, that these foci correlate with H3K4me3-marked sites, that the PHD3-H3K4me3 interaction mediates chromatin binding while the NUP98moiety drives phase separation in vitro, that foci in cells overlap genes heavily decorated with H3K4me3, and that H3K4me3-marked genes are those found to be upregulated in patient datasets. These are important mechanistic findings and of interest to the community.

Still, the functional/causal link is a bit more tentative, as the data is mostly correlative, since it is not directly established that there is feedback between H3K4 methylation, NUP98KDM5Arecruitment, phase separation and target gene overexpression. An experiment that could further bolster this claim would be a direct test of whether NUP98-KDM5A expression drives overexpression of bound genes, e.g. expression of the fusion protein vs PHD- and NUP98-mutant variants, followed by qPCR of target genes, such as the HOX cluster, and possibly H3K4me3 ChIP at the same loci. As all the constructs and cell lines exist, this could be feasible and would substantially strengthen the manuscript.

We thank the reviewer for the thoughtful evaluation and agree that the causal link betweenH3K4me3 recognition, condensate formation, and gene activation remains partly correlative. Experiments such as qPCR or ChIP following expression of WT versus mutant constructs would indeed strengthen the causal chain. However, performing these assays across multiple constructs and loci in a physiologically relevant system is not feasible within the current revision cycle. We have added text acknowledging this limitation and clarifying that our study focuses on establishing the biophysical mechanism of targeting, while functional consequences are inferred from patient datasets rather than new perturbation experiments.

We modified the Study limitations section of the Discussion to add:

Similarly, our study focuses on the biophysical mechanism of chromatin targeting by the NUP98::KDM5A oncofusion, with functional consequences inferred from patient-derived datasets rather than defined perturbation experiments.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

(1) Figure 1 - Figure Supplement 2: show co-localisation of large droplets of the chromatin-binding mutant with nucleoli by co-immunostainings, using nucleolar markers.

(2) Figure 1E: Show another histone mark as a control, ideally H3K27me3.

(3) The authors are encouraged to stick to the recommended nomenclature of fusion genes(https://www.nature.com/articles/s41375-021-01458-0).

Reviewer #2 (Recommendations for the authors):

(1) The phase-separation behaviour clearly depends on the fusion FP, as the switch from mEGFP to mCherry shifts the apparent saturation concentration to ~200 nM. The authors should discuss, and where possible test, how much of the observed behaviour is attributable to the FP itself: what happens in the absence of any FP, and which construct (mCherry- or mEGFP-tagged) the authors consider the better proxy for the untagged fusion protein.

(2) This is particularly relevant given that the in-cell dynamics (FRAP) are reported to depend on the position of the EGFP tag, and no corresponding mCherry experiments are shown.

(3) The authors claim a patient NUP98-KDM5A concentration of ~180 nM. This is based on data for NUP98 in HEK293 cells, with the added correction factor for having only one allele mutated in patients. Still, the direct jump from HEK293 to patient leukemia cells is big; so the authors should formulate this more carefully: no conc. determination was done directly in patients.

(4) Chicken erythrocyte chromatin: What is the endogenous H3K4me3 level in the chicken erythrocyte chromatin used here?

(5) Figure 1 - Figure Supplement 3B. The legend defines the square as the mean but does not explain the remaining symbols (vertical lines, horizontal lines, grey versus black dots). Please define these plot elements.

(6) In the reconstitution experiments, the arrays and the condensates appear to colocalize only weakly, and, a bit counterintuitively, the strongest apparent colocalization is seen for the NUP98KDM5A binding mutant. Can the colocalization be quantified and interpreted?

(7) Figure 2 - Figure Supplement 3. Raising the chromatin concentration from 12.5 to 100 nM produces little visible change in the images. Where does the additional chromatin go? Does it stay in solution? As in point 4, the best colocalization again appears to be obtained with the mutant, which is unexpected and a bit difficult to reconcile with the proposed H3K4me3-PHDdrivenrecruitment model.

(8) That highly expressed genes in patients carry H3K4me3 is basically expected and is not by itself informative. The relevant question is the origin of the effect: is H3K4me3, and the associated upregulation, a consequence of NUP98-KDM5A condensate formation? Analyzing the direct effects of NUP98-KDM5A (and mutants) on specific target gene expression profiles could help to dissect this further.

We thank the reviewers for these recommendations. We have made the following revisions to the manuscript:

We replaced NUP98-KDM5A with NUP98::KDM5A, in accordance with the recommended nomenclature. We made this change in both the text and the figures. We also corrected other NUP98 on confusions to reflect this recommendation.

We have revised Figure 1 – Figure Supplement 3B and the accompanying legend to describe the symbols used in this figure.

eLife Assessment

This valuable manuscript contains exciting data that has the potential to advances understanding of how psilocybin acutely alters decision making by modulating the impact of outcomes on future choices and altering how prefrontal cortex neurons encode outcomes. While the experimental paradigm, behavioral and neural analysis are elegant, the strength of evidence is, at present, incomplete given the lack of a saline treated comparison. This work will be of interest to behavioral and computational neuroscientists studying decision making, flexibility, or psychedelics.

Reviewer #1 (Public review):

Summary:

Jacobs et al. report how acute psilocybin influences value-based decision-making and associated neural correlates in the medial prefrontal cortex (mPFC) of rats performing a probabilistic two-choice task with reward probability varying in blocks. The authors highlight that acute psilocybin does not influence basic outcome-sensitive choice strategies (i.e., win-stay/lose-shift), but find that psilocybin increases choice variability and optimal choices in low-uncertainty contexts. Using RL modeling, the authors chose one of five models examined because it represented the behavior around transitions. This recapitulates their behavioral results, showing changes in choice variability (entropy) specifically under high-uncertainty conditions and improving asymptotic performance under low uncertainty. Single-unit recording from mPFC during the task revealed that psilocybin may enhance the representation of reward-preferring neurons and suppress non-reward-preferring neurons, with no changes to outcome encoding and total/relative value signals largely intact. The authors interpret these findings as evidence that psilocybin rebalances mPFC processing to favour learning from positive outcomes, potentially offering a candidate neural coding mechanism for psilocybin's purported ability to increase cognitive flexibility acutely. While the research question is highly significant, the experimental design, analysis, and interpretations have critical shortcomings.

Strengths:

Understanding how psilocybin influences behavioral effects like reward learning and cognitive flexibility is critical to understanding its potential use in the clinic. The behavioral, neural, and modeling studies use a powerful set of approaches that are useful for understanding this complex problem.

Weaknesses:

(1) Design and sample size were major concerns. Only 8 rats were used, and the primary comparison is within-subject psilocybin vs. the immediately preceding "baseline" day rather than a fully counterbalanced or randomized drug/vehicle design. Since psilocybin sessions always followed baseline sessions (never the reverse), order effects, practice effects, or day-to-day drift cannot be fully excluded. Additionally, a saline control was only available for 5 of 8 animals, on a haphazard schedule ("due to unanticipated technical issues"), so the claim that effects are pharmacologically specific to psilocybin rather than to injection/novelty stress rests on a small, incomplete control dataset.

(2) There are a number of major concerns regarding behavioral confounds. Results about the number of trials performed in each condition seem buried in the methods instead of being provided with the rest of the behavioral data. Critically, the results show a 40% reduction in the number of trials completed. This is important to highlight and consider in the context of reduced motivation/hunger, and also in terms of the associated neural activity. Additionally, the percentage of rewarded trials is not reported - this is important for the interpretation of the behavioral results. For example, is the increased choice variability following psilocybin treatment measured by entropy advantageous? Does this result in more or fewer rewarded trials?

(3) Any differences in neural data are greatly diminished by the behavioral differences. Given the extraordinary long latency (averages of up to 20s) to reward receipt, it is unclear how the neural data in Figure 3 referencing reward is meaningful given the time frame is a maximum of 1.5s post-response. How can neurons be labeled non-reward-preferring based on their activity well before they retrieved the reward? Further, it is unclear how the post-response neural activity should be interpreted in the context of this long and variable latency to retrieve reward. The abnormally long latency to retrieve a reward is not discussed, especially in the context of response latencies on the order of 2 to 4s. Furthermore, the difference between groups (7 versus 20 s) should be discussed because this may represent a difference in motivation/hunger/response drive between groups and influence the interpretation of both behavioral and neural data.

Additionally, it is unclear how to interpret the neural data given the offset timing between groups because of the large behavioral differences. For example, given the decrease in latency to respond following the psilocybin treatment, the neural data shown (aligned to the response; e.g., Figure 6A, B) is largely offset from the time of trial initiation. There is no discussion of what is occurring behaviorally during the latency, which may be the cause of the different neural signatures.

(4) For the RL Modeling, there is circularity in the author's approach to choosing a model - namely, based on the "posterior predictive ability to recover psilocybin effects". The author's process for arriving at Model 3 as the specific "preferred" model over the intermediate options (particularly Model 4 with stickiness) is not well-described. The validation step (of recovering the key psilocybin behavioral transition signatures) is circular and not sound evidence for model adequacy rather than independent confirmation. I would want to see the actual PSIS-LOO comparison table (with SEs and Pareto k diagnostics) from Supplemental Figure 2A before fully accepting that Model 3 was clearly the best-justified choice rather than simply the most convenient one given convergence constraints. There was no prior predictive check with the parameters being chosen from the literature. Additionally, the Bayesian parameter effects rely on a relatively lenient 80% HDI threshold (rather than 95%).

(5) Some minor concerns:

a) The order of the Type I and II neurons appears to be reversed in this sentence based on the rest of the results/figures. "These neurons had distinct temporal patterns, showing reward-related increases immediately at action or later in the action/outcome epoch and are depicted as Type I and Type II, respectively".

b) Connecting results from Figures 3 and 4 would be interesting; in other words, showing how Type I/II neurons look in Figure 4 plots.

c) In Figure 1H, are the post-transition trials collapsed across high and low next blocks, since it seems from the depicted trial order that low uncertainty blocks can be followed by either a high uncertainty block or another (reversed) low uncertainty block.

d) The methods by which the authors arrived at the binned value neurons are not well-described, and more clarity on how authors arrived at relative and total value estimates would be helpful.

Reviewer #2 (Public review):

Summary:

The authors set out to examine how psilocybin acutely alters learning in rats in a reward-based probabilistic choice task. Additionally, they asked how psilocybin acutely alters encoding of decision variables in the prefrontal cortex (PFC).

Strengths:

The behavioral paradigm is robust. Thorough and rigorous computational analyses suggest that psilocybin may differentially modulate the impact of rewarded and unrewarded outcomes and enhance exploration under uncertainty. Single-unit neural activity in PFC suggests a potential mechanism for these effects, with psilocybin inducing opposing effects on encoding of rewarded and unrewarded actions. These findings are potentially exciting and important.

Weaknesses:

The major weakness of this work is the lack of a saline-treated control. All conclusions rest on analyses that are limited to within-animal comparisons between a baseline session and a later session in which rats were injected with psilocybin prior to session start. There is no control for the injection stress or passage of time. The manuscript mentions that 5 of 8 mice were also tested in a session following a saline injection, but because data were lost for some animals, this was not pursued. Figure S1G reports very limited analysis of this data to suggest that there is no effect. At a minimum, this must be analyzed in full (behavioral and neural data) and included in the manuscript to support the central claim that the effects reported are indeed attributable to psilocybin.

Beyond this, it would be of interest to examine if the effect of psilocybin is restricted to the acute session or if any effects persist.

Reviewer #3 (Public review):

Summary:

Current theories suggest psilocybin promotes meta-learning, modifies learned associations, and promotes flexibility, contributing to its transdiagnostic clinical relevance. However, the effects of psilocybin on neural state dynamics during learning and decision-making are unclear. Jacobs et al. use a rat model to examine how psilocybin changes prefrontal cortex encoding during a value-based probabilistic choice task. The authors found psilocybin increased high-value choices when uncertainty was low and increased choice variability when uncertainty was high, compared to baseline performance. A reinforcement learning model found psilocybin decreased side bias and increased reward learning while decreasing the unrewarded learning rate. Encoding of rewarded actions was enhanced in the mPFC while unrewarded actions were attenuated.

Strengths:

The neural correlates and computations underlying the impacts of psychedelics are poorly understood, and this study examines these through well-established decision-making approaches. These data show changes in mPFC activity during a dynamic task, allowing for analysis across different contexts and events. The data presented provide some support for leading theories on psilocybin's therapeutic use in promoting meta-learning and flexibility. The combination of techniques, including decision-making, computational modeling, and neural recordings, is appropriate, powerful, and needed to make sense of the underlying change in experience triggered by psilocybin.

Weaknesses:

An unfortunate weakness in the design that limits the interpretability of the findings is a lack of control data under saline administration, and a lack of comparison of psilocybin to saline rather than baseline. Psilocybin performance is compared to a single baseline session, which lacks the effect of an injection that would be present in a saline control. There is some evidence that mild stress can itself impact outcome encoding. Supplemental data in this manuscript suggest there is no significant difference in probability of selecting the optimal choice on low uncertainty trials between baseline and saline; however, the direction of impact of the saline injection is in the same direction as psilocybin in the main figures. This means that the mild stress of simply receiving an injection of saline may not be totally neutral with regard to these measures, and may contribute to the effects attributed to psilocybin at both a behavioral and neural level.

Author response:

I am writing to provide a description of our planned revisions. We thank the reviewers for their thoughtful comments on the manuscript. All three reviewers had the primary concern of “lack of saline-treated control.” We had, in fact, included saline-treated control in our design, and shown some of this data in the original submission (Supplemental data Fig 1G). We agree with the Editor’s note that more analysis for saline sessions should be provided. The planned revisions will include additional saline data to compare saline injections on behavior (as shown for psilocybin on figure 1), on reward-preferring and non-preferring cluster analysis (as shown for psilocybin on figure 3), and on action decoding and encoding (as shown for psilocybin on figure 6).

It should be emphasized that the work involved outbred rats performing a complex probability-shifting decision-making task while freely moving in an operant box. Some of the comments by reviewer 1 (such as considering long/changing latencies to retrieve reward as “extraordinarily” or “abnormally” long) suggest that there may have been a mix-up with what is common in freely-moving rats versus simpler and head-fixed mouse paradigms. This distinction is relevant to some of the concerns raised by reviewer 1. First, our n is well within the range seen in electrophysiological studies in behaving rats as most these studies utilize n=5-9 (e.g., Sias et al., Nature Neuro, 2024; Maor et al., Nature Comm, 2026). Second, our analysis dealt with shifting latencies accordingly (e.g. time-locking to action) and focused on within subject performance because in complex cognitive tasks such as ours, there is a considerable amount of baseline performance variability in outbred rats. It is, therefore, common, and clinically relevant, to focus on within subject analysis of drug or vehicle effects. This has been the case with previous studies on the effect of psilocybin on cognitive flexibility (e.g. Torrado Pacheco et al, NPP 2023) and was our approach here. The additional saline data will hopefully convince the reviewers and interested readers that the reported effects of psilocybin were not an artifact of i.p. injections

eLife Assessment

This study provides valuable insights into humans' ability to generalize learned structural knowledge to new experiences that share underlying graph components but involve different stimuli. The behavioral and MEG evidence is solid, broadly supporting the transfer of abstract structural knowledge across contexts, although the findings speak more directly to the reuse of individual structural components than to their recombination. This study will be of interest to cognitive neuroscientists studying structure learning, abstract representations, and the generalization of knowledge across contexts.

Reviewer #1 (Public review):

Summary:

The paper presents an elegant task designed to investigate humans' ability to generalize knowledge of learned graph structures to new experiences that share the same structure but are built from different stimuli. Using behavior and MEG recordings, the authors test evidence for neural representation and application of structural knowledge.

Comments on revised version.

The authors have addressed my concerns and have largely toned-down the overstatement of their findings which was pervasive in the previous version of the manuscript. However, the revised title still seems rather misleading given the findings presented. I suggest the authors revise this to: 'A neural mechanism for transfer of abstract structural knowledge in humans'.

Reviewer #2 (Public review):

The authors track structural role representations during stimulus processing, which is relevant to study how prior experience shapes new learning. They employ a sequence learning task based on graph factorization and record brain activity using MEG to investigate whether graph factors are reused during learning and inference in a new graph. MEG data were recorded during passive stimulus presentation, while behavior is assessed in a subset of probe trials testing either experienced or inferred successions in the graph. The authors report neural similarity among stimuli conditional on their role in subprocesses, and a correlation between evidence for role representations and performance in experience probes, but not in inference probes.

The interpretation of "knowledge transfer" in the title and abstract is, however, constrained by the absence of a corresponding relationship with performance in the inference probes, which provide the behavioral measure of transfer. While the authors discuss why role representations might relate more closely to experience than inference probes in this task, this distinction makes the link between the neural findings and transfer weaker than the current framing suggests. (E.g., "was positively related to behavioral success in subprocess knowledge transfer" in the abstract could be misleading)

More broadly, the framing of the results as a mechanism for structure transfer or structural scaffolding appears stronger than warranted by the evidence. The findings indicate representations consistent with reusable roles, but do not establish that these representations functionally facilitate learning and transfer. The claims should align with this distinction between evidence for representational reuse and structural scaffolding.

Also, apart from abstract roles, simpler categorical grouping of different stimuli belonging to the same subprocess could produce increased similarity.

Relatedly, although the MEG-behavior (experience probes) correlation is significant for Component 6 but not Component 4, the difference between the two correlations is not significant. Interpreting the correlation as selective to the structurally relevant graph is thus difficult to statistically substantiate.

Overall, the study provides a well-designed paradigm for investigating compositional representations during structural generalization at high temporal resolution. However, the findings do not provide sufficient evidence to frame them as a mechanism for structure transfer.

Reviewer #3 (Public review):

Summary:

This study investigates how task components can be learned and transferred across different task contexts. The authors designed two consecutive sequence learning tasks, in which complex image sequences were generated from the combination of two graph-based structural "building blocks". One of these components was shared between the prior and transfer task environments, allowing the authors to test compositional transfer. Behavioral analyses using generalized linear models (GLMs) assessed participants' sensitivity to the underlying structure. MEG data were recorded and analyzed using classifications and feature representational similarity analysis (RSA) to examine whether neural similarity increased for stimuli sharing the same relational structure. The paper aims to uncover the neural dynamics that support compositional transfer during learning.

Strengths and weaknesses:

I found the methods and task design of this paper difficult to follow, particularly the way stimuli were constructed and how the experimental sequences were generated from the graph structures. These aspects would be hard to replicate without some clarification. I appreciate the integration of behavioral and neuroimaging data. The overall approach, especially the use of compositional graph structures in sequence learning, is interesting and could be used and revised in further studies in compositionality and transfer learning. I appreciated the authors' careful interpretation of their findings in the discussion. However, I would have liked a similar level of caution in the abstract, which currently overstates some claims.

Comments on revised version.

No further comments. I thank the authors for their diligence in addressing my comments.

Author response:

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

eLife Assessment:

This study provides valuable insights into humans' ability to generalize knowledge of learned graph structures to new experiences that share the same structure but are built from different stimuli. However, the evidence for the authors' claims is incomplete, with the main claims of structural generalization and compositionality only partially supported by MEG and behavioral data. This study will be of interest to cognitive neuroscientists studying structure learning and generalization.

We thank the editor and reviewers for a positive evaluation of our work and constructive suggestions to improve our manuscript.

The reviewers' comments motivated textual modifications to the manuscript, changes and additions to figures and conduct of further analyses.

We believe that the changes made in response to the reviewers' comments and suggestions provide for a substantially improved revised manuscript.

Public Reviews:

Reviewer #1 (Public review):

Summary of the paper:

The paper presents an elegant task designed to investigate humans' ability to generalize knowledge of learned graph structures to new experiences that share the same structure but are built from different stimuli. Using behavior and MEG recordings, the authors test evidence for neural representation and application of structural knowledge.

We appreciate this overall positive evaluation of our work and suggestions to strengthen the empirical support for the claims made in the paper.

(1) While the task design is elegant, it isn't clear to me that the data support all the claims made in the paper. I have detailed my concerns below.

The authors claim that their findings reveal "striking learning and generalization abilities based on factorization of complex experiences into underlying structural elements, parsing these into distinct sub processes derived from past experience, and forming a representation of the dynamical roles these features play within distinct subprocesses." And "neural dynamics that support compositional generalisation, consistent with a structural scaffolding mechanism that facilitates efficient adaption within new contexts".

(a) First, terms used in these example quotes (but also throughout the paper) do not seem to be well supported by data or the task design. For example, terms such as 'compositional generalisation' and 'building blocks' have important relevance in other papers by (some of) the same authors (e.g., Schwartenbeck et al., 2023), but in the context of this experiment, what is 'composition'?

We agree that our use of composition is less technical and formal than the nested and recursive forms of composition using “building blocks” reported in previous work (e.g., Schwartenbeck et al., 2023). In this experiment we study composition of smaller graphs via superimposition of their outputs. The superimposed “compound” outputs can be viewed as outputs of a single, more complex, graph. This is a special case of all possible ways that graphs can be composed from other graphs, and does not depend on a complex syntax for their combination (although more complex combination rules, other than multiplication are theoretically possible). Nevertheless, in accordance with common definitions of compositionality, this leads to a generative setup, given that a huge number of graphs can be expressed as the product of smaller graphs in this way.

Figure panel (Fig 1A) illustrates this precise point. Importantly, we argue this kind of composition is ubiquitous in real-world learning scenarios where multiple independent processes contribute to the holistic experience of a situation.

Based on the reviewer’s suggestion we have now amended the manuscript introduction to render the distinction between different definitions, and usage, of the term composition more salient.

It is our understanding that this form of flexible, combinatorial, reuse of factorized prior knowledge is considered a specific instance of composition. Indeed, this understanding is supported by a cognitive-computational neuroscience literature that adheres to similar conceptualizations of compositionality, e.g., Barron et al., 2013 Nat Neuro; Rubino et al., 2023 PsyArxiv; Luettgau et al., 2024 PNAS.

We add a paragraph to the introduction section elaborating on what we mean by graph factorization and its relevance for questions of compositional generalization (reproduced below for convenience).

To address these questions, we developed a sequence learning paradigm based on a graph theoretical concept of graph factorization, emulating environments whose dynamics are the product of multiple simultaneous subprocesses. “Product graphs (Cartesian product), wherein the joint dynamics of two or more graph factors create complex product state space dynamics that can be factorized (and simplified) into constitutive factors, provides a compelling and flexible test bed for probing compositional generalization of multiple (independent) and simultaneously unfolding subprocesses.” Examples of such graph products and their factorizations, from a vast space of possible combinations, are provided in Figure 1A. Here we take inspiration from prior studies in neuroscience (e.g., Barron et al., 2013; Rubino et al., 2023; Schwartenbeck et al., 2023) and a prior cognitive literature, e.g., recursive operations over hierarchies (O’Donnell et al., 2009, 2011), hierarchical function composition, or rich forms of semantic composition in language to create complex concepts (Murphy, 1988).

(2) Can the authors demonstrate clear behavioural or neural evidence for compositional use of multiple graph structures, or alternatively remove reference to these terms? In the current paper, it seems to me that the authors are investigating abstract knowledge for singular graph structures (together with the influence of prior learning), as opposed to knowledge for the compound, more complex graph formed from the product of two simpler graphs.

We thank the reviewer for this important clarification, and agree our language should more precisely distinguish what the task design requires from what our data directly demonstrate. The reviewer is correct that our analyses primarily provide evidence for the abstraction and transfer of a single previously-experienced graph structure to novel stimuli, rather than direct evidence that participants combine two independently learned structures into a novel compound representation. We agree this distinction is important, and we have revised our language throughout the manuscript accordingly, reserving "compositional generalization" to describe the broader task structure (in which every compound stimulus is the joint product of two simultaneously evolving subprocesses, and the specific compound experienced during transfer learning was never observed during prior learning) and using more precise language: "decomposition," "abstraction," and "transfer" of single structural components to describe what our behavioral and neural evidence actually establishes. We have not claimed direct evidence that participants compositionally construct novel joint representations from two previously separate structures; rather, our evidence concerns the selective reuse of one such component when it recurs across learning phases, embedded in a novel pairing with another, entirely new structural component. We have adjusted the manuscript throughout and changed the title of the paper to “A neural mechanism for compositional structure transfer in humans”, to more adequately reflect the evidence presented.

(3) While I would like to be convinced that this data provides evidence for the transfer of abstract, structural knowledge, I think the authors either need to provide more convincing evidence or tone down their claims.

Can the increase in neural similarity between stimuli mapping to the same abstract structural sub-process not be explained by temporal proximity in experiencing the transitions (e.g., Cai et al., 2016)? Indeed, behavior seems to be dominated by direct experience of the structure as opposed to applying abstract knowledge of equivalent structures (and, as a result, there is little difference in behavioural performance between experience and inference probes).

The reviewer makes an astute observation while providing a constructive challenge. We agree that distinguishing transfer of abstract structural knowledge from effects driven by direct temporal experience is critical, and we appreciate this opportunity to clarify our reasoning on this point.

The reviewer suggests that increased neural similarity between stimuli that map to the same abstract structural sub-process might be explained by a temporal proximity in experience, as shown in prior work (e.g., Cai et al., 2016). We agree that temporal proximity is a potentially powerful determinant of representational similarity. Likewise, we do not dispute that temporal statistics of experience play an important role in shaping such neural representations. Nevertheless, we believe these accounts do not provide a sufficient explanation for our results, which we consider reflects a consequence of being exposed to the outputs produced by sequences drawn from an underlying graph structure and a prior learning of these sequences.

In our task, the temporal statistics of stimulus experience are fully dictated by the transition structure of the graph. Different graph structures necessarily impose different constraints on temporal proximity (e.g., how many steps it takes to revisit a given node, which nodes are encountered at similar temporal distances, etc.). Thus, if stimuli that occupy equivalent structural roles across different graphs come to be represented more similarly, this is precisely because learners have abstracted over the specific transition structure that governs these temporal relationships. In other words, consistent with our previous computational findings that models based on temporal difference learning mechanisms account well for behavioural response patterns (Luettgau et al., 2024 PNAS), temporal proximity is not a confound but instead is a likely mechanism through which structural knowledge is acquired. Critically, our key claims concern transfer across learning contexts, and not sensitivity to temporal proximity within a single experience stream. By design, the selective learning improvements observed in each group during transfer cannot be explained by temporal proximity alone, because the relevant temporal statistics in the transfer phase are novel and differ across groups. The fact that prior learning selectively facilitates subsequent learning depending on an earlier experience of an abstract structure indicates participants are not simply tracking local temporal regularities, but instead are reusing previously learned structural knowledge to guide learning within a new context.

We agree that behavior in our task is strongly influenced by direct experience, and we do not claim that participants rely exclusively on abstract structural knowledge. Indeed, the relatively small differences between experience and inference probes are consistent with the view that abstraction complements and facilitates, rather than replaces, learning from direct experience. Our claims are therefore deliberately modest: we argue for the transfer of abstract structural knowledge that shapes learning dynamics and neural representations, not for behavior dominated by explicit inference over abstract graphs.

To avoid any criticism related to over-interpretation, we have toned down our language and clarify throughout introduction, results and discussion section of the manuscript that (i) temporal proximity is an intrinsic feature of graph structure, (ii) neural effects reflect (in part) abstraction over these structured temporal relationships, and (iii) that our results provide evidence for transfer of structural knowledge that biases learning, rather than for fully context-independent representations or strong behavioral generalization.

(4) The strongest evidence for neural representation of abstract task structures seems to be the increase in similarity by transition type. But this common code for 'transition type' is only observed for 6-bridge graphs and only for experienced transitions. There was no significant effect in inference probes. Therefore, there doesn't seem to be evidence for the application of a knowledge scaffold to facilitate transfer learning. Instead, the data reflects learning from direct experience and not generalisation.

We appreciate the reviewer's point, and agree it deserves a careful investigation. We carefully considered how this correlation could be interpreted under a non-transfer, pure direct-experience account. Our interpretation is that even if all graph learning relies entirely on direct experience, a correlation is still expected due to individual differences in learning ability. Under this framework, higher-ability participants are simply more likely to directly discover the same structural component (the 6-bridge) across both sessions (prior and transfer learning), which in turn supports higher MEG transition decodability. Concurrently, these same participants will exhibit higher behavioral accuracy due to their superior learning capacity. Importantly, under this account, these higher-ability participants should also show higher behavioural accuracy during the transfer phase for the 4-cycle component, predicting (a) a positive correlation between our transition decoding signal and 4-cycle behavioural accuracy, and (b) a positive correlation between 4-cycle and 6-bridge behavioural accuracy themselves, within each prior learning group.

We tested both predictions directly. For (a), in the 6-bridge prior group, average decoding strength at timepoints showing a selective significant group difference correlated with experience probe performance on the 6-bridge graph during transfer learning (ρ = .48, p = .016, surviving correction for two comparisons, α = .025), but the corresponding correlation with 4-cycle behavioural accuracy did not reach significance (ρ = .38, p = .071). We acknowledge, however, that a direct comparison of these two dependent correlations (Steiger's test) did not reach significance (p = .66).

For (b), we directly correlated 4-cycle and 6-bridge transfer accuracy within each prior learning group. For experience probes, this correlation was significant in the 4-cycle prior group (ρ = .51, p = .010) but not in the 6-bridge prior group (ρ = .24, p = .257); for inference probes, the pattern reversed, with a significant correlation in the 6-bridge prior group (ρ = .42, p = .035) but not the 4-cycle prior group (ρ = .16, p = .452). In neither probe type did a direct comparison of the two groups' correlation coefficients reach significance (experience probes: Z = -1.06, p = .291; inference probes: Z = .95, p = .340). Given that two of these four single-group correlations were nominally significant (without correction for multiple comparisons) while the corresponding between-group comparisons were not, and given the inconsistency in which group/probe-type combination shows the effect, we consider this pattern inconclusive rather than supportive of a clear general-ability signal linking performance on the two graph components.

We therefore cannot statistically rule out the general-ability account with the current sample, and have revised our interpretation accordingly: rather than claiming strict specificity to the 6-bridge structure, we now describe the correlation as selective to 6-bridge experience probes in this sample, while acknowledging that distinguishing structural reuse from a more general learning ability to represent novel relational structure remains an open question. We have added the decoding-strength comparison as a new Supplementary Figure 5, showing decoding-strength correlations with both 4-state and 6-bridge graph performance side by side, and report the direct behavioural correlation between graph components in the text, and have revised the relevant main-text passage to match these updated values.

We also wish to address why this correlation was selective to experience probes and did not extend to inference probes. We agree that, under an assumption of perfectly learned and context-independent abstract representations, one might expect performance on experience and inference probes to relate equally to the decoding signal. We believe the reviewer's concern is valid only under an assumption of perfect abstraction. In our task, abstract roles correspond to latent structural variables that are learned gradually and with uncertainty. Under such conditions, role inference is inherently context-dependent: the role of a given stimulus is not inferred in isolation, but in relation to the compound within which it appears. For example, if one element of a compound is already known with high confidence to belong to a particular subgraph, the role of the second element can be inferred more reliably; in contrast, when both elements are themselves uncertain, role inference becomes less precise. Critically, learning is unlikely to be uniform across compounds, as structural knowledge may be acquired with variable precision for different stimulus pairings depending on their frequency, co-occurrence statistics, and position within a sequence. This variability in learning precision means we would expect corresponding variability in a participant's ability to infer dynamical roles across different compound types, and consequently in their ability to make accurate predictions about those compounds. Under this account, our MEG signal, which indexes role-based representational structure specifically for experienced compounds, most directly reflects knowledge precision for those items; a correlation with inference probe accuracy requires an additional inferential step from experienced to novel compounds, introducing noise proportional to uncertainty in the underlying learned representations. We have revised the Discussion to make this reasoning more explicit.

(5) The authors frequently suggest that they are providing insight into temporal dynamics, but there is no mention of particular oscillations or particular temporal sequences of neural representation that support task performance.

We thank the reviewer for highlighting this. We agree our use of the term "temporal dynamics" is potentially misleading, and we have now revised the manuscript accordingly. Our analyses did set out to identify specific neural oscillations, rhythmic processes, or temporally ordered stages of computation that support task performance. Rather, time-resolved MEG analyses is a method that allows us to track when representational content is expressed during stimulus processing, serving three research questions about subprocess-specific representational similarity, dynamical role representations, and relationship to behavioral transfer performance; as opposed to characterizing the temporal organization of underlying computations per se.

We have now removed or replaced references to "temporal dynamics" and "neural dynamics" throughout the manuscript where these might be interpreted as claims about oscillatory or sequentially structured neural mechanisms. Instead, we now frame our conclusions in terms of representational content and its modulation by prior learning experience. We believe this aligns our language with what the data actually support and removes potential confusion regarding the scope of our claims.

(6) Regardless of point (b), can the authors provide more convincing evidence for a graph structure being represented per se (regardless of whether this representation is directly experienced or inferred)? From Figure 3C, it seems that the model RDM doesn't account for relative distance within the graph. Do they see evidence for distance coding? Can they reconstruct the graph from representational patterns using MDS?

We thank the reviewer for raising this point. The reviewer is correct that the model RDM shown in (previous) Figure 3C does not specifically capture relative distance within the graph and asks whether there is evidence for distance coding or whether the graph can be reconstructed using techniques such as MDS. We agree that distance-based representations are one possible signature of graph knowledge, with such effects reported in a prior literature. However, our theoretical focus is on an investigation of a more functionally relevant aspect of graph structure learning and reuse of abstracted knowledge: the representation of graph nodes according to their structural roles or positions within a graph.

To address the question more directly, we ran an additional sensor-level RSA analysis using the localizer task data, comparing representational changes from PRE to POST learning. Specifically, we examined whether stimuli in the 6-bridge graph that map onto the same structural roles (e.g., boundary, intermediate, central) within the relevant graph structures show greater neural similarity from before to after learning, and whether this effect depends on the graph structure experienced during prior learning. This new analysis revealed a cluster-based, permutation-corrected, interaction effect between prior learning group and graph factor, consistent with a representational similarity increase selectively for stimuli sharing the same structural role in a graph structure relevant to each group’s prior experience. Importantly, this effect occurs at similar time points as our earlier RSA effects based on overall stimulus-set similarity, suggesting that both analyses capture related aspects of the same underlying representational processes.

We believe this role-based representational similarity provides more direct evidence for graph-structured representations and their transfer across learning than distance-based or MDS reconstructions in our task context. First, distance within the graph is necessarily correlated with temporal proximity of experience, making it difficult to disentangle distance coding from local transition statistics. Second, multiple distinct graph structures used in our experiment (e.g., cycles, lines, and bridges) share overlapping distance relationships but differ in their higher-order structural roles. Our analysis therefore targets those aspects of structure that are invariant across surface features and that are most relevant for transfer across learning contexts.

For these reasons, while we do not claim we can reconstruct full graph geometry from neural patterns, our results are consistent with an organization of neural representations according to abstract structural positions within a graph, that is independent of specific stimulus identities. We have added a new panel to Figure 3 illustrating this role-based RSA effect and also expanded the Methods and Results sections to describe this analysis in detail. We believe this addition provides more convincing evidence that participants represent graph structure at an abstract level, even when that structure is not explicitly experienced in the same form during the localiser task. We thank the reviewer for encouraging this more specific analysis, as it significantly strengthens our claims.

“To test whether localiser representations of abstract graph structure generalised across sets experienced in the main task, we performed RSA separately for PRE and POST localisers. Structural roles were derived from each participant’s learning-phase transition sequences. For 4-state graphs, nodes were assigned to two matched role bins (boundary-like vs central-like) based on first-encounter order within each phase. For 6-state graphs, roles were determined either from node degree (6-line: boundary, intermediate, central). For 6-bridge graphs we used a predefined structural template. Specifically, the first eight valid transitions in the learning phase were mapped onto a fixed sequence of structural roles (central–intermediate–boundary–central–central–intermediate–boundary–central). Each node was assigned a role based on the majority role it occupied across its occurrences in these eight steps (ties resolved in favour of central > intermediate > boundary). For the localiser data, at each time point and sensor we estimated condition-specific activations using a GLM with 20 dummy-coded regressors (10 identities × 2 sets). This yielded sensor-wise beta patterns for each condition. Cross-set representational dissimilarities were computed using an L1 distance metric (mean absolute difference across sensors), resulting in 4×4 and 6×6 cross-set RDMs at each time point. We then summarised each RDM according to structural role: within-role distances (same role across sets), between-role distances (different roles), and their difference (delta = between − within). To control for unequal pair counts, between-role comparisons were subsampled to match the number of within-role pairs (200 repetitions). This produced time-resolved structural generalisation indices for 4-state and 6-state graphs, which were submitted to group-level statistical analyses as described above.”

(7) In general, the figures are not very clear, and the outcome from statistical tests is not graphically shown. The paper would be easier to digest if, for example, Figures 1-2 were made clearer and statistical significance relative to chance were indicated throughout. To give two examples: (i) Figure 1 should clearly indicate what is meant by observed and held-out transitions and whether it is just the transition or also the compound that is new to the participant. (ii) Figure 2D-E could be shown with relevant comparisons and simpler statistical comparisons. Currently, it is hard to follow without carefully reading the legend.

We have taken the reviewer’s point on board and have revised the figures to improve clarity. Specifically, we have improved labeling and visual annotation in Figure 1 to clearly distinguish observed versus held-out transitions and to highlight the key experimental manipulations. We have also simplified the presentation of behavioral results in Figure 2D–E and made the most relevant statistical comparisons more explicit, with statistical outcomes now indicated directly in the figures where appropriate.

We believe these changes improve readability and reduce the burden on the reader, while still relying on the figure captions to convey essential methodological details and interpretation, following standard practice. See Figure 1 and Figure 2.

Reviewer #1 (Recommendations for the authors):

(1) For the classifier: what was the overall accuracy of the classifier when tested on held-out examples of the same stimuli? Relative to chance? Relative to entirely different stimuli that have the same transition type (e.g., line 424)?

(2) In the Results paragraph 2 can you state how many subjects had prior exposure to 4-state vs how many had prior exposure to 6-state?

This has now been added these data to the results section, reproduced below for convenience:

In brief, participants “(N = 50, N = 25 subjects per condition)” performed a sequence learning task, observing image sequences produced by walks on product graphs constructed from smaller graph factors (Fig. 1A and B).

Reviewer #2 (Public review):

Summary:

The authors aimed to investigate the temporal dynamics of how prior experiences shape learning in new complex environments by examining whether the brain reuses abstract structural components from those experiences. They employed a sequence learning task based on graph factorization and recorded neural activity using magnetoencephalography (MEG) to investigate how the underlying graph factors are reused to support learning and inference in a new graph. MEG data was derived from passive stimulus presentation trials, and behavior was assessed through a small number of probe trials testing either experienced or inferred successions in the graph. Representational similarity analysis of the MEG data was performed at a quite aggregated level (the principal components explaining 80% of the variance). The authors report (1) enhanced neural similarity among stimuli that belong to the same graph-factor as well as (2) a correlation between abstract role representations, corresponding to particular positions in the graph, and performance in experience-probes but not in inference-probes.

Overall, the findings are congruent with prior research on neural correlates of structural abstraction. They offer an elegant, well-suited task design to study compositional representations, replicating the authors' earlier finding and providing temporal information on structural generalisation in a sequence learning task.

We thank the reviewer for this positive evaluation of our work and insightful comments and suggestions.

(1) The first finding is considered evidence for representational alignment of the graph factors. However, alignment seems to be just one possible arrangement underlying the increased similarity between stimuli of the same vs different graph factors. For instance, a simple categorical grouping of stimuli belonging to the same graph, rather than their structural alignment, could also underlie the reported effect. The wording should be adjusted to avoid over interpretation.

We agree with the reviewer that increased neural similarity between stimuli associated with the same graph factor does not uniquely imply representational alignment in its strongest sense, and could in principle also reflect a more general categorical grouping of stimuli based solely on belonging to the same graph. We revised the manuscript to tone down the wording and avoid overinterpretation, replacing references to “representational alignment” with more neutral language that refers to increased within-factor representational similarity.

At the same time, we believe that additional analyses introduced in the revised manuscript constrain alternative accounts. In particular, our new analysis on localiser data, examining representational changes from PRE to POST learning for stimuli that map onto the same structural roles within relevant sub-process graph structures. This analysis showed the presence of a cluster-based, permutation-corrected, interaction effect between group and graph factor, indicating that representational similarity increases selectively for stimuli sharing the same structural role, rather than simply increased similarity for all stimuli associated with the same graph.

Because these role-based groupings cut across specific stimulus identities and cannot be reduced to a single categorical label for an entire graph, we believe this result provides specific evidence that neural representations are shaped by internal structural properties of the graph, rather than reflecting the emergence of a coarse categorical distinction between graph types. We have added this analysis as a new panel to Figure 3 and expanded the Results and Methods sections accordingly (reproduced below for convenience):

“The observed increase in neural similarity between stimuli associated with the same graph factor could, in principle, reflect a more general categorical grouping of stimuli belonging to the same graph factor, rather than reflecting an abstract graph representation. To address this alternative account, we examined whether neural similarity in the localizer task is specifically structured according to the graph-theoretic role each stimulus occupies within its graph, using a finer-grained theoretical RDM that distinguishes between boundary, intermediate, and central positions. Structural roles were derived from each participant's learning task sequences and defined by graph topology. For Component 6, roles were assigned using a bridge-graph traversal template (for Condition 6, whose prior learning involved the bridge graph) or degree-based assignment (for Condition 4, whose prior learning involved a line graph): boundary nodes were defined as nodes with degree 1 (graph periphery); intermediate nodes as neighbours of boundary nodes; and central nodes as the remaining, more densely connected nodes. For Component 4, two roles were distinguished, boundary-like nodes (those encountered first and last in the sequence traversal) and central-like nodes (those encountered second and third), reflecting structural position within the graph rather than arbitrary temporal sequence. For further details of role assignment, see Methods. This analysis revealed a cluster-based permutation-corrected interaction (P_FEW < 0.05, non-parametric permutation test, Figure 3E) between prior learning group and graph factor at overlapping time windows with the previously reported PCA and sensor-level analyses investigating feature-mappings to abstract graph factors. We did not observe significant and multiple comparisons corrected prior learning group and graph factor main effects (all P_FEW > 0.90, non-parametric permutation test). These results indicate that representational similarity increased selectively for stimuli sharing the same structural role in the graph structure relevant to each group’s prior experience. Critically, this effect was not observed uniformly for all stimuli associated with the same graph, arguing against a purely categorical grouping explanation. Instead, these findings provide evidence that learning induces representations that reflect higher-order relational structure, capturing graph-theoretic roles that are invariant across surface features of the embedding stimuli and support transfer across learning contexts.

Taken together, our findings indicate a task-induced, learning-related increase in neural similarity for stimuli sharing a structural role in subprocesses common across task phases and is compatible with a structural scaffolding mechanism, though we do not claim these representations are fully context-independent.

“To test whether localizer representations of abstract graph structure generalised across sets experienced in the main task, we performed RSA separately for PRE and POST localisers. Structural roles were derived from each participant’s learning-phase transition sequences. For 4-state graphs, nodes were assigned to two matched role bins (boundary-like vs central-like) based on first-encounter order within each phase. For 6-state graphs, roles were determined either from node degree (6-line: boundary, intermediate, central). For 6-bridge graphs we used a predefined structural template. Specifically, the first eight valid transitions in the learning phase were mapped onto a fixed sequence of structural roles (central–intermediate–boundary–central–central–intermediate–boundary–central). Each node was assigned a role based on the majority role it occupied across its occurrences in these eight steps (ties resolved in favour of central > intermediate > boundary). For the localiser data, at each time point and sensor we estimated condition-specific activations using a GLM with 20 dummy-coded regressors (10 identities × 2 sets). This yielded sensor-wise beta patterns for each condition. Cross-set representational dissimilarities were computed using an L1 distance metric (mean absolute difference across sensors), resulting in 4×4 and 6×6 cross-set RDMs at each time point. We then summarised each RDM according to structural role: within-role distances (same role across sets), between-role distances (different roles), and their difference (delta = between − within). To control for unequal pair counts, between-role comparisons were subsampled to match the number of within-role pairs (200 repetitions). This produced time-resolved structural generalisation indices for 4-state and 6-state graphs, which were submitted to group-level statistical analyses as described above.”

(2) The second finding of abstract role representations is indeed expected for structural generalisation. While the data presents an interesting indication, its interpretability is constrained by a lack of testing for generalization of the effect to other graph structures (e.g., to rule out graph-specific strategies) as well as the absence of a link to transfer performance in inference-probes. The authors argue that the experienced transitions the classifier was trained on might be more similar in process to the experience-probes than the inference-probes. However, as inference-probes are the key measure of transfer, one could argue that if abstract role representations truly underlie transfer learning, they should be evident in the common neural signal.

The reviewer provides a thoughtful comment and we agree that, under an assumption of perfectly learned and context-independent abstract representations, one might expect this would be related equally to performance in experience and inference probes. Below, we clarify the scope of our claims, including presenting additional neural evidence for abstract role representations, and argue that an absence of a strong neural-behavioral correlation for inference probes is not inconsistent with our proposed abstract structural knowledge account of transfer learning.

First, we acknowledge that our experiment tests generalization only across two graph families, rather than across a broader set of graph structures. While we think it unlikely that our findings are idiosyncratic to cycles or bridge-like graphs, a we now limit our claims and emphasise a future need for testing generalization across additional graph families (see discussion section as reproduced below)

“We concede that the present experiment examines generalization across two families of graph structures, rather than across a broader space of possible graphs. While we consider it unlikely that our findings are idiosyncratic to cycles, lines and bridge-like graphs, we acknowledge that further work will be needed to support abstraction beyond the limited graph structures studied here.”

Second, to provide more direct neural evidence that abstract structural roles within the studied graphs are represented per se, we add a new analysis that uses independent localizer task data. In this analysis, we performed a sensor-level RSA contrasting pre- and post-learning representations. Specifically, we test whether stimuli that occupied the same role within learned graph substructures became more neurally similar after learning. This localizer RSA revealed a cluster-corrected interaction effect between prior learning group and graph factor, expressed in an increased representational similarity for stimuli mapping onto equivalent structural roles, independent of specific stimulus identities and outside the task context of the main learning phase. This result supports the proposal that participants form abstract, role-based, representations of graph structure where these representations are not tied to a particular transition sequence or probe type. This finding adds an important insight in that it goes beyond establishing role representations for the 6-bridge graph factor, enabling us to make claims about generalization of structural knowledge of the 4-cycle graph factor in the 4-cycle prior condition (thereby to some extent showing a wider generalization across more structural forms).

“To test whether localizer representations of abstract graph structure generalised across sets experienced in the main task, we performed RSA separately for PRE and POST localisers. Structural roles were derived from each participant’s learning-phase transition sequences. For 4-state graphs, nodes were assigned to two matched role bins (boundary-like vs central-like) based on first-encounter order within each phase. For 6-state graphs, roles were determined either from node degree (6-line: boundary, intermediate, central). For 6-bridge graphs we used a predefined structural template. Specifically, the first eight valid transitions in the learning phase were mapped onto a fixed sequence of structural roles (central–intermediate–boundary–central–central–intermediate–boundary–central). Each node was assigned a role based on the majority role it occupied across its occurrences in these eight steps (ties resolved in favour of central > intermediate > boundary). For the localiser data, at each time point and sensor we estimated condition-specific activations using a GLM with 20 dummy-coded regressors (10 identities × 2 sets). This yielded sensor-wise beta patterns for each condition. Cross-set representational dissimilarities were computed using an L1 distance metric (mean absolute difference across sensors), resulting in 4×4 and 6×6 cross-set RDMs at each time point. We then summarised each RDM according to structural role: within-role distances (same role across sets), between-role distances (different roles), and their difference (delta = between − within). To control for unequal pair counts, between-role comparisons were subsampled to match the number of within-role pairs (200 repetitions). This produced time-resolved structural generalisation indices for 4-state and 6-state graphs, which were submitted to group-level statistical analyses as described above.”

Third, in relation to the relationship between neural role representations and transfer performance during inference probes, we believe the reviewer's concern is valid only under an assumption of perfect abstraction. In our task, abstract roles correspond to latent structural variables that are learned gradually, and with uncertainty. Under such conditions, role inference is inherently context-dependent: the role of a given stimulus is not inferred in isolation, but in relation to the compound within which it appears. For example, if one element of a compound is already known with high confidence to belong to a particular subgraph, the role of the second element can be inferred more reliably. In contrast, when both elements are themselves uncertain, role inference becomes less precise. Critically, learning is unlikely to be uniform across compounds, as structural knowledge may be acquired with variable precision for different stimulus pairings depending on their frequency, co-occurrence statistics, and position within a sequence. This variability in learning precision means we would expect corresponding variability in a participant's ability to infer dynamical roles across different compound types, and consequently in their ability to make accurate predictions about those compounds.

Our MEG signal indexes role-based representational structure specifically for experienced compounds, and most directly reflects the precision of structural knowledge for those items. Therefore a correlation between a MEG signal and accuracy on inference probes is expected to be weaker, since it requires an additional inferential step from experienced to novel compounds, a step that is likely to introduce additional noise proportional to the uncertainty in the learned representations. Ideally we would expect such a correlation to be present for inference probes as well, since better structural knowledge about experienced compounds should in principle support better generalisation to novel ones. However, we consider we are insufficiently powered to detect this weaker relationship and interpret an absence of a significant correlation for inference probes as not inconsistent with a structural scaffolding account.

Finally, we emphasize that the absence of a strong correlation between aggregate neural decoding measures and aggregate inference-probe performance should not be taken as evidence against abstract learning. Inference performance is likely to be supported by multiple sources of information (including partial structural knowledge and heuristic strategies), whereas neural decoding captures the stability of representational geometry at a population level. The newly added localizer RSA results demonstrate that abstract role representations are present in the neural signal, while at the same time their task-dependent expression may not always be directly mirrored at the level of inference-probe behavior.

We have an added a paragraph caveating our findings to the discussion section (reproduced below for convenience):

“Addressing our third research question, whether the strength of identified neural representations relate to behavioral transfer, we found a positive relationship between decoding strength and experience probe performance that was numerically larger for the structurally relevant (6-bridge) graph than for the irrelevant (4-cycle) graph, though this difference did not reach significance; no corresponding relationship was evident for inference probes.” Thus, a behavioral accuracy group difference in predicting successive states for the 6-bridge inference probe performance was of negligible magnitude and did not align with the robust effect observed for experience probes (Fig. 2D), where the latter replicates and extends a previous finding using a 6-state cycle graph and predicted by computational simulations using successor feature models (Luettgau et al., 2024). We speculate that the disparity between probe types is likely due to experience probes and inference probes engaging dissociable cognitive processes – the former relying on predictive representations acquired from experience, the latter relying on an ability to perform mental operations on learned representations of independent feature transitions. This may increase the likelihood of errors due to complexities of the inferred state spaces. However, the observed differences could also relate to previously reported effects of explicit knowledge on performance levels (Luettgau et al., 2024) or be driven by capacity limitations in working memory taxing larger subprocesses more strongly (Collins et al., 2017; Collins & Frank, 2012). “It is also likely that structural knowledge is not acquired uniformly across compounds. Because abstract roles are learned gradually and with uncertainty, role inference is inherently context-dependent: the role of a given stimulus is not inferred in isolation, but in relation to the compound within which it appears. When one element of a compound is already known with high confidence, the role of the second element can be inferred more reliably; when both elements are themselves uncertain, inference becomes less precise. Learning precision may therefore vary across stimulus pairings depending on their frequency, co-occurrence statistics, and position within a sequence, and we would expect corresponding variability in participants' ability to infer dynamical roles, and hence to predict outcomes, across different compound types. This means that an MEG signal, which indexes role-based representational structure specifically for experienced compounds, most directly reflects knowledge precision for those items. A correlation with inference-probe accuracy requires an additional inferential step, from experienced to novel compounds, introducing noise proportional to the uncertainty in the learned representations. In principle, better structural knowledge about experienced compounds should support better generalisation to novel ones, but we might be insufficiently powered to detect this weaker relationship. We therefore interpret an absence of a significant correlation for inference probes not as inconsistent with a structural scaffolding account, and suggest that a null result here should be interpreted with caution rather than as evidence against it. More broadly, an absence of a strong correlation between aggregate neural decoding measures and aggregate inference-probe performance should not itself be taken as evidence against abstract learning. Inference performance is likely supported by multiple sources of information, including partial structural knowledge and heuristic strategies, whereas neural decoding captures the stability of representational geometry at the population level. The localizer RSA results reported above demonstrate that abstract role representations are present in the neural signal, even though their task-dependent expression may not always be directly mirrored in inference-probe behavior.”

(3) The authors write, "we observed a qualitative pattern indicative of increased neural similarity between stimuli that adhered to the same underlying subprocess across task phases. (...) There was a statistically significant interaction effect of condition x graph factor spanning approximately 300 - 680 ms post-stimulus onset". I conclude there was no significant main effect of graph factor, but the relevant statistics are not reported. The authors should report and discuss the complete statistics.

Thank you for this comment. The reviewer is correct in their conclusion that the main effects were not significant and we agree the main effect of graph factor and group (as well as their significance) should be reported. We now do so in the figure caption of Figure 3D, as well as in the results section. The section reads as follows:

We did not observe prior learning group and graph factor main effects (all PFWE > 0.118, non-parametric permutation test).

(4) The RSA is performed on highly aggregated data (the PCs that explained 80% of the variance). Could the authors include their rationale for this choice (e.g. over-analysis of sensor-level data)? In case sensor-level analyses have been conducted as well, maybe there are comparisons or implications of the chosen approach that are useful to mention in the discussion. The authors should provide the average and distribution of the number of PCs underlying their analyses.

We thank the reviewer for raising this important methodological concern. We agree that the use of PCA-based dimensionality reduction in RSA requires further justification. Our primary motivation for using PCA, and retaining components that jointly explain 80% of the variance, was based on denoising rather than aggregation per se. Sensor-level MEG data are high-dimensional and highly correlated, and representational analyses that operate directly on the full sensor space may be susceptible to noise amplification and overfitting, particularly when estimating distance measures across time. Retaining components up to a fixed variance threshold is a widely used approach in MEG decoding and RSA to preserve the dominant signal structure while attenuating sensor-specific noise. Importantly, this procedure discards only the lowest-variance 20% of the signal and on this basis does not impose strong constraints on the representational geometry.

To address the reviewer’s concern more directly, and to ensure our findings are not driven by the PCA procedure itself, we conducted additional robustness checks. Specifically, we performed sensor-level RSA analyses without PCA, using multiple distance metrics (including L1 and L2 distances). These sensor-level analyses yielded highly similar temporal profiles and statistical effects to the PCA-based results, including for the key contrasts reported in the main text. We now include these results in a new supplementary figure and reference them explicitly in the Results section.

In the revised manuscript, we clarify that PCA was used as a principled denoising step, not as a means of imposing structure on the data, and we explicitly show that our conclusions are robust across analytic choices.

“Importantly, to verify that the chosen PCA approach did not impose artificial structure within the data, or introduce other potential confounds, we also conducted the above analyses on sensor level MEG data. Highly similar results were obtained, with consistent results for different similarity metrics (absolute distance (L1 norm) and Euclidean distance (L2 norm), Supplementary Figure 3)), demonstrating that our results and conclusions are robust across analytic choices.”

Methods

“In addition to the PCA-based feature analysis, we performed a sensor-level RSA. At each time point, we estimated condition-specific activation patterns using a GLM applied separately to each sensor (dummy-coded stimulus identity). RDMs were then computed across stimuli using mean absolute difference (L1 norm) as the primary distance metric. For completeness, we additionally evaluated Euclidean distances (L2 norm), which yielded qualitatively similar results (Supplementary Figure 3).”

(5) While the paper is well-written overall, it would benefit from more explicitly identifying the concrete research question and advancing through the results. The authors state their aim as understanding the "temporal dynamics of compositional generalisation", revealing "at which moment during neural information processing are they assembled". They conclude with "providing evidence for temporally resolved neural dynamics that support compositional generalization" and "we show the neural dynamics (...) presented across different task phases...". It remains somewhat vague what specific insight about the process is provided through the temporal resolution (e.g., is the time window itself meaningful, if so, it should be contextualized; is the temporal resolution critical to dissociate subprocesses). The different task phases -initial learning and transfer- are the necessary conditions to investigate transfer learning, but do not by themselves offer a particularly resolved depiction of the process.

We thank the reviewer for these thoughtful and constructive comments. We agree that in the original version of the manuscript our framing did not sufficiently specify the concrete insight that might be gained from neural data, nor did it clearly delimit the scope of our claims. Our primary research question did not relate to characterizing fine-grained temporal mechanisms or sequential subprocesses underlying compositional generalization (e.g., oscillatory dynamics, replay, or temporally ordered stages of computation). Rather, our goal is to determine whether abstract compositional structural representations are detectable neurally during stimulus processing, and how these representations are shaped by prior learning experience.

We address three specific research questions: whether subprocess-specific representational similarity emerges following learning that is specific to a subprocess experienced during prior learning; whether fine-grained dynamical role representations are acquired during stimulus processing; and whether their strength relates to a behavioral transfer success. Importantly, the time-resolved nature of the analyses is a feature of the method that allows us to track the expression of representational content at a sub-trial timescale, rather than an end in itself. We agree with the reviewer that the temporal resolution does not, in itself, resolve distinct computational subprocesses, nor does it provide a mechanistic decomposition of compositional operations. Accordingly, we have revised the manuscript to more explicitly articulate these three concrete research questions in the Introduction, as well as clarify that the reported time windows should be interpreted as markers of representational availability rather than discrete processing stages. We also tone down any language that might be interpreted as making claims about temporally structured mechanisms. Finally, we agree that the distinction between task phases (prior learning vs. transfer) is not intended to depict a detailed temporal process, but rather to provide the necessary contrast to test whether neural representations reflect abstract structural knowledge that goes beyond direct experience. We have revised the Results and Discussion to make this logic more explicit and to better guide the reader through how each analysis contributes to addressing the three research questions.

We believe these revisions better align our claims with the level of insight the data support and make the contribution of each analysis more legible to the reader.

Reviewer #2 (Recommendations for the authors):

(1) Could the authors explain in the manuscript how the held-out transitions were chosen?

We agree that the stimulus construction and sequence generation logic were not described with sufficient clarity. In the revised manuscript we add an explicit description of how compound images were constructed from two independently varying visual features, each corresponding to one graph factor, and introduced the distinction between directly experienced, implied-but-unobserved, and never-valid transitions earlier in the text, so that the probe logic is easier to follow from the outset.

We reproduce parts of the revised methods section describing the held-out stimulus set below:

“[...] Because sequences were drawn from a subset of the product graph, only a portion of all 24 possible compound images (and their transitions) were encountered directly. Transitions between compound images could thus be: (i) directly experienced in a sequence, (ii) never presented in a sequence but implied by the factorization structure, or (iii) transitions between compound images from different subgraphs, which were never valid. The held-out set of transitions described in (ii) and (iii) between compound stimuli (partially observed during sequence presentations, partially never observed during sequence presentations) provided a context for assessing knowledge of graph factorization.”

(2) It might be easier to follow the results section (without going back to Figure 1) if the mentioning of specific graph factors (6-bridge, 4 cycle) would occasionally be accompanied by their function/condition in the experiment.

Thanks for this suggestion, we have implemented this in the manuscript.

Reviewer #3 (Public review):

Summary:

This study investigates how task components can be learned and transferred across different task contexts. The authors designed two consecutive sequence learning tasks, in which complex image sequences were generated from the combination of two graph-based structural "building blocks". One of these components was shared between the prior and transfer task environments, allowing the authors to test compositional transfer. Behavioral analyses using generalized linear models (GLMs) assessed participants' sensitivity to the underlying structure. MEG data were recorded and analyzed using classifications and feature representational similarity analysis (RSA) to examine whether neural similarity increased for stimuli sharing the same relational structure. The paper aims to uncover the neural dynamics that support compositional transfer during learning.

I found the methods and task design of this paper difficult to follow, particularly the way stimuli were constructed and how the experimental sequences were generated from the graph structures. These aspects would be hard to replicate without some clarification. I appreciate the integration of behavioral and neuroimaging data. The overall approach, especially the use of compositional graph structures in sequence learning, is interesting and could be used and revised in further studies in compositionality and transfer learning. I appreciated the authors' careful interpretation of their findings in the discussion. However, I would have liked a similar level of caution in the abstract, which currently overstates some claims.

We thank the reviewer for a positive evaluation coupled with helpful suggestions to improve the clarity of our presentation. We are particularly pleased the reviewer found the integration of behavioral and neuroimaging data, the compositional graph structure approach, and our interpretation of findings in the discussion to be particular strengths of the manuscript.

We address the two specific concerns raised below.

We agree that the stimulus construction and sequence generation logic were not described with sufficient clarity. In the revised manuscript we have: (i) added an explicit description of how compound images were constructed from two independently varying visual features, each corresponding to a single graph factor; (ii) defined the Cartesian product graph operationally, clarifying also that transitions between compound images are valid when one factor changes state while the other remains fixed; (iii) included a worked example tracing a sample sequence through both factors; and (iv) introduced the distinction between directly experienced, implied-but-unobserved, and never-valid transitions earlier in the text, so that the probe logic is easier to follow from the outset.

We reproduce parts of the revised methods section describing stimulus construction below:

“Compound images were constructed by overlaying two independently varying visual features drawn from separate feature sets. Each feature set corresponded to a single graph factor, such that the state of each graph factor independently determines one feature of the compound image. For example, a compound image in state (A, 3) would display feature A from the cyclic factor and feature 3 from the path factor, superimposed or juxtaposed into a single image.

The set of all compound images and their transitions was defined by the Cartesian product of two graph factors: transitions between compound images were valid if, and only if, one factor alone changed state according to its own transition rules while the other factor remained fixed.

Observed sequences of image compositions composed from 10 individual features. Each sequence consisted of 8 compound images with 12 compound images in total across sequences.

[...] For instance, in condition 1, a valid sequence might proceed: (A,1) → (A,2) → (A,3) → (B,3) → (C,3) → (C,4) → (D,4) → (D,5), where each step reflects a transition in either the cyclic (letter) or path (number) factor, but never both simultaneously.

[...] Because sequences were drawn from a subset of the product graph, only a portion of all 24 possible compound images (and their transitions) were encountered directly. Transitions between compound images could thus be: (i) directly experienced in a sequence, (ii) never presented in a sequence but implied by the factorization structure, or (iii) transitions between compound images from different subgraphs, which were never valid. The held-out set of transitions described in (ii) and (iii) between compound stimuli (partially observed during sequence presentations, partially never observed during sequence presentations) provided a context for assessing knowledge of graph factorization.”

We thank the reviewer for flagging the mismatch between the strength of interpretations made in the abstract and discussion section. We have revised the abstract to bring the language in line with the level of caution exercised in the discussion, avoiding overstated interpretations of the findings.

The revised abstract is reproduced below:

“A human ability to adapt to the dynamics of novel environments relies on abstracting and generalizing from past experiences. Previous research has focused on how humans generalize from isolated sequential processes, yet we know little about mechanisms that enable adaptation to more complex dynamics, including those that govern much everyday experience. Here, using a novel sequence learning task based on graph factorization, coupled with simultaneous magnetoencephalography (MEG) recordings, we ask whether reuse of experiential "building blocks" enable inference and generalization. Behavioral results were consistent with participants decomposing task experience into sub processes, abstracting their dynamical structure away from their sensory specifics and transferring these to a new task environment. Neurally, this transfer was associated with a learning-induced, condition-specific, increase in neural similarity among stimuli conditional on their structural role in abstract sub processes that were shared across task phases. Consistent with this, a complementary dynamical role decoding analysis revealed enhanced neural decodability for stimuli sharing identical structural transition types across experiential contexts, specifically when prior experience included the relevant graph structure. Decoding strength for these role representations was positively related to behavioral success in sub process knowledge transfer, a relationship that will require future independent replication. These findings provide neural evidence consistent with the idea that a structural scaffolding mechanism supports generalization of experience to new contexts.”

(1) While the introduction mentions brain areas implicated in the low-dimensional representation of task knowledge, the current study uses M/EEG and does not include source reconstruction. As a result, the focus is primarily on the temporal dynamics of the signal rather than its spatial origins. Although I am not suggesting that the authors should perform source reconstruction in this study, it would strengthen the paper to introduce the broader M/EEG literature on task-relevant representations and transfer. The same applies to behavioral studies looking at structural similarities and transfer learning. I encourage the authors to integrate relevant literature to better contextualize their results.

Duan, Y., Zhan, J., Gross, J., Ince, R. A. & Schyns, P. G. Pre-frontal cortex guides dimension-reducing transformations in the occipito-ventral pathway for categorization behaviors. Current Biology 34, 3392-3404 (2024).

Luyckx, F., Nili, H., Spitzer, B. & Summerfield, C. Neural structure mapping in human probabilistic reward learning. eLife 8, e42816 (2019). (This is in the references but not in the text).

Zhang, M. & Yu, Q. The representation of abstract goals in working memory is supported by task-congruent neural geometry. PLoS biology 22, e3002461 (2024).

L. Teichmann, T. Grootswagers, T. Carlson, A.N. Rich Decoding digits and dice with magnetoencephalography: evidence for a shared representation of magnitude Journal of cognitive neuroscience, 30 (7) (2018), pp. 999-1010

Garner, K., Lynch, C. R. & Dux, P. E. Transfer of training benefits requires rules we cannot see (or hear). Journal of Experimental Psychology: Human Perception and Performance 42, 1148 (2016).

Holton, E., Braun, L., Thompson, J., Grohn, J. & Summerfield, C. Humans and neural networks show similar patterns of transfer and interference during continual learning (2025).

We thank the reviewer for providing these helpful suggestions. We have taken these references on board and now introduce these findings in the introduction of the manuscript (reproduced below for convenience):

“In parallel, M/EEG studies have shown that abstract, format-independent representations can emerge dynamically over time. For example, multivariate analyses of MEG data demonstrate cross-decoding of numerical magnitude across distinct symbolic formats (e.g., digits and dice), indicating a shared, format-invariant representation that unfolds with distinct temporal dynamics (Teichmann et al., 2018). More broadly, M/EEG evidence suggests that task-relevant information can be dynamically prioritized, compressed, and reformatted, indicative of a rapid temporal evolution of abstract and behaviorally relevant representations even when spatial localization remains coarse. Recent MEG data suggests that high-dimensional sensory inputs are progressively transformed into lower-dimensional, task-relevant, feature manifolds across distinct processing stages, reflecting dynamic gating and reformatting of information in support of flexible categorization (Duan et al., 2024). Complementary EEG–fMRI evidence also indicate abstract task goals are encoded in a low-dimensional neural geometry that emerges in frontal cortex and communicated to posterior regions to guide implementation of stimulus-specific representations, with the fidelity of this geometry predicting behavioral performance (Zhang & Yu, 2024). Likewise, converging behavioral and computational data indicates successful generalization depends on detecting and encoding abstract relational structure rather than stimulus surface features. In effect, training enables transfer when task phases share an underlying rule or abstract relationship, even across modalities, but not when an overlap is purely perceptual (Garner et al., 2016).”

“At the same time, reuse of shared abstract representations introduces a computational trade-off, such that overlapping codes that facilitate transfer can also induce interference, as observed in both humans and artificial neural networks (Holton et al., 2025). However, while there is now much work elucidating how learned representations are generalized, it remains unclear whether and how abstracted structural representations are expressed neurally during ongoing stimulus processing, and how they are shaped by prior structural experience. Thus there is a need to characterize not only where abstract representations are located in the brain but also how they dynamically unfold and are reused across contexts in support of compositional generalization.”

(2) I found it interesting that the authors chose to perform PCA for dimensionality reduction prior to conducting RSA; however, I haven't seen such an approach in the literature before. It would be helpful to either cite prior studies that have employed a similar method or to include a comparison with more standard approaches, such as sensor-level RSA or sensor-searchlight analysis.

(3) Connected to the previous point, the choice to use absolute distance as a dissimilarity measure is not justified. How does it compare to standard metrics such as correlation distance or Mahalanobis distance? The same applies to the use of Kendall's tau.

We thank the reviewer for raising these methodological concerns and agree methodological choices should be explicitly motivated and justified. Our primary motivation for using PCA, and retaining components that jointly explain 80% of the variance, was denoising. Sensor-level MEG data are high-dimensional and highly correlated, and representational analyses that operate directly on the full sensor space may be susceptible to noise amplification and overfitting, particularly when estimating distance measures across time. Retaining components up to a fixed variance threshold is a widely used approach, both in MEG decoding and RSA, to preserve the dominant signal structure while attenuating sensor-specific noise. Importantly, this procedure discards only the lowest-variance 20% of the signal and therefore does not impose strong constraints on the representational geometry.

To address the reviewer’s concern regarding methodological choices and what this means for the conclusions drawn more directly, and to ensure our findings are not driven by specific analytic choices, we conducted additional robustness checks, availing of more “standard” approaches. Specifically, we performed the same RSA on localizer tasks but using sensor-level data without PCA, employing multiple distance metrics (including absolute distance/L1 norm and Euclidean distance/L2 norm). These similarity metrics are commonplace in RSA, providing linear, per-dimension differences between neural patterns (L1) or robust small outlier differences along individual dimensions compared to squared differences (Euclidean/L2), providing simple and interpretable similarity metrics.

These sensor-level analyses yielded highly similar temporal profiles and statistical effects to the PCA-based results, including the key contrasts reported in the main text. We now include these results in a new Supplementary Figure and reference them explicitly in the Results section.

In the revised manuscript, we clarify that PCA was used as a principled denoising step, not as a means of imposing structure on the data, and explicitly show that our conclusions are robust across analytic choices.

“Importantly, to verify that the chosen PCA approach did not impose artificial structure within the data, or introduce other potential confounds, we also conducted the above analyses on sensor level MEG data. Highly similar results were obtained, with consistent results for different similarity metrics (absolute distance (L1 norm) and Euclidean distance (L2 norm), Supplementary Figure 3)), demonstrating that our results and conclusions are robust across analytic choices.”

Methods

“In addition to the PCA-based feature analysis, we performed a sensor-level RSA. At each time point, we estimated condition-specific activation patterns using a GLM applied separately to each sensor (dummy-coded stimulus identity). RDMs were then computed across stimuli using mean absolute difference (L1 norm) as the primary distance metric. For completeness, we additionally evaluated Euclidean distances (L2 norm), which yielded qualitatively similar results (Supplementary Figure 3).”

Our choice of Kendall’s tau rank correlation was motivated by its common use for linking neural and theoretical RDMs (e.g., Luyckx et al., 2019 eLife; Nili et al., 2014 PLoS CB). We have now added these citations to the manuscript.

(5) The analysis described in the "Abstract representation of dynamical roles in subprocesses" does not appear to convincingly test the stated prediction of a structural scaffolding account. The authors hypothesize that if structure and dynamics from prior experiences are repurposed, then stimuli occupying the same "dynamical roles" across different sequences should exhibit enhanced neural similarity. However, the analysis seems to focus on decoding transitions rather than directly assessing representational similarity. Rather, this approach may reflect shared temporal representation in the sequences without necessarily indicating that the neural system generalizes the abstract function or position of a stimulus within the graph. To truly demonstrate that the brain captures the dynamical role across different stimuli, it would be more appropriate to directly assess whether neural patterns evoked by stimuli, in the same temporal part of the sequence, with shared roles (but different visual identities) are more similar to each other than to those from different roles.

We thank the reviewer for this careful and constructive critique and analytical suggestion. The reviewer is correct that in the bridge graph traversal, structural roles map onto fixed serial positions within a sequence, meaning that same-role stimulus pairs also tend to share the same temporal position in the sequence. This is a potential confound: neural similarity between same-role stimuli could in principle reflect similarity related to time-in-sequence rather than an abstract structural role per se.

The original role-based RSA does not fully resolve this concern, and we have now conducted, based on the reviewer’s suggestion, an additional analysis that directly addresses it. Rather than comparing stimuli that share a structural role, we examine whether the neural representation of a single central node stimulus is similar across its two distinct serial positions within the prior and transfer learning sequences. Specifically, in the bridge graph traversal, each central node appears twice per sequence: the left central node occurs at steps 1 and 4, the right central node occurs at steps 5 and 8. We estimate neural patterns for each central node separately at its first and second occurrence, using occurrence-specific GLMs, and then compute cross-phase similarity between prior and transfer learning: the neural pattern evoked by a central node at its first occurrence in prior learning is compared to its pattern at the second occurrence in transfer learning, and vice versa.

For the 6-bridge prior group, whose prior learning also involved a bridge graph, central nodes were identified using the same bridge-template positions (steps 1/4 and 5/8) in both prior and transfer learning phases. For the 4-cycle prior group, whose prior learning involved a line graph rather than a bridge graph, central nodes in the prior phase were identified using degree-based role assignment (the two nodes with the highest degree that are not boundary or intermediate nodes), ordered by first appearance in the prior trial sequence. Transfer learning central nodes were identified from the bridge graph template in both groups, since transfer learning always involved a bridge graph. The group comparison thus tests whether cross-phase similarity between central node representations is specific to participants whose prior experience included a bridge graph relative to participants whose prior experience involved a structurally different graph, for whom "central" nodes in prior and transfer are central in qualitatively different senses. This approach directly controls for the time-in-sequence confound, since the same node is compared across different serial positions rather than comparing different stimuli that share a role and therefore also share a sequence position.

We compared this cross-phase similarity between the 6-bridge prior and 4-cycle prior groups. We observe a significant group difference at a late stage of the trial (approximately 1.125–1.25 sec post-stimulus onset, P^FEW = 0.043, two-sample non-parametric permutation test, one-tailed), with additional differences at earlier time points that did not survive correction for multiple comparisons (0.125 sec: t = 2.18, p = .034; 0.79 sec: t = 2.84, p = .007). This indicates increased neural similarity for central node representations across phases in participants who experienced these roles consistently during prior learning.

We wish to be entirely transparent, however, about two limitations of this additional analysis with respect to the original dynamical role decoding results reported in Figure 4. First, the temporal profile of the cross-phase similarity effect does not perfectly overlap with the time windows identified in the original decoding analysis, which limits the strength of the convergence argument. We therefore cannot fully rule out a possibility that the earlier decoding effects partly reflect temporal position within the sequence rather than abstract structural role coding per se. Second, this analysis is necessarily based on a smaller and less variable set of comparisons than either the original decoding analysis or the role-based RSA, since the bridge graph contains only two central nodes whose repeated occurrences can be directly compared in this way. This limits the statistical power of the analysis and our ability to draw strong conclusions from it in isolation. We acknowledge both limitations explicitly in the revised manuscript. Nonetheless, the observation of a group-specific effect in the cross-phase analysis, which by design controls for serial position, provides some support for the structural scaffolding account, and the late corrected cluster suggests that role-consistent representations do emerge across learning phases.

In addition to this online analysis, we conducted a parallel role-based RSA on data from the localizer task. This localizer-based analysis examines representational changes from PRE to POST learning for stimuli that share the same structural role within the relevant sub-process graph structures. We find a cluster-based permutation-corrected interaction effect between prior learning group and graph factor, indicating that representational similarity increases selectively for stimuli sharing the same structural role, rather than uniformly for all stimuli associated with a given graph. Together with the localizer-based RSA results, we believe these converging analyses strengthen the overall interpretation, while we acknowledge the temporal confound as a remaining limitation that future work with designs explicitly counterbalancing serial position across roles could address more definitively. We reproduce below the main text results, supplementary methods and Supplementary Figure (Supplementary Figure 4) detailing this complimentary analysis.

Main manuscript

“To further test the robustness of the above effect, and to assess whether abstract dynamical roles are reflected in neural representational geometry during task performance, we conducted a complementary cross-phase representational similarity analysis. This analysis addresses directly whether neural representations of central nodes are stable across their distinct serial positions within a sequence, an expected property if representations reflect abstract structural identity rather than time-in-sequence. Specifically, for each participant we estimated neural patterns for each central node separately at its first and second occurrence (phase) within the sequence traversal, using occurrence-specific GLMs, and computed cross-phase similarity between prior and transfer learning phases. This approach controls for a potential time-in-sequence confound inherent in role-based comparisons, since now, the same functional node is compared across different serial positions (rather than comparing different stimuli with a common sequence position, by virtue of sharing role). Consistent with the decoding results, we observed increased cross-phase similarity in participants with prior experience of Component 6 relative to those with prior experience of Component 4 (Supplementary Figure 4). Specifically, a significant group difference was observed at a late stage of the trial (1.125–1.25 sec post-stimulus onset, P_FWE = 0.043, two-sample non-parametric permutation test, one-tailed), with additional differences at earlier timepoints that did not survive correction for multiple comparisons (0.125 sec: t = 2.18, p = .034; 0.79 sec: t = 2.84, p = .007). These effects indicate increased neural similarity for central node representations across phases in participants who experienced these roles consistently during prior learning. We consider these results complementary to the main analyses and interpret them with caution given that the group difference was only corrected for multiple comparisons at the late cluster. Importantly, the temporal profile of the cross-phase similarity effect does not perfectly overlap with the time windows identified in the original dynamical role decoding analysis, which limits the strength of the convergence argument. We therefore cannot fully rule out the possibility that the earlier decoding effects reported in Figure 4 partly reflect temporal position within the sequence rather than abstract structural role coding per se. We additionally note that this analysis is based on a smaller and less variable set of comparisons than either the original decoding analysis or the role-based RSA, since the bridge graph contains only two central nodes whose repeated occurrences can be directly compared in this way, limiting its statistical power in isolation. Nonetheless, the observation of a group-specific effect in the cross-phase analysis which directly controls for serial position provides support for the structural scaffolding account, and the late corrected cluster suggests that role-consistent representations do emerge across learning phases, albeit potentially at a later stage of processing than the decoding results alone would suggest.”

Supplementary Information

“Cross-phase representational similarity analysis

Complementary to the decoding analysis in Figure 4, we conducted a cross-phase representational similarity analysis (RSA) to test whether neural representations of central nodes were stable across their distinct serial positions within a sequence, a property expected under abstract structural role coding but not under serial position coding. We focused on central nodes of the 6-state graph, as these nodes appear twice within each sequence traversal of the bridge graph: the left central node occurs at steps 1 and 4, the right central node occurs at steps 5 and 8, allowing first and second presentations to be dissociated.

Structural roles were assigned separately for prior and transfer learning phases. For line graphs (experienced by the Condition 4 during prior learning), central nodes were identified from node degree in the adjacency matrix (boundary: degree 1; intermediate: neighbors of boundary; central: remaining nodes), and the two central nodes were ordered by first appearance in the trial list. For bridge graphs, central nodes were identified from the predefined 8-step traversal template.

At each time point and sensor, condition-specific activations were estimated using occurrence-specific GLMs: one GLM fitted to trials corresponding to the first occurrence of each number in the trial list, and one fitted to trials corresponding to the second occurrence. Component 6-specific beta estimates were extracted from each GLM separately, yielding four sets of neural patterns per subject: prior-phase first occurrence, prior-phase second occurrence, transfer-phase first occurrence, and transfer-phase second occurrence.

Cross-phase representational dissimilarity was then computed using an L1 distance metric (mean absolute difference across sensors) across four cross-presentation pairs: Left central (prior first, transfer second), left central (prior second, transfer first), right central (prior first, transfer second), and right central (prior second, transfer first). These pairings compare the same node across different serial positions and different learning phases, directly controlling for time-in-sequence as a potential confound. The mean L1 distance across the four pairs constituted the cross-presentation dissimilarity index at each time point, where lower values indicate greater neural similarity. Timecourses were baseline-corrected by subtracting the pre-stimulus mean (timepoints 1–25, corresponding to −250 to 0 ms) per participant, and analyses were restricted to post-stimulus timepoints. Group differences between Condition 6 and Condition 4 participants were assessed using a two-sample non-parametric cluster permutation test (cluster-forming threshold p = 0.10, cluster-level alpha = 0.05, 1,000 permutations, one-tailed, consistent with the directional prediction that Condition 6 participants would show greater cross-phase neural similarity for central nodes). The results are presented in Supplementary Figure 4.”

(4) In the following section, the authors correlate decoding accuracy with participants' behavioral performance across different conditions. However, out of the four reported correlations and the additional comparison of differences between conditions, only one correlation and one correlation difference reach significance, and only marginally so. The interpretation of this finding should therefore be more cautious, especially if it is used to support a link between neural representations and behavior. Additionally, it is possible that correlation with a more clearly defined or targeted neural signature, more directly tied to the hypothesized representational content, could yield stronger or more interpretable correlations.

We agree with the reviewer that the reported brain-behavior correlations should be interpreted cautiously. Only one of the four correlations reached significance, with effect sizes of medium magnitude. However, we emphasize this correlation is significant when corrected for multiple comparisons for two tests, experience and inference probes (ɑ = 0.025). Additionally, it is important to note, we did not expect all four correlations to be significant. The conditions differ in the extent to which behavior is predicted to depend on the specific neural representation indexed by the decoding analysis, and only a subset of conditions are theoretically expected to show a brain–behavior relationship. The absence of significant correlations in the remaining conditions is therefore not inconsistent with our account. For example, we did not expect the two 4-cycle correlations (depicted in green, Fig. 4C) to be significant.

We agree with the reviewer that the reported brain-behavior correlations should be interpreted cautiously, and have revised the manuscript to reflect a more measured interpretation. We note that the correlation value reported in the revised manuscript (ρ = .48, p = .016) differs slightly from the originally submitted version (ρ = .41, p = .041), due to a correction in the timepoint-averaging window used to derive the decoding strength measure. We consider the revised value the authoritative one and have updated all relevant text accordingly.

Only one of the four correlations reached significance, with an effect size of medium magnitude. We would not characterise this correlation as marginal: it survives correction for multiple comparisons across the two theoretically motivated tests (experience and inference probes; corrected α = 0.025), and a medium-magnitude Spearman correlation (ρ = .48) in an individual-differences analysis with n = 25 is consistent with a meaningful brain-behavior relationship, even if the sample size limits the precision of the estimate. We agree that independent replication is needed to establish robustness.

Additionally, it is important to note that we did not expect all four correlations to be significant. The conditions differ in the extent to which behavior is predicted to depend on the specific neural representation indexed by the decoding analysis, and only a subset of conditions are theoretically expected to show a brain–behavior relationship. The absence of significant correlations in the remaining conditions is therefore not inconsistent with our account. For example, we did not expect the two 4-cycle correlations (depicted in green, Fig. 4C) to be significant, since the 4-cycle-prior group did not experience the bridge graph structure during prior learning and therefore would not be expected to show the same alignment between dynamical role representations and transfer learning performance.

In any event, based on the reviewer’s (and comments from other reviewers about these correlations), we have revised the manuscript to tone down the interpretation of these results (reproduced below):

“These findings are consistent with, though not conclusive evidence for, dynamical role representations conveying behavioral benefits. Speculatively, such knowledge representation might enable individuals to constrain their predictions of successive states. For example, upon transitioning to a bridge node one can predict that the successor state will involve "the other side of the bridge". While we acknowledge that the correlation results appear inconsistent across experience and inference probes, it is important to consider that our MEG classifiers were trained and tested on experienced transitions alone, such that enhanced classification accuracy is more reflective of a neural representation of experienced transitions. Under this assumption, the MEG classifier would more closely align with behavioral accuracy for the experience probes (testing actually experienced transitions) than for the inference probes (testing never experienced but only implied transitions on the graph). See Discussion for further elaboration of this point, including why this asymmetry is expected under a model of gradual, imperfect structural learning.

The observed correlation is also consistent with a general learning-ability account, whereby higher-ability participants both better discover the Component 6 structure (that happens to be shared) across sessions and show higher behavioral accuracy overall. For example, higher-ability participants may rediscover this structure during transfer, without necessarily benefitting from having been exposed to it during primary learning. This ability-based account predicts that decoding strength should also relate to Component 4 accuracy, and that Component 4 and Component 6 accuracy should correlate within each prior learning group. Neither prediction was consistently supported (Supplementary Figure 5): The Component 4 decoding strength and Component 4 correlation was not significant (ρ = .38, p = .071), and did not differ significantly from the Component 4 decoding strength and Component 6 correlation (Steiger's test, p = .656); accuracy correlations between Component 4 and Component 6 were significant in only one prior learning group for each probe type, with no significant between-group differences (all p > .29).

Given this inconsistent pattern, we consider the general-ability account unsupported, though our low per-condition sample size limits power to rule it out. We therefore treat the brain-behavior correlation as selective to Component 6 experience probes in this sample, and treat it as supportive rather than primary evidence for a link between role representations and transfer, pending independent replication.”

We agree that a more targeted neural signature could in principle yield a more interpretable correlation with behavior. Our interpretation of the reviewer's comment is that this implicitly requires a measure that captures representational content at the level of individual trials or items, for example, a trial-by-trial estimate of how strongly a given stimulus is represented in its abstract structural role, which could then be correlated with individual participants' behavioral performance. Our current decoding measure does not provide this: it indexes classification accuracy averaged across all trials of a given transition type, yielding a single scalar value per participant per timepoint rather than a trial- or item-level representational estimate.

Computing such a trial-level measure from our data would require a fundamentally different analysis approach, for example, single-trial decoding confidence scores or RSA-based measures of individual stimulus representational fidelity, which goes beyond what our current classifier-based pipeline and data structure can deliver. We have instead conducted additional analyses that more directly index neural representational content (role-based representational similarity analyses, reported in the revised manuscript), which provide converging support for the proposed computational account.

Minor Comments:

“During preprocessing, sensors were excluded based on an identified noise level. However, the authors do not specify the threshold used to define this noise level, nor do they report how many sensors were excluded per participant. It would be helpful to have these details. Additionally, it is unclear why the authors opted to exclude sensors rather than removing noise with MaxFiltering or interpolating bad sensors. Finally, the authors should report how many trials were discarded on average (and standard deviation) per participant.”

We thank the reviewer for these comments. We have updated the methods section to address the points raised.

Regarding the noise threshold and number of excluded sensors/trials: sensors were excluded using the default osl_detect_artefacts settings, which apply a Generalised Extreme Studentised Deviate (GESD) test at a significance level of 0.05. On average across all MEG runs and participants, 8.67 sensors (SD = 7.36) and 9.13 trials (SD = 11.46) were excluded. These details have been added to the methods section.

Regarding the choice to exclude sensors rather than applying MaxFilter or interpolation: MaxFilter was not applicable here as our data were acquired on a CTF system, for which MaxFiltering is not designed. Sensor interpolation was not performed as bad channels are excluded prior to ICA decomposition following standard OSL pipeline practice. Given that excluded sensors represented on average ~3% of the total 273 sensors, we do not expect this to have meaningfully affected our results.

The updated methods section now reads as follows:

“On average across all MEG runs and participants, 8.67 sensors (SD = 7.36) and 9.13 trials (SD = 11.46) were excluded (using default osl_detect_artefacts settings; channel significance level for Generalised Extreme Studentised Deviate = 0.05). We did not apply MaxFilter as our data were acquired on a CTF system, for which MaxFiltering is not designed. Sensor interpolation was not performed as bad channels are excluded prior to ICA decomposition following standard OSL pipeline practice; given that excluded sensors represented on average ~3% of the total 273 sensors, we do not expect this to have meaningfully affected the results.”

Reviewer #3 (Recommendations for the authors):

(1) The manuscript seems to use the terms generalization, transfer, compositionality, and compositional generalization somewhat interchangeably. While I appreciate that both compositionality and transfer learning are central to the task design, I believe it's important to distinguish these concepts more clearly. Specifically, generalization typically refers to the ability to apply learned structure within a task to novel but similar instances, whereas transfer involves applying learned components or representations across different tasks or contexts. Clarifying these distinctions - ideally with explicit definitions early in the introduction - would help avoid confusion and better situate the contribution of this work within the broader literature.

(2) The manuscript frequently refers to abstract knowledge or abstract representations. However, I would argue that abstract knowledge is typically characterized by its ability to generalize or transfer across different tasks or domains. In the current paradigm, while the structural relationships between stimuli change, the task demands and sensory modalities remain constant. It would strengthen the conceptual clarity of the paper if the authors could more carefully define what they mean by "abstract".

We thank the reviewer for these suggestions. We agree that “abstract” can carry different meanings across the literature. We have now added an explicit definition to the introduction clarifying our usage. We use 'abstract' to refer to representations that are invariant to surface sensory features while capturing relational structure, consistent with findings in a representation learning literature (e.g., Bernardi et al., 2020; Behrens et al., 2018). While the reviewer raises the valid point that task demands and sensory modality remain constant across learning phases, we highlight that the stimuli themselves are entirely novel in the transfer phase, such that any representational similarity must reflect relational rather than perceptual overlap.

We have added explicit definitions of abstract, generalization, transfer, and compositional generalization to the introduction, clarifying how each term applies to our task design and how they relate to one another. We have also aimed for more consistent usage throughout the manuscript.

This added paragraph in the Introduction reads as follows:

“Throughout this work, we use "generalization" to refer to the ability to apply learned relational structure to novel stimulus compositions within the same task context, and "transfer" to refer to the reuse of (abstracted) representations across distinct task contexts or learning phases. Note we use “abstract” to refer to representations that are invariant to the sensory specifics of individual stimuli, capturing relational structure (e.g., the role a stimulus occupies within a graph) rather than surface features. Our task is structured compositionally in that every experienced compound stimulus is the joint product of two simultaneously evolving subprocesses, while the specific compound encountered during transfer learning is never observed during prior learning. We use 'decomposition' and 'transfer' to refer to the abstraction and reuse of a single such subprocess across learning phases, and reserve “compositional generalization” for the broader phenomenon whereby task performance depends on appropriately decomposing and recombining structural knowledge, a capacity our design requires of participants. Here we acknowledge that our study speaks most directly to decomposition and transfer of individual structural components rather than to the act of recombination itself.”

(3) The supplementary materials include a section on replay analyses, but it is unclear how these analyses relate to the main text. I could not find a corresponding discussion or reference to replay in the results or discussion sections.

We thank the reviewer for flagging this. The replay analysis was referenced in the Discussion of the previous version of the manuscript. To make this analysis more visible, we have also added a brief summary of the null replay result to the Results section, with a pointer to the Supplementary Results where the full analysis and discussion of limiting factors are reported.

eLife Assessment

This study provided useful information regarding the role of a subpopulation of cortical interneurons (Chrna2-expressing Martinotti cells) in motor learning and cortical dynamics. However, while some behavior and imaging data are compelling, limitations in the motor-learning task design and data analysis make it difficult to support the central conclusions.

Reviewer #2 (Public review):

Summary:

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

Strengths:

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

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

Weaknesses:

I appreciate the authors' efforts to address the remaining issues raised by the reviewers and agree that it becomes clearer by refocusing the study away from learning towards effects of Ma2 cell activation on motor execution. I also appreciate added clarifications throughout the manuscript.

That said the study still relies heavily on measuring changes throughout "performance refinement", which is just another word for learning. Given their task design, it is hard to assess any learning or refinement over the course of training. This makes it hard to interpret why certain changes in neuronal activity occur at the defined sessions. The main conclusion thus is that there is "a complex interplay between Mα2 activity, oscillatory brain dynamics, and motor function", as the authors state as well in their discussion. I would encourage a further focus or restriction of the data presentation and interpretation on motor execution and limit any plasticity findings on training session 6, where the reaching task properly starts, such as in the study the authors reference (10.1016/j.neuron.2021.07.030).

Relatedly, the authors cannot say that "enhancing Mα2 cell excitability improved the performance of a motor task relying on an already refined skill (Figure 6C)" (lines 1134-1136). Since there is no improvement in success ratio throughout the training prior to CLZ injection (Supp Fig. 7A), there is no evidence for learning or skill refinement in the period where the pellet is presented on the plate.

Regarding the Ca2+ image analysis, the baseline subtraction approach the authors use only appropriately reduces part of the technical variations seen with 1p Ca2+ imaging. It only partially accounts for differences in GCaMP expression or different illumination across cells, or slower task-independent fluctuations in neuronal activity. Given the goal of the analysis is to show activity modulation in the peri-trial interval around prehension onset, some version of trial-wise baseline normalization (which is commonly used) is necessary and would likely result in less confusing results.

Minor:

- Unclear which session traces in figure 1I are from.

- Label in Supp Fig. 7A should say "switch from spoon to plate".

Author response:

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

eLife Assessment:

This valuable study addresses a critical question regarding the role of a subpopulation of cortical interneurons (Chrna2-expressing Martinotti cells) in motor learning and cortical dynamics. However, despite the inclusion of interesting behavioral and imaging data, significant limitations remain, even after revision, in the design of the motor learning task and the associated data analyses. As a result, the presented data are incomplete to support the central conclusions.

We appreciate the valuable feedback. To address these concerns, we have rephrased all conclusions to explicitly focus on motor execution consolidation rather than learning, as our task design inherently emphasizes performance refinement over acquisition. This reframing is now fully supported by compelling evidence from our data and analysis, demonstrating that Mα2 cell activation reduces pyramidal cell population plasticity and improves motor execution.

Public Reviews:

Reviewer #1 (Public review):

In line 331, the authors refer to session 5 as “training,” describing it as the final spoon session, and session 6 as “re-training,” because it is the first session in which the pellet is presented on the plate rather than on the spoon. However, in Fig. 1F-H, even in the Ctrl group, it is clear that the performance drops significantly in session 5, which is supposed to be the easiest session before switching to the more difficult plate condition.

As described in the methods, the task is made progressively harder (as evidenced by the success ratio curve), therefore a lower performance in session 5 (the hardest spoon session) is expected, as now explicitly stated in the results (lines 816-825).

In the classic pellet-reaching task, the spoon sessions would typically be considered “shaping”, while the plate sessions would represent the actual training phase. However, in this manuscript, the authors still insist on referring to session 2 as “learning” and session 5 as “training.” I don’t understand the difference between session 2 and session 5, especially when session 5’s performance is lower than session 2 (even in Fig 1H when you compare succ ratio).

The experiments were designed to investigate the effect of Ma2 cell activation on learning and training outcomes. As we wanted to compare completely naive mice with mice that had been trained and that had learned the task, we included the first sessions with the spoon in our analysis. In many other studies or experimental set-ups those sessions would have been considered shaping. The reason for this design is that we did not want to miss early changes related to training and learning that could have a great impact on the final outcome, as most dendritic spines that correlate with learning are formed after the initial training session and are maintained in the second training session (10.1016/j.neuron.2021.07.030). During pilot studies, several different strategies were tested and it was discovered that the mice needed to succeed several times in the initial sessions in order to keep trying. Hence, the pellet presentation needed to be less challenging in the beginning than in the final sessions in which the pellet was presented on the plate. Therefore, as explained above, the task was made progressively harder throughout the session and from session to session, thus mice expectedly had more difficulty in session 5. We agree that the session names could be misleading, particularly in the context of our task design, and therefore changed all session labels to session numbers.

Since session 6 (on the plate) is considered as “retraining,” why don’t the authors present the behavioral results beyond session 6? As a result, it remains unclear whether the animals improved their performance during the retraining phase.

We agree, and now show behavior performance from sessions 1 to 10 for the behavioural group. Mice displayed a clear improvement after session 6, reaching a plateau after session 7 (Figure 1F-H). Unfortunately, due to technical challenges with reattaching the miniscope multiple times and following the same neurons over time, we do not have miniscope data from beyond session 6, which is why we initially only presented behavioural data up to session 6.

Lastly, in Fig. 4B the authors present only the success ratio and claim that performance improves with CLZ application. However, when comparing sessions 8-10 between the Ctrl and Cre+ groups, there already appears to be a baseline difference. CLZ treatment in Cre+ mice seem to bring performance only to the WT level rather than producing a clear improvement beyond baseline.

In the previous version (first submitted), we showed the success rate (number of success per minute), to the same conclusion. Note that our conclusion is supported by the effect of treatment found in Cre+, meaning that Cre+ improved upon CLZ (Ma2 cells increase in excitability) compared to NaCl (No change in Ma2 cell excitability). The control group is a control for the Ma2 cell activation specificity, and as expected it shows no treatment effect (since there is no hM3Dq expression, as already reported, Supplemental Table 1L). Further, the reviewer’s suggestion of a baseline difference between groups is not supported by our data, since no statistical differences were found on these comparisons (t-test eff. size < -1.587; p > 0.293; as already stated in the methods, the absence of significance bars “for any comparison means no significant difference was found”).

Regarding the alignment between imaging and behavior, the authors report ~100 prehensions per minute. However, the calcium imaging traces show fewer than 20-30 spikes over 150 seconds (~2.5 min; Fig. 1E). This discrepancy raises concerns about whether the authors can truly isolate calcium signals corresponding to individual prehension events (either successful ones or multiple combined events for unsuccessful attempts). The manuscript still does not present behavioral data that directly aligns prehension events with calcium imaging activity. Although the authors performed analyses suggesting that prehension-related activity does not systematically alter nonprehension epochs, this claim is difficult to evaluate without seeing the underlying traces. It is therefore unclear how the authors selected the example calcium traces aligned to prehension onset, given that there are more than 100 prehension events per minute.

If we understand this comment correctly, the reviewer presumes that there should be a direct relationship between upper motor neuron activity and prehension activity (i.e. muscle activity). Such one-to-one correspondence between prehension events and calcium transients in individual pyramidal neurons should not be expected.

Upper motor neurons do not fire in a direct, event-by-event relationship with movements. Instead, motor actions emerge from distributed population activity, where individual neurons contribute sparsely and variably. Classical work described in the book Corticospinal function and voluntary movement by Porter and Lemon (1993) show that the relationship between single-neuron firing and muscle contraction is indirect and probabilistic. This circumstance required spike-triggered averaging based on several thousands of recordings to reveal consistent relationships between neuronal firing and motor output, precisely because individual spikes are neither necessary nor sufficient for a movement instance.

Calcium imaging accentuates this dissociation. Calcium signals are a temporally filtered and nonlinear proxy of spiking activity, with limited sensitivity to single spikes and a bias toward burst-like or temporally clustered firing. As a result, many individual action potentials, especially those occurring in isolation, may not generate detectable calcium transients. Therefore, observing fewer calcium events than behavioral actions is expected and does not imply a failure to capture relevant neural dynamics.

Thus, motor cortex representations during skilled behaviors such as prehension are distributed across neuronal ensembles rather than encoded at the level of single neurons. Each prehension recruits a subset of neurons, and each neuron contributes only intermittently across trials. Consequently, the appropriate level of analysis is not event-by-event alignment for individual cells, but rather population-level dynamics (e.g., ensemble activity patterns, dimensionality reduction, or trial-averaged responses aligned to behavior).

We have not selected “example calcium traces aligned to prehension onset”. We performed all quantifications on the average of all trials, and we thus show one example of such traces in Figure I and all average traces in Supplemental Figure 3I. All trials were included, therefore there is no “selection”, as already explained in the methods section (lines 607-611).

Taken together, the apparent discrepancy between behavioral event rate and calcium event frequency reflects known principles of motor cortex coding and calcium imaging limitations, rather than a methodological inconsistency. This explanation is now added to the discussion section (lines 1186-1211).

Regarding the baseline effect analysis (Supplemental Figure 3I), we do not understand the claim that the underlying traces for the quantification shown cannot be seen, when they are explicitly shown in the same figure (Supplemental Figure 3A-E). We agree that the figure legends for Supplemental Figures 2 and 3 were missing the description of the traces shown as average traces of all trials for all neurons, with the shade showing the standard error of the mean. We now added this information to the Supplemental Figures 2 and 3 legends.

In Fig. 1I, the authors also did not address why neural activity during successful trials is already lower one second before movement onset. The longer traces provided do not help to explain this observation or clarify the origin of this pre-movement reduction in activity. It actually further suggests that there may be some artifacts in the imaging that could affect the analysis.

We understand that the example shown in Figure 1I can cause confusion, therefore we chose a clearer example neuron. As explained in the previous comment, no quantifications were made at the trial level, therefore Figure 1I shows the average +SEM of all trials.

Overall, because it remains difficult to understand exactly what the authors are analyzing (and because the definitions of the motor learning stages appear arbitrary) it is difficult to agree with the authors’ conclusion that Ma2s cells reduce PyrN cell assembly plasticity during learning, thereby possibly facilitating already acquired motor skills.

We rewrote our conclusions to better fit our results and thoroughly explained the calcium imaging analysis concerns above. No quantifications (nor claims) were done at single-level trial, and we clearly show the exact traces from where the quantifications were derived (Figure 1E, 1I; Supplemental Figure 2A-H and 3A-H.

Reviewer #2 (Public review):

While the authors addressed some of the concerns raised by the reviewers, several major limitations still exist in the revised manuscript.

(1) I appreciate the authors now showing more measures of the prehension task (total reaches, success reaches/min, and success ratio) and providing more details on the task design. However, it is unclear why the authors chose a task design that is somewhat different from the commonly used approach. Here they increase the distance of the food pellet each session and are thus making the task increasingly harder, whereas commonly the target distance is kept stable (See 10.1038/nature08389 for example). The result is that important readouts of learning (e. g. success rate) thus remain stable, making it impossible to judge if learning has occurred, without a control group of non-trained mice. This makes it impossible to judge if the task is affected by increased Ma2 cell excitability, since there is no reference of how these measurements are supposed to change in a mouse that learns or doesn’t learn the task.

The experiments were designed to investigate the effect of Ma2 cell activation on learning and training outcomes. As we wanted to compare completely naive mice with mice that had been trained and that had learned the task, we included the first sessions with the spoon in our analysis. In many other studies or experimental set-ups, those sessions would have been considered shaping. The reason for this design is that we did not want to miss early changes related to training and learning, that could have a great impact on the final outcome, as most dendritic spines that correlate with learning are formed after the initial training session and are maintained in the second training session (10.1016/j.neuron.2021.07.030). During pilot studies, several different strategies were tested and it was discovered that the mice needed to succeed several times in the initial sessions in order to keep trying. Hence, the pellet presentation needed to be less challenging in the beginning than in the final sessions in which the pellet was presented on the plate. On previous pilot experiments, when the spoon distance was kept stable across sessions, most animals did not perform the task upon switching to the plate position (a large and abrupt increase in difficulty), whereas progressively increasing the difficulty led to more mice reaching for the plate target. We agree that this task design better reflects motor execution than learning, and have stated throughout the manuscript that learning is unaffected. We also reframed conclusions about learned motor execution to better fit our task design.

(2) Regarding the analysis of the calcium imaging data, it is still unclear why the authors cannot report a commonly used dF/F0 or z-score value, as recommended by both reviewers. The authors state the 1 sec time window prior to the prehension cannot be used as a baseline (F0), as there might be preparatory motor activity. In that case an even earlier window (such as -2 to -1sec) or z-scores should be used. The current version relabeling the background subtracted fluorescence signal as dF/F0 is misleading. Relatedly, it is unclear why the authors don’t think the 1 sec window before prehension cannot be used as baseline, but at the same time use the difference in calcium activity before and after prehension onset as a cut-off criterion for defining cells as modulated during prehension and including in the analysis.

As explained in the previous review round, we performed baseline subtraction as described in the methods section: "The baseline components were estimated using a running 8th percentile derived from kernel density estimation over a 250-frame window, and subtracted from each respective trace. Although true F0 estimation is not possible on 1-photon data, the baseline-subtracted trace is here referred to as ΔF/F. The fact that the label ΔF/F is commonly used in 1-photon data is misleading, since studies analyzing miniscope data using the label ΔF/F define F0 as the raw video frame minimum (e.g., 10.1371/journal.pbio.3000584, 10.1111/ejn.70190), mean of “non-spiking” periods for all neurons (e.g., 10.1016/j.crmeth.2025.101221), or providing no F0 definition (e.g., 10.1002/adfm.202511627, 10.3389/fnagi.2025.1550673, 10.1016/j.bpsgos.2025.100531, 10.1038/s41586-025-10096-w), none of which are true F0. We reiterate, as in the previous review round, that any use of the label ΔF/F is technically incorrect, not only here but in any study using 1-photon data.

After the previous review round we defined F0 as the baseline components estimated as described above, which is a much more reliable and standardized measure of baseline components estimation compared to the measures cited above, using an extensively validated algorithm (10.7554/eLife.38173), and labeled our data as ΔF/F since they are already properly scaled, as previously done (10.3389/fnagi.2025.1550673). However, after this discussion, we suggest reverting back to the more correct technical term “ΔF detr.”. This also aligns with previous other similar work (e.g., 10.1073/pnas.2503139122, 10.3389/fnsys.2026.1765204), and that is what our traces (and any traces from any study derived from 1-photon data) actually are.

We searched and found some studies analysing miniscope data that have z-scored fluorescence traces, but none of those provided a clear justification for doing so (e.g., 10.1016/j.isci.2026.115514, 10.1016/j.neuron.2025.07.020). We do not think that z-scoring appropriately scales fluorescence traces, particularly when baseline components are already subtracted. Fluorescence traces exhibit strong positive skewness, characteristic of calcium imaging data where the baseline is stable (after removing baseline components as described above) but transient events create a heavy tail, and z-scoring relies on the mean and standard deviation. In highly skewed distributions, both the mean and standard deviation are strongly influenced by the magnitude and frequency of these transients. Consequently, z-scoring the traces artificially compresses the baseline and distorts the relative amplitude of events, leading to biased comparisons between different cells by penalizing cells according to their own activity (more active cells will have more transients, therefore larger mean, and lower z-scored traces), instead of normalizing to a biologically meaningful scale. This can be easily demonstrated, take for example the signal fragment in Author response image 1.

Author response image 1.

The A panel shows a fluorescence trace from one neuron. If we slice this trace in half (B), the background component is the same, since they are derived from the same cell. If we z-score those halves, the half with higher activity is penalized (C, blue), while the less active half is artificially increased (C, orange). Since both signals are from the same cell, if z-scoring the traces was an appropriate normalization procedure, the differences between them should remain, which is not the case. Therefore, it is clear that z-scoring does not properly scale fluorescence traces for the purpose of comparisons between cells.

Regarding the inclusion criterion, we used the difference between before and after activity precisely to include only neurons related to the task, which does not mean that the "before" epoch is baseline, only that it is not the same as the "after" epoch, as shown by data already included in the manuscript. Further, we already ran the suggested analysis (after/before ratio) and presented the full results set: “We found a three-way interaction effect of genotype, session, and prehension accuracy (Supplemental Table 2K; Supplemental Figure 4), with genotype effects for both accuracies on the naive session and successful prehensions during the learning session (Supplemental Table 2L-N), and accuracy effects in control mice across the learning and training sessions (Supplemental Table 2O-P). This indicates that Mα2 cell activity might selectively refine the neural signal balance during planning/execution phases of motor execution particularly on initial training sessions.”

(3) While the authors have improved their statistical reporting, key information is still missing in several places. For example, no N-numbers are listed in legends for figure 3, and there is no mention of the number of mice for analysis in figures 2 and 3. For clarity, the authors should also include the statistical test performed in the figure legends for any p-values shown in the figure.

We thank the reviewer for the attention to details, and clarified that Figure 2 shows latency analysis for the same dataset as Figure 1, we now explicitly state the number of mice and assemblies in Figure 3 results section, and added the statistical tests performed in each panel’s legend.

eLife Assessment

This paper presents a collection of analyses relating structure and function in the whole-brain Drosophila EM connectome and whole-brain calcium imaging data. The linkage of detailed anatomical structure with population activity is of broad interest in circuit neuroscience in light of increasingly detailed brain maps, but the methods used made the evidence incomplete due to indirect consideration of biological detail such as neurotransmitter identity and assumptions about the relationship between structural correlations and functional correlations. The conclusions are useful for specific network observations, but a more thorough analysis of the anatomical and functional data is needed to support the overall claims.

Reviewer #1 (Public review):

In the previous revision the authors addressed some of the points, but they also make some technical errors. My overall view of the manuscript hasn't changed since the original evaluation.

Previously I noted that SC sparsity presents an issue when comparing to full FC matrices. They authors misinterpreted the Honey et al paper. They resampled ALL entries of the SC matrix (including zeros) from a Gaussian distribution. In effect, this assigns zeros small (but uniform) weights. In Honey et al, the authors resampled only existing edge weights from a gaussian distribution (the rationale at the time was that there might be pushback against the extremely heavy-tailed edge weight distribution). In other words, the zeros are still zeros following this resampling procedure.

That said, I agree that the log transform is likely useful or necessary given edge weight distributions.

In short, I still think that the approach is interesting and meritorious, I just don't think the execution is correct.

Comments on revised version.

In this revision the authors addressed my remaining points.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

In this revision the authors address some of the points, but they also make some technical errors. My overall view of the manuscript hasn't changed since the original evaluation.

Previously I noted that SC sparsity presents an issue when comparing to full FC matrices. They author misinterpreted the Honey et al paper. They resampled ALL entries of the SC matrix (including zeros) from a Gaussian distribution. In effect, this assigns zeros small (but uniform) weights. In Honey et al., the authors resampled only existing edge weights from a gaussian distribution (the rationale at the time was that there might be pushback against the extremely heavy-tailed edge weight distribution). In other words, the zeros are still zeros following this resampling procedure.

That said, I agree that the log transform is likely useful or necessary given edge weight distributions.

In short, I still think that the approach is interesting and meritorious, I just don't think the execution is correct.

Thank you very much for pointing this out. Regarding Gaussian resampling, we recalculated the results to avoid including zeros of the SC matrix in the resampling. As before, we confirmed that when the ROI count is small (200 or less), Gaussian resampling tended to produce a higher SC-FC correlation (Extended Data Figure 2-2a), as the ROI count increased so too did the Sparsity Rate (Extended Data Figure 2-2b), and the log transform tended to produce a higher SC-FC correlation (Extended Data Figure 2-2a). However, for both methods we still observed a tendency for the FC-SC correlation to decrease as the ROI count increased. Therefore, we will continue to use the log transform, but we will update Ext.Data.Fig.2-2 and now explicitly state in the text (lines 132-137) that there is a tendency for FC-SC correlation to be higher as the Sparsity Rate increases.

Reviewer #3 (Public review):

Summary:

In this manuscript, Okuno et al. re-analyze whole-brain imaging data collected in another paper (Brezovec et al., 2024) in the context of the two currently available Drosophila connectome datasets: the partial "FlyEM" (hemibrain) dataset (Scheffer et al., 2020) and the whole-brain "FlyWire" dataset (Dorkenwald et al., 2024). They apply existing fMRI signal processing algorithms to the fly imaging data and compute function-structure correlations across a variety of post-processing parameters (noise reduction methods, ROI size), demonstrating an inverse relationship between ROI size and FC-SC correlation. The authors go on to look at structural connectivity amongst more polarized or less polarized neurons, and suggest that stronger FC-SC correlations are driven by more polarized neurons.

Strengths:

(1) The result that larger mesoscale ROIs have higher correlation with structural data is interesting. This has been previously discussed in Drosophila in Turner et al., 2021, but here it is quantified more extensively.

(2) The quantification of neuron polarization (PPSSI) as applied to these structural data is a promising approach for quantifying differences in spatial synapse distribution. The revision now uses morphological cable length for some analyses rather than straight-line distance, which improves the realism and interpretability of these results.

Weaknesses:

One should not score noise/nuisance removal methods solely by their impact on FC-SC correlation values, because we do not know a priori that direct structural connections correspond with strong functional correlations. In fact, work in C. elegans, where we have access to both a connectome and neuron-resolution functional data, suggests that this relationship is weak (Yemini et al., 2021; Randi et al., 2023). Similarly, I don't think it's appropriate to tune the confidence scores on the EM datasets using FC-SC correlations as an output metric. While it is likely that some FC-SC relationship does exist at large scales, it does not in my view justify use of this metric for evaluating noise removal methods, since such methods may inadvertently remove real neural correlates. This concern remains unaddressed in the revision.

Thank you for pointing this out. Firstly, regarding C. elegans, and mentioned in our Discussion section, the FC of calcium dynamics in the soma was compared with the SC of the whole cell in C. elegans studies (Yemini et al., 2021; Randi et al., 2023). Neurons in C. elegans are non-spiking, and calcium dynamics have been shown to differ between compartments (soma and other parts) in a neuron (Hendricks, et al., 2012). While fly and human studies compared the SC and FC of synaptic population dynamics in brain regions, C. elegans studies compared the SC of cells (not regions) with the FC of dynamics in the cell body. This discrepancy is thought to explain the differing results.

In support of our approach, numerous studies (cited in our manuscript Introduction) have demonstrated statistically significant FC-SC relationships across species and modalities, including human rs-fMRI and diffusion MRI, marmoset retrograde tracing and rs-fMRI, and Drosophila calcium imaging and connectomics. These findings suggest that still SC provides an important anatomical constraint on FC.

We also used FC-SC detection in addition to FC-SC correlation. Correlation measures similarity by normalizing the covariance between FC and SC values by the product of their standard deviations. On the other hand, FC-SC detection measures the detection accuracy (AUC) of FC. Since FC and log-transformed SC are fundamentally different, both correlation and detection are considered useful indicators. In both indicators, high correlation and high AUC were observed using the same nuisance removal method. Our results suggested that head movement and the attenuation of fluorescent molecules due to long-duration imaging are the sources of noise, and this provides supporting evidence that the procedure may be improving signal quality. Therefore, we consider FC-SC correspondence to be a biologically motivated validation metric for nuisance removal and a way to validate nuisance removal factors.

However, we agree that FC-SC correspondence may not be regarded as a ground truth objective, and that biologically meaningful functional relationships may also arise through indirect pathways, network effects, neuromodulation, and other mechanisms not captured by direct SC. As examples, neuropeptides diffuse to transmit signals, while gap junctions tightly and directly connect neurons, and there is also a pathway mediated by glia. Although SC cannot be equated with FC, we believe that the relationship between the two remains strong. To investigate this matter further, we consider it necessary to conduct additional research involving the simultaneous measurement of whole-brain FC (calcium) and ‘non-synaptic’ signals.

We have rewritten the manuscript to say that SC is not a ground truth but a biologically motivated evaluation metric (line 102) for nuisance removal, and we have added content to the discussion around the points discussed here (lines 296-306).

Any discussion of FC-SC comparisons should include an analysis of excitatory/inhibitory neurotransmitters, which are available in the fly connectome dataset. The authors examine the ratios of input and output neurotransmitters in different defined regions. However, I think it would be more useful to integrate the neurotransmitter information more fully into the assessment of SC, for instance: examining the signed weight (excitatory - inhibitory), or by examining the excitatory and inhibitory networks separately.

Thank you for suggesting the idea. This matter has not been thoroughly investigated in this study. Figure 3g shows the correlation between neurotransmitters (rate of neurons) and FC-SC correlation, but we have not been able to analyze the effects on FC-SC in detail. Due to this, we think that this topic requires further analysis as part of future work.

Comparisons between fly and human MRI data are also premature here. Firstly, the fly connectomes, which are derived from neuron-scale EM reconstructions, are a qualitatively different kind of data from human connectomes, which are derived from DSI imaging of large-scale tracts. Likewise, calcium data and fMRI data are very different functional data acquisition methods-the fact that similar processing steps can be used on time-series data does not make them surprisingly similar, and does not in my view constitute evidence of "similar design concepts."

We are not saying we have presented evidence; we are merely saying that it is possible that it is a similar design concept. That is, the fundamental calculation is performed analog through local addition, subtraction, and multiplication by multiple synapses, and when it is necessary to transmit the data over a distance, it is converted to digital by convolution. This is the idea of an analogue-digital-hybrid circuit (Shirahata et al., Euro. J. Neurosci., 2025). Although C. elegans lacks sodium channels and most of neurons are non-spiking, AWA cell has been shown to exhibit action potential mediated by calcium spikes (Liu et al., Cell, 2018). This suggests that even C. elegans utilize an analog-digital-hybrid circuit. ‘FC’ is correlation between local activity of synapses, and FC-SC showed a high correlation between regions. This result was similar in both flies and mammals. In other words, functions are implemented locally (analog) and connected over a wide area (digitally). We believe this design is common across multiple species, even at different brain size scales. To address your concern, we modified the text on line 320 to give a more conservative tone.

The comparison of FlyEM/FlyWire connectomes concludes that differences are more likely a result of data processing than of inter-individual variability. If this is the case, the title should not claim that the manuscript covers individual variability.

The analysis of the wedge-AVLP neuron strikes me as highly speculative, given that the alignment precision between the connectome and the functional data is around 5 microns (Brezovec* et al, PNAS 2024).

Thank you for pointing that out. We have changed the title because the original title was considered too broad and general.

Recommendations for the authors:

Reviewer #3 (Recommendations for the authors):

(1) I still feel that the title and introduction overstate the results, despite the small changes to the introduction.

Thank you for pointing that out. We have changed the title because the original title was considered too broad and general.

(2) Re: the use of a custom FC-SC score in this manuscript, I agree with the authors that the overall approach is common. However, I would like to see some evidence that the qualitative conclusions in the paper are robust to the way this score has been designed, particularly because it deviates from metrics used in prior work.

In relation to this, we also performed an additional analysis of the FC-SC correlation using the Gaussian resampling method (Honey et al., 2009) [6]. Compared to the log transform of the SC matrix, Gaussian resampling showed a higher correlation when the ROI count was 200 or less, and as the ROI count increased and the Sparsity Rate rose, the log transform showed a higher correlation (Ext.Data.Fig.2-2a and b). However, in both methods we still observed a tendency for the FC-SC correlation to decrease as the ROI count increased. This content has been added to the main text (lines 132-137).

eLife Assessment

This valuable study compares orthogonal approaches for detecting RNA chemical modifications and provides a helpful framework for improving the reliability of RNA modification identification by direct RNA sequencing. The evidence supporting the technical benchmarking claims is solid, with multiple complementary controls, including in vitro-transcribed RNA, chemical inhibition of METTL3, and orthogonal m6A mapping. However, claims regarding relative method accuracy and cell-type- and isoform-specific biological differences remain less well supported due to limited biological replication and validation.

Reviewer #1 (Public review):

[Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

The authors have substantially strengthened the manuscript in response to my comments. In particular, they have improved the treatment of non-canonical calls, added further illustrative examples at the level of individual RNA variants, repeated pathway enrichment using an appropriate expression-matched background, assessed treatment-associated cell viability, and moderated several biological claims. The revised manuscript is now most compelling as a technical benchmarking and best-practice resource for high-confidence prediction of a common RNA chemical modification from nanopore direct RNA sequencing.

The matched-input comparison of two methods, use of unmodified RNA controls, and inhibitor treatment are important strengths. Together, these experiments provide solid support for the central conclusion that careful filtering and appropriate controls can substantially reduce likely false-positive calls and yield a reproducible set of high-confidence predictions. The added analyses also provide useful illustrations of the capacity of long-read sequencing to assign predictions to individual RNA variants, although they do not establish broad superiority over the short-read comparison method.

Reviewer #2 (Public review):

The revised manuscript has improved substantially. The authors have addressed the main concerns raised during the previous round by introducing an IVT-based calibration framework, evaluating background calls across probability thresholds, examining basecalling mismatches, and replacing "stoichiometry" with the more appropriate term "predicted modification proportion." The expanded discussion of non-DRACH predictions is also useful and appropriately balanced. Overall, the study now provides a valuable practical framework for Dorado-based RNA modification analysis.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

The authors have substantially strengthened the manuscript in response to my comments. In particular, they have improved the treatment of non-canonical calls, added further illustrative examples at the level of individual RNA variants, repeated pathway enrichment using an appropriate expression-matched background, assessed treatment-associated cell viability, and moderated several biological claims. The revised manuscript is now most compelling as a technical benchmarking and best-practice resource for high-confidence prediction of a common RNA chemical modification from nanopore direct RNA sequencing.

The matched-input comparison of two methods, use of unmodified RNA controls, and inhibitor treatment are important strengths. Together, these experiments provide solid support for the central conclusion that careful filtering and appropriate controls can substantially reduce likely false-positive calls and yield a reproducible set of high-confidence predictions. The added analyses also provide useful illustrations of the capacity of long-read sequencing to assign predictions to individual RNA variants, although they do not establish broad superiority over the short-read comparison method.

A few claims remain insufficiently supported. First, the discussion of non-DRACH sites should cite Cancer mutations rewire the RNA methylation specificity of METTL3-METTL14 (Science Advances, DOI: 10.1126/sciadv.ads4750). In particular, Figure 3C of that study shows that the normal METTL3-METTL14 enzyme retains measurable activity on non-DRACH substrates; this provides a useful precedent for describing the inhibitor-sensitive GGAUU calls as plausible candidates rather than treating non-DRACH calls, especially those detected only by GLORI, as presumptively artefactual. The revised manuscript appropriately notes technical and biological explanations for non-canonical calls and identifies the GGAUU sites as candidates requiring further validation, but this additional context would improve interpretation.

We thank the reviewer for reminding us of this interesting publication. We have included the citation and adjusted our discussion accordingly (lines 576-579).

Second, the comparison between Dorado and GLORI remains overinterpreted. The matched-input design and inhibitor perturbation are important strengths, but the current analysis does not establish that sites detected only by GLORI are false positives or that Dorado is a reference standard. This would require, for example, direct comparison of inhibitor responsiveness among GLORI-only and shared sites, while accounting for coverage, conversion efficiency, and other method-specific biases; therefore, phrases such as "GLORI exhibits reduced specificity," "largely spurious," and "validated by an orthogonal approach" are not fully supported by the data presented.

Thank you for this comment. We agree that our original language potentially overinterpreted the GLORI-Dorado comparison. In our GLORI dataset, the relatively high gene-level non-conversion rate (>5%) particularly limits interpretation of low-proportion sites. Although this background has little effect at high m⁶A/A proportions, it can substantially affect calls in the 10-20% range. Consistent with this limitation, GLORI-only sites were predominantly low-stoichiometry and more often occurred outside DRACH motifs, whereas shared sites were enriched at higher predicted m⁶ A/A proportions. GLORI coverage was similar between categories, and all sites analyzed were STM2457-responsive. However, residual non-conversion may disproportionately affect low-fraction calls. Thus, our findings do not establish that GLORI-only sites are false positives but instead indicate unresolved method-specific differences. We have adjusted the section accordingly (lines 327-382) and added two additional subplots in Fig. S21 (lines 1193-1202 contain the new figure legends).

Similarly, the thresholding approach should not be described as "unbiased." The analysis is useful and empirically informed, but it relies on selected parameters, a manually specified penalty, and predetermined decision criteria. Describing the procedure as "data-driven" or "empirically calibrated against matched unmodified RNA controls" would more accurately convey both its value and its limitations.

Thank you for this comment. We agree that this language should be used carefully. To our knowledge, we have already refined the relevant wording in the current revised version.

The additional RNA-variant examples are helpful, but they demonstrate an advantage of long-read data rather than broadly proving that Dorado "outperforms" GLORI for RNA-variant-level quantification. A more accurate formulation would be: "Direct RNA sequencing provides RNA-variant-resolved predictions that are not readily assignable using short-read GLORI data, as illustrated at several loci."

We have revised the wording to emphasize that DRS enables RNA-variant-resolved predictions that are not readily assignable using short-read GLORI data (lines 384-385 and 404-406).

Finally, the authors now acknowledge that the functional significance of cell-type differences remains unresolved, but the abstract still presents the results as evidence of "finely tuned" regulation, and the discussion links them to neuronal function. The data provide solid evidence for differences in predicted modification proportions between cell types, but the evidence is incomplete for claims about functional consequences; these statements should therefore be framed as context or hypotheses for future study rather than as inferences from the present experiments.

Thank you for this comment. In the revised manuscript, we do not describe the observed cell-type differences as evidence of “finely tuned” regulation or infer functional consequences for neuronal function. Any potential relevance is framed only as context for future investigation. We refer to lines 43-46 and 602-606, where we describe our observations about increased m⁶A/A site ratios in HD10.6 cells compared to NHDFs.

Reviewer #2 (Public review):

The revised manuscript has improved substantially. The authors have addressed the main concerns raised during the previous round by introducing an IVT-based calibration framework, evaluating background calls across probability thresholds, examining base calling mismatches, and replacing "stoichiometry" with the more appropriate term "predicted modification proportion." The expanded discussion of non-DRACH predictions is also useful and appropriately balanced. Overall, the study now provides a valuable practical framework for Dorado-based RNA modification analysis.

I have only a few remaining comments, all of which should be addressable through clarification and minor textual revisions.

(1) The comparison between NHDF and HD10.6 cells does not appear to include independent biological replicates. The observed differences in predicted m6A/A proportions should therefore be presented as differences between the two datasets rather than as general cell-type-specific effects. The interpretation is already cautious in the Discussion, but similar wording should be used in the Abstract, Results heading, and title if applicable.

We thank the reviewer for this suggestion and agree that the wording should avoid overinterpretation. We consider the title appropriate, as the manuscript reports differences in predicted m⁶ A proportions between fibroblast and neuronal datasets. However, to ensure greater precision, we revised the relevant sentence in the Abstract (lines 43-46) and two Results section headings (lines 130-131 and 408-409).

(2) The direction of the comparison in Figure 3D should be clarified. The percentages represent the fraction of GLORI sites that is also detected by Dorado. Thus, the analysis shows that Dorado increasingly recovers high-proportion GLORI sites. It does not directly demonstrate that high-proportion Dorado predictions are increasingly validated by GLORI. The text should reflect this distinction, and the figure should clearly identify which circle represents each method.

We thank the reviewer for this clarification. We revised the text in lines 367-369 to clarify that Dorado increasingly recovers high-proportion GLORI sites, and we adjusted Fig. 3D to improve clarity.

(3) The ratio of calls detected in IVT relative to biological samples is a useful empirical measure of background, but it is not formally a false-positive rate because the true status and testability of every site are not known. A term such as "IVT-to-sample call ratio" or "empirical background-call ratio" would be more precise and would not alter the main conclusion regarding the value of stringent filtering.

The reviewer’s point is well taken, and we have replaced the use of false-positive with ‘background signal’ throughout the manuscript when dealing with our approach and results. We have however retained the use of false-positive as a term when directly quoting from other published studies (regardless of whether they used this term correctly).

(4) The description of the change-point framework still contains the placeholder "at least X-fold more signal." The actual enrichment threshold should be provided. The authors should also ensure that a log2 ratio threshold of 1.5 is not described as 1.5-fold enrichment, since these values are not equivalent.

We thank the reviewer for catching this. We have corrected the sentence on lines 739-740 to note that the fold enrichment is 1.5-fold which equates to a log2 ratio threshold of ~0.585.

These are relatively minor issues. The authors have responded constructively to the previous reviews, and I do not consider further experimental work necessary.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

(1) Add the Science Advances reference, DOI: 10.1126/sciadv. ads4750, to the discussion of non-DRACH sites and use it to contextualize inhibitor-sensitive GGAUU sites as plausible candidates requiring independent validation.

(2) Revise language implying that GLORI-only sites are false positives, that GLORI has reduced specificity, or that Dorado serves as the validating reference method; instead, describe the observed agreement and acknowledge unresolved method-specific differences.

(3) If feasible, compare inhibitor responsiveness directly among shared and GLORI-only sites, reporting coverage and conversion-related quality measures for those categories.

(4) Replace "unbiased" thresholding language with "data-driven" or "empirically calibrated against matched unmodified RNA controls."

(5) Revise broad "outperforms" language about RNA-variant-level quantification to specify that long-read sequencing enables predictions at individual RNA variants in the illustrated examples.

(6) Remove or further qualify statements that infer finely tuned regulation or neuronal functional effects from differences in predicted modification proportions alone.

We have addressed these recommendations as described above.

eLife Assessment

This study reports the relative importance of Tie1 and Tie2 signaling for atrial versus ventricular trabeculation. It is an important study and is one of the few works to date that have carefully and simultaneously analyzed these two processes. In line with a previous study in zebrafish, the authors demonstrate key differences between atrial and ventricular trabeculation. While the imaging and quantitative analyses are conducted using solid and validated methodologies throughout the manuscript, future studies using additional experimental approaches will be invaluable to further establish the underlying mechanisms.

Reviewer #1 (Public review):

Summary:

In this manuscript, Ding et al. use genetic mouse models to demonstrate that atrial trabeculation is more dependent on Tie1/Tie2 signaling than ventricular trabeculation. With additional experimentation that would support the current claims, the results may hold significant value, as atrial trabeculation remains an understudied phenomenon in cardiac biology with potential implications for atrial cardiomyopathy and atrial fibrillation.

Strengths:

Detailed characterization of atrial versus ventricular trabeculation across different developmental timepoints, and the use of appropriate animal models to address the scientific question at hand.

Comments on revised version.

I thank the authors for their detailed comments on the reviewers' comments. There however remain some issues that should be properly addressed to make the manuscript stronger.

(1) The authors state that Tie1 and Tek levels are noticeably higher in atrial vs ventricular endocardial cells. Can the authors provide the appropriate statistics to give strength to this claim, since the Supplemental Figure 1F does not indicate robust chamber-specific differences? As normal statistics (t-test, Mann-Whitney, ANOVA) cannot be applied to raw RNAseq data, the authors should collect independent samples and perform qRT-PCR analyses on these.

(2) When Tek is depleted in endothelial cells at E12.5, ventricular defects are also observed at later stages. Presumable, this could mean that earlier depletion of Tek, prior to initiation of ventricular trabeculation, might have resulted in equally severe atrial and ventricular defects. Similarly, they observe that, when Tie1 is deleted early on (tamoxifen administration at E10.5), combined with the heterozygous deletion of Tek, both atrial and ventricular defects are observed. While their postnatal studies do suggest a differential impact of Tie2 on atrial vs ventricular trabeculation (or the maintenance thereof), a more thorough analysis is needed to truly support the authors' claim that Tie1 and Tie2 synergistically impact atrial trabeculation more dramatically than ventricular trabeculation. However, to give weight to this claim, they should perform EC-specific depletion of Tie1 and Tie2 (or at least 3 out of 4 alleles) at different developmental stages, where tamoxifen is administered prior to ventricular or atrial trabeculation.

(3) Earlier reports indicate that a part of Tie1 null mutants (app. 15%) die in utero. Did the authors also observe this phenomenon in their mutant lines, and did these animals have more severe atrial defects?

(4) Although the reviewers did not request detailed mechanistic insights into the molecular mechanisms that drive Tie1-mediated atrial trabeculation, several of the reviewers' requests (e.g. YAP/TAZ staining) were unfortunately not performed. This is a shame, as peer review is instrumental to strengthen manuscripts and increase their validity. It would have therefore been nice if the authors had at least made an attempt to address some of outstanding reviewers' questions.

Reviewer #2 (Public review):

Ding et al. examine the role of TIE1 in cardiac chamber morphogenesis using genetic mouse models targeting Tie1, Tek, or both, and analyzing endocardial cell-mediated chamber formation across multiple embryonic developmental and postnatal stages, supported by analysis of published single cell datasets and new bulk RNA seq analyses of murine cardiac tissue. The authors find that Tie1 and Tek expression is higher in atrial than ventricular endocardial cells. Notably, endothelial TIE1 is required for atrial trabeculation at E12.5, but is less critical in ventricular trabeculation. TIE1 also acts synergistically with TIE2 during atrial trabeculation. While Tie1 deficiency alone does not cause defects at E10.5, combined heterozygous deletion of Tek disrupts both atrial and ventricular development at E10.5. This synergy is further supported by analyses at later embryonic stages and in postnatal hearts.

Strengths:

The study is well designed, clearly written, and supported by high-quality figures. The performed experiments demonstrate a previously unrecognized role for TIE1 in cardiac development and identify synergistic control of cardiac morphogenesis by TIE1 and TIE2. This synergy is consistent with the previously identified roles of TIE1 and TIE2 in venous development and with TIE1 involvement in angiopoietin-dependent postnatal vascular and lymphatic remodeling. Together, these findings support a role for TIE1 as a contributor to ANGPT1-TIE2 signaling during heart development.

Weaknesses:

The manuscript does not include direct mechanistic studies; however, RNA seq analysis of atria and ventricles showed reduced expression of Tek, Dll1, and Notch1 upon Tie1 deficiency in developing hearts. Although no direct mechanistic experiments are performed, potential mechanisms based on previously published work such as TIE1-TIE2 heterodimer formation, effects on endothelial junctions, migration, survival and venous development are discussed in the context of TIE1-TIE2 synergy.

Reviewer #3 (Public review):

Summary:

Ding et al. investigate the roles of TIE1 and TEK (Tie2) in mouse cardiac development, with a particular focus on atrial trabeculation. The authors employ multiple genetic models, including Tie1ICDflox/flox (with Cdh5-CreERT2), a knockout-first allele (EUCOMM, Tie1 tm1a/tm1a), and a Tek deletion model.

Based on the dataset from Feng et al. 2022 Nat Commun, the authors report increased expression of Tie1 and Tek transcripts in atrial endocardial cells compared to ventricular cells at embryonic day (E) 14.5. Loss of Tie1 leads to early atrial trabeculation defects detectable at E12.5, whereas ventricular defects appear later and are less pronounced at E14.5. Chamber-specific RNA sequencing reveals stronger transcriptional changes in atrial tissue.

Conditional deletion of Tek results in a similar phenotype, with more pronounced atrial defects. Combined deletion of Tie1 and Tek (Tie1 ΔICD/ΔICD; Tek+/-) leads to earlier and more severe defects in both atrial and ventricular trabeculation and results in embryonic lethality around E12.5, suggesting a synergistic interaction between the two genes.

Conditional endothelial deletion of Tie1 combined with heterozygous global Tek at later embryonic stages allows analysis at later time points and again shows more severe defects in atrial trabeculation. Postnatal analysis of this model reveals reduced heart-to-body weight ratios and potential mild atrial abnormalities.

Strength:

(1) The authors address chamber-specific signaling mechanisms underlying atrial versus ventricular trabeculation, an area of high developmental and clinical relevance.

(2) The study provides a comprehensive temporal analysis across multiple embryonic stages.

(3) The use of multiple genetic models strengthens the overall conclusions and allows comparative interpretation.

(4) While focusing on trabeculation, the authors also include observations on coronary vessel development, increasing the broader relevance of the work. The findings are therefore of interest to the wider cardiovascular research community.

Weakness:

(1) Limited mechanistic insight<br /> While the authors aim to investigate underlying mechanisms, the current study is largely descriptive and based on mRNA expression and genetic interaction analyses (Tie1/Tek co-deletion). Direct mechanistic insights into signaling pathways remain limited. However, the dataset provides a valuable foundation for future mechanistic studies, which should be more clearly acknowledged in the discussion.

Author response:

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

eLife Assessment:

This study reports the relative importance of Tie1 and Tie2 signaling for atrial versus ventricular trabeculation. It is an important study and is one of the few works to date that have carefully and simultaneously analyzed these two processes. In line with a previous study in zebrafish, the authors demonstrate key differences between atrial and ventricular trabeculation. While the imaging and quantitative data were conducted with solid and validated methodology throughout the manuscript, the work would benefit from more rigourous approaches where Tie1/2 signaling is disrupted prior to the onset of atrial/ventricular trabeculation, to allow for a more direct comparison.

We thank the editors for the eLife assessment. We would like to request that the following statement be modified: “…the work would benefit from more rigourous approaches where Tie1/2 signaling is disrupted prior to the onset of atrial/ventricular trabeculation, to allow for a more direct comparison”. We request this change for the following reasons:

We utilized two distinct genetic mouse models in this study (as summarized in Fig. 7I), comprising conventional knockouts (Tie1^tm1a/tm1a, Tie1^ΔICD/ΔICD and Tie1^ΔICD/ΔICD;Tek^+/-) and inducible gene deletion models (Tek^iECKO, Tie1ICD^iECKO, and Tie1ICD^iECKO;Tek^+/-) [1-3].

The Tie1^tm1a/tm1a line is equivalent to the previously published Tie1^-/- mouse line as demonstrated in our prior work and by others [1, 4-6]. For the Tie1^ΔICD/ΔICD line, as described in the previous report [2], the expression level of Tie1^ΔICD decreased dramatically compared with that of wild-type mice, which may result from the nonsense-mediated mRNA decay. Based on the lymphatic and venous phenotypes [1, 2], we conclude that the Tie1^ΔICD/ΔICD line is almost equivalent to Tie1 null mice.

Therefore, the Tie1 or Tek alleles were inactivated prior to the onset of atrial and ventricular trabeculations, as shown in Fig. 1, Fig. 2, Fig. 3, Fig. 5A-D, and Supplemental Fig. 3. Based on these findings, we propose that TIE1 is differentially required for atria versus ventricle morphogenesis, and acts synergistically with TIE2 during cardiac trabeculation.

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this manuscript, Ding et al. use genetic mouse models to demonstrate that atrial trabeculation is more dependent on Tie1/Tie2 signaling than ventricular trabeculation. With additional experimentation that would support the current claims, the results may hold significant value, as atrial trabeculation remains an understudied phenomenon in cardiac biology with potential implications for atrial cardiomyopathy and atrial fibrillation.

Strengths:

Detailed characterization of atrial versus ventricular trabeculation across different developmental time-points, and the use of appropriate animal models to address the scientific question at hand.

Weaknesses:

The authors have consistently treated mice with tamoxifen after ventricular, but not atrial, trabeculation has already started. As such, the observed cardiac phenotypes - where predominantly atrial trabeculation is affected - might be a mere consequence of the precise time window in which Tie1/2 signaling was impaired, rather than a direct measurement of its relative importance for atrial versus ventricular trabeculation. The conclusions of the paper may thus be significantly strengthened by depleting Tie1/2 signaling prior to the onset of ventricular trabeculation, as is done for atrial trabeculation.

We thank the reviewer for the comments.

Regarding the timeline of gene deletion and tamoxifen treatment, we would like to provide the following clarification.

Fig. 1-3: As described in the Methods and Materials, Tie1^tm1a/tm1a is a knockout first mouse model established from EUCOMM embryonic stem cells (EPD0735-3B07) targeting Tie1 gene. Therefore, the Tie1^tm1a/tm1a line is equivalent to the previously published Tie1 null mice (Tie1^-/-). The Tie1^Flox/Flox mouse line (with exon 8 floxed) was generated when the lacZ reporter and neo-cassette were excised using the FLPeR mice.

Fig. 5A-D: To investigate the synergy of TIE1 and TIE2 in cardiac trabeculation, we utilized the Tek^+/- and Tie1^ΔICD/+ mouse lines and they were crossbred to generate double mutant mice harboring a homozygous Tie1 mutation and a single null Tek allele (Tie1^ΔICD/ΔICD;Tek^+/-). Although no obvious defects were observed in atrial or ventricular structures following Tie1 deficiency alone at E10.5, both atrial and ventricular development were disrupted in Tie1^ΔICD/ΔICD;Tek^+/- mutants at the same stage (Fig. 5A-D).

Supplemental Fig. 3: To verify the role of TIE1 in atrial development, we employed alternative knockout mouse line targeting the Tie1 intracellular domain by floxing exons 15 and exon 16 (Tie1ICD^Flox/Flox). Mutants harboring these null alleles are designated as Tie1^{ΔICD/ ΔICD}. As detailed in the previous publication, the line is also equivalent to the previously published Tie1 null mice (Tie1^-/-). The cardiac phenotypes shown in Supplemental Fig. 3 are indeed similar to those of Tie1^tm1a/tm1a mutant mice.

For the inducible knockouts targeting Tie1, Tek and both, the results are shown in Fig. 4, Fig. 5E-H, Fig. 6, and Fig. 7.

Fig. 4: As mice homozygous for Tek mutation (Tek^-/-) die before E10.5 [3, 7], we performed studies using the inducible knockout line targeting Tek (Tek^Flox/-;Cdh5-Cre^ERT2 named as Tek^iECKO), as shown in Fig. 4.

Fig. 5-7: To investigate the synergy of TIE1 and TIE2 in the cardiac trabeculation at the later stages of embryogenesis (Fig. 5E-H, Fig. 6) and the postnatal stage (Fig. 7), we used the inducible knockout models targeting Tie1/Tek, including Tie1ICD^iECKO (Tie1ICD^Flox/-;Cdh5-Cre^ERT2) and Tie1ICD^iECKO;Tek^+/- (Tie1ICD^Flox/-;Cdh5-Cre^ERT2;Tek^+/-).

Reviewer #2 (Public review):

Summary:

Ding et al. examine the role of TIE1 in cardiac chamber morphogenesis using genetic mouse models targeting Tie1, Tek, or both, and analyzing endocardial cell-mediated chamber formation across multiple embryonic developmental and postnatal stages, supported by analysis of published single-cell datasets and new bulk RNA seq analyses of murine cardiac tissue. The authors find that Tie1 and Tek expression is higher in atrial than ventricular endocardial cells. Notably, endothelial Tie1 is required for atrial trabeculation at E12.5, but is less critical in ventricular trabeculation. TIE1 also acts synergistically with TIE2 during atrial trabeculation. While Tie1 deficiency alone does not cause defects at E10.5, combined heterozygous deletion of Tek disrupts both atrial and ventricular development at E10.5. This synergy is further supported by analyses at later embryonic stages and in postnatal hearts.

Strengths:

The study is well-designed, clearly written, and supported by high-quality figures. The performed experiments demonstrate a previously unrecognized role for Tie1 in cardiac development and identify synergistic control of cardiac morphogenesis by Tie1 and Tie2. This synergy is consistent with the previously identified roles of Tie1 and Tek in venous development and with Tie1 involvement in angiopoietin-dependent postnatal vascular and lymphatic remodeling. Together, these findings support a role for Tie1 as a contributor to Ang1-Tie2 signaling during heart development.

Weaknesses:

The manuscript does not include direct mechanistic studies; however, RNA seq analysis of atria and ventricles showed reduced expression of Tek, Dll1, and Notch1 upon Tie1 deficiency in developing hearts. Although previously reported mechanisms, such as TIE1-TIE2 heterodimer formation and effects on endothelial junctions, migration, or survival are discussed, no direct mechanistic experiments are performed. Addressing some of these mechanisms would have clarified the basis of Tie1-Tie2 synergy. As two distinct Tie1 models are used, including one targeting the kinase domain, the authors should state whether phenotypes differed or were similar between models.

We thank the reviewer for this comment. In this study, we provide genetic evidence that TIE1 is differentially required for atrial versus ventricular trabeculation. Although the precise molecular mechanisms underlying TIE1 function warrant further investigation, we have provided compelling genetic evidence of its synergistic role with TIE2 during this process. Crucially, the two genetic models targeting Tie1 (Tie1^tm1a/tm1a, Tie1^ΔICD/ΔICD) produced consistent cardiac and vascular phenotypes, as shown in both this study and our previous work [1, 2].

Regarding endocardial TIE1 in atrial morphogenesis, fully elucidating its downstream mechanism may require the isolation of primary endocardial cells for further study. However, our previous work using HUVECs have demonstrated that TIE1 acts synergistically with its homologue TIE2 to regulate AKT-mediated COUP-TFII protein stability [1, 3]. The essential roles of COUP-TFII in venous and atrial development have been well-documented in the literature [8, 9]. We have expanded our Discussion section to reflect these mechanistic insights.

Reviewer #3 (Public review):

Summary:

Ding et al. investigate the roles of TIE1 and TEK (Tie2) in mouse cardiac development, with a particular focus on atrial trabeculation. The authors employ multiple genetic models, including Tie1ICDflox/flox (with Cdh5-CreERT2), a knockout-first allele (EUCOMM, Tie1 tm1a/tm1a), and a Tek deletion model.

Based on the dataset from Feng et al. 2022 Nat Commun, the authors report increased expression of Tie1 and Tek transcripts in atrial endocardial cells compared to ventricular cells at embryonic day (E) 14.5. Loss of Tie1 leads to early atrial trabeculation defects detectable at E12.5, whereas ventricular defects appear later and are less pronounced at E14.5. Chamber-specific RNA sequencing reveals stronger transcriptional changes in atrial tissue.

Conditional deletion of Tek results in a similar phenotype, with more pronounced atrial defects. Combined deletion of Tie1 and Tek (Tie1 ΔICD/ΔICD; Tek+/-) leads to earlier and more severe defects in both atrial and ventricular trabeculation and results in embryonic lethality around E12.5, suggesting a synergistic interaction between the two genes.

Conditional endothelial deletion of Tie1 combined with heterozygous global Tek at later embryonic stages allows analysis at later time points and again shows more severe defects in atrial trabeculation. Postnatal analysis of this model reveals reduced heart-to-body weight ratios and potential mild atrial abnormalities.

Strengths:

(1) The authors address chamber-specific signaling mechanisms underlying atrial versus ventricular trabeculation, an area of high developmental and clinical relevance.

(2) The study provides a comprehensive temporal analysis across multiple embryonic stages.

(3) The use of multiple genetic models strengthens the overall conclusions and allows comparative interpretation.

(4) While focusing on trabeculation, the authors also include observations on coronary vessel development, increasing the broader relevance of the work. The findings are therefore of interest to the wider cardiovascular research community.

Weaknesses:

(1) Timing of recombination vs. trabeculation onset

Ventricular trabeculation begins earlier than atrial trabeculation. Since tamoxifen (in contrast to 4-hydroxytamoxifen) requires metabolic activation, Cre-mediated recombination will occur with a delay. This suggests that atrial trabeculation may be targeted before its onset, whereas ventricular trabeculation may already be underway for 2-3 days at the time of effective gene deletion.

How do the authors account for this discrepancy in their interpretation?

Have earlier induction time points been tested to better capture the onset of ventricular trabeculation? This limitation should be explicitly discussed.

(2) Clarity of genetic models and experimental design

The study employs several genetic constructs. It would improve clarity if, for each experiment, the specific genetic model and tamoxifen regimen were clearly described before presenting the results.

We thank the reviewer for the detailed and constructive comments. As detailed above, for studies employing the inducible gene deletion mouse models (models (Tek^iECKO, Tie1ICD^iECKO, and Tie1ICD^iECKO;Tek^+/-), the genetic models and tamoxifen treatment schemes have been provided in the related figures.

For the investigation of TIE1 or TIE2 in atria versus ventricle morphogenesis at earlier embryonic stages, we utilized the conventional knockouts targeting Tie1 (Tie1^tm1a/tm1a, Tie1^ΔICD/ΔICD) or the combined deletion of Tie1 and Tek (Tie1^ΔICD/ΔICD;Tek^+/-), as detailed above.

(3) Tie1 tm1a/tm1a phenotype vs. known global knockout

Previous studies (PMID: 8846781, 7596437) show that complete Tie1 loss leads to severe edema, vascular rupture, and embryonic lethality around E13.5-E14.5.

How does the Tie1 tm1a/tm1a allele differ, given that animals appear to survive longer? Is this allele hypomorphic rather than a full knockout?

This point requires clarification.

Tie1^tm1a/tm1a is equivalent to the full Tie1 knockout (Tie1^-/-). As demonstrated in our prior work, the Tie1^ΔICD/ΔICD model produced lymphatic and blood vascular phenotypes similar to those of Tie1^-/- mutants [1, 2, 5, 6].

(4) Limited mechanistic insight

While the authors aim to investigate underlying mechanisms, the current study is largely descriptive and based on mRNA expression and genetic interaction analyses (Tie1/Tek co-deletion). Direct mechanistic insights into signaling pathways remain limited. However, the dataset provides a valuable foundation for future mechanistic studies, which should be more clearly acknowledged in the discussion.

We thank the reviewer for the comment. We have incorporated this discussion into the revised manuscript

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

The conclusions of the paper may be significantly strengthened by addressing several of the following comments:

Major critiques:

The authors state that Angpt1 is more prominently expressed in atrial versus ventricular myocardium and that this might explain the larger impact of Tie1 deletion on atrial versus ventricular trabeculation. However, other ligands for Tie1/Tie2 have been described, such as Svep1, Angpt2, and Lect2. It would be great if the authors could show the relative expression of these ligands in atrial versus ventricular cardiomyocytes over time.

We appreciate the reviewer's valuable suggestion. To address this, we analyzed the scRNA-seq dataset to compare the expression profiles of key ligands (Angpt1, Angpt2, Svep1, and Lect2) between atrial and ventricular cardiac cells during embryonic development. These new data have been integrated into Supplemental Fig. 1F. Our analysis reveals that Angpt1 is differentially enriched in atrial cells relative to ventricular cells. However, no distinct inter-chamber differences were detected for Svep1 or Angpt2. Notably, Lect2 transcripts were virtually absent in both chambers; this is fully congruent with our bulk RNA-seq data, which showed no detectable Lect2 expression in whole-heart lysates.

Likewise, the authors indicate that ANGPT1-TIE2 signaling is stronger in atria than ventricles at E14.5. Could this reflect a temporal rather than spatial difference? How does ANGPT1-TIE2 signaling compare between atria and ventricles during different steps of the trabeculation process?

We thank the reviewer for this insightful question regarding spatial versus temporal dynamics. To address whether the observed differences reflect a temporal shift or a sustained spatial pattern, we evaluated Angpt1 expression across a developmental timeline spanning the major stages of trabeculation, from initiation through remodeling (E10.5 to E18.5).

As shown in the newly added Supplemental Fig. 1F, Angpt1 is consistently and differentially enriched in atrial cells relative to ventricular cells throughout all examined embryonic stages. This sustained divergence confirms that elevated ANGPT1-TIE2 signaling in the atria is a distinct spatial feature rather than a transient temporal shift. This finding aligns with our ligand-receptor interactome analysis between cardiomyocytes and endocardial cells (Supplemental Fig. 1E). Furthermore, a specialized role for ANGPT1-mediated signaling in atrial morphogenesis has been independently demonstrated by Kim et al. (Cell Rep, 2018), supporting our model that this pathway is uniquely critical for the assembly of the atrial muscular network [10].

Ding and colleagues demonstrate that loss of endothelial/endocardial Tie1 results in absent atrial trabeculae, while ventricular trabeculae are only moderately impacted. As ventricular trabeculation is initiated much earlier than atrial trabeculation, this discrepancy might, however, be secondary to the precise time point within trabecular network formation at which Tie1 was deleted. The make sure that Tie1 does not prevent ventricular trabeculation to an equal extent as atrial trabeculation, the authors should conduct additional experiments in which Tie1 is deleted before the onset of ventricular trabeculation (i.e., E8.5). Consistent with this, Tie1 knockout mice display substantial ventricular trabeculation defects (Figure 5). This comment extends to the analysis of the Tie2 knockout mice, where knocking out Tie2 starting from E12.5 does not affect ventricular trabeculation at E14.5 but impairs ventricular trabeculation when hearts are analyzed later (at E17.5). Additionally, Figure 5 clearly demonstrates that ventricular trabeculation is impacted in the combined absence of Tie1 and Tie2. Overall, a more detailed analysis of the relative importance of Tie1/Tie2 signaling for atrial versus ventricular trabeculation with more appropriate tamoxifen regimens needs to be performed to support the claims made by the authors.

While one can deduce from the Methods section that tamoxifen was administered at E10-5-E11.5 for the results depicted in Figure 1, no information is provided regarding the tamoxifen regimens for the following figures, obscuring proper interpretation of the results. The authors should clearly state the precise time window in which they depleted endothelial Tie1 for each Figure in both text and legends, as this essential information is missing in Figures 2 and 3.

We thank the reviewer for pointing this out. To rigorously investigate the differential requirement of TIE1, as well as its synergy with TIE2, during the distinct stages of atria versus ventricle trabeculation, we utilized both conventional (constitutive) and inducible gene knockout models. For the inducible genetic models, the exact timelines of gene deletion and tamoxifen treatments are detailed within their respective figure panels and legends. Below is a brief overview of the constitutive null models used in this study.

Fig. 1-3 (constitutive null models): As described in the Materials and Methods, the Tie1^tm1a/tm1a line is a “knockout-first” mouse model generated from EUCOMM embryonic stem cells (EPD0735-3B07) targeting Tie1 gene. Consequently, the Tie1^tm1a/tm1a line is functionally equivalent to previously published Tie1 null mice (Tie1^-/-). The Tie1^Flox/Flox mouse line (harboring a floxed exon 8) was generated by excising the lacZ reporter and neo-cassette using FLPeR delete mice. As described above, the Tie1^ΔICD/ΔICD line is also equivalent to Tie1 null mice.

Fig. 5A-D: To investigate the genetic synergy between TIE1 and TIE2 during the initiation phase of cardiac trabeculation, we crossbred Tek^+/- and Tie1^ΔICD/+ lines to generate compound mutants harboring a homozygous Tie1 mutation paired with a single null Tek allele (Tie1^ΔICD/ΔICD;Tek^+/-). While Tie1 deficiency alone yielded no obvious defects in atrial or ventricular structures at E10.5, compound Tie1^ΔICD/ΔICD;Tek^+/- mutants exhibited severe disruption in both atrial and ventricular development at the same developmental stage (Fig. 5A-D).

In the last paragraph, Ding et al. propose a model in which Tie1/Tie2 signaling drives atrial versus ventricular endothelial/endocardial specification. To substantiate this claim, the authors should identify several genes that differentiate between these two and demonstrate that their asymmetric expression pattern is perturbed in Tie1/2 mutant embryos. Alternatively, single-cell RNA-seq in control and Tie1/2 mutant embryos could provide a more holistic view of endothelial specification processes and their impact on myocardial shape and function.

We thank the reviewer for these constructive suggestions. We agree that a comprehensive single-cell RNA-seq analysis of Tie1/Tek mutant embryos would provide a holistic view of endocardial specification. While performing this extensive sequencing is beyond the scope of our current study, we believe our present combination of chamber-specific bulk RNA-seq and functional genetic models provides a strong foundation for these mechanisms. Further studies are certainly warranted to provide deeper single-cell insights in this direction, and we have added this point to our revised Discussion section.

A recent publication (Albu et al., Nature Communications) found no evidence for Nrg/Erbb and Notch signaling in atrial trabeculation, but reported that YAP signaling was implicated in the latter. This raises important questions: Did the authors observe Yap-mediated signaling in their RNAseq datasets? Could Yap/Taz be investigated in control and Tie1/2 mutant mice? Is there a link between Tie1/2 and Yap/Taz signaling in other systems, such as venous versus arterial endothelial specification?

We thank the reviewer for this insightful comment. Based on our bulk RNA-seq data, we did not observe an obvious change in Yap1 and Wwtr1 (Taz) transcript expression in the hearts of Tie1 null mice compared to their littermate controls. However, this transcriptomic data does not exclude the possibility of post-translational regulation occurring at the protein level, such as alterations in protein phosphorylation, nuclear translocation, or the activation of downstream YAP/TAZ target genes. We agree that investigating potential alterations in YAP signaling following the induced endothelial deletion of Tek or compound Tie1/2 represents an important direction for future study.

In contrast to the study by Albu and colleagues, the manuscript presented here observes Notch1 downregulation in the atria but not ventricles (Figures 3F, G). What could be the reason for these discrepancies? Does loss of Tie1 and/or Tie2 impair Nrg1, VEGF, or is Notch signaling more impactful in atria versus ventricles at the protein levels?

We thank the reviewer for the insightful comment. Using a Zebrofish model, Albu et al., (Nat Comm 2024) [11] reported no evidence for Nrg/ErbB or Notch signaling in atrial muscle network formation. Based on our bulk RNA-seq analysis of heart tissues from Tie1 deficient mice compared to littermate controls, we demonstrated a decrease in Erg1 and Notch1 transcripts in Tie1 mutants. In contrast, we detected an increase in Vegfa transcript levels, which may result from compensatory changes in cardiac angiogenesis induced by hypoxic stress upon TIE1 deficiency. The biological significance of these transcriptomic changes requires further investigation. We have modified the text to include this discussion in the revised manuscript.

In Figures 5 and 6, it is not clear which genotype was used as a control. Are these controls Tek+/-, Tie1ICDECKO, or WT mice? If there were any differences between these genotypes, these should be reported with the appropriate statistical tests.

We thank the reviewer for pointing this out. We have added the information in the revised manuscript. Related to Fig. 5 and Fig. 6, the genotype for the control mice used are described in the Methods and Materials section. The genotypes of the Tie1/Tek knockout and control mice are as follows: Tie1ICD^Flox/-;Cdh5-Cre^ERT2 named as Tie1ICD^iECKO, littermates Tie1ICD^Flox/+ or Tie1ICD^Flox/+;Cdh5-Cre^ERT2 mice named as Control. Tek^Flox/-;Cdh5-Cre^ERT2 named as Tek^iECKO, littermates Tek^Flox/+ or Tek^Flox/+;Cdh5-Cre^ERT2 mice named as Control. Tie1ICD^Flox/-;Cdh5-Cre^ERT2;Tek^+/- named as Tie1ICD^iECKO;Tek^+/-, littermates Tie1ICD^Flox/+;Cdh5-Cre^ERT2;Tek^+/-, Tie1ICD^Flox/+;Tek^+/-, Tie1ICD^Flox/+;Cdh5-Cre^ERT2;Tek^+/+ or Tie1ICD^Flox/+;Tek^+/+ named as Control.

Minor points:

The resolution of the retinal stainings should be improved.

We thank the reviewer for pointing this out. We performed the analysis of retinal vascular phenotype in this study, which is for the verification of gene deletion efficiency.

Sarcomeric defects have been observed in tie1-/- zebrafish. Can the authors check if such sarcomeric defects are also present in their mouse models and if they are more pronounced in atrial versus ventricular cardiomyocytes?

We thank the reviewer for the interesting comment. We have not studied whether there is any sarcomeric defect in Tie1 null mice, which requires further investigation.

If Tie1/2 signaling is more critical for atrial than ventricular development and function, is there any evidence for an association between atrial cardiomyopathy and/or atrial fibrillation in humans carrying pathogenic Tie1/Tie2 mutations?

We thank the reviewer for this insightful comment. Whether human patients carrying pathogenic Tie1/Tie2 mutations have atrial cardiomyopathy and/or atrial fibrillation warrants further investigation.

Bulk RNAseq reveals that Tie1tm1a/tm1a mice have reduced Nrg1 mRNA levels in the heart (Figure 2I), but this is not reflected in the atrial or ventricular transcriptomics (Figure 3). Could the authors explain this discrepancy in more detail?

We thank the reviewer for pointing this out. As described above, there was a decrease of Nrg1 transcript level in hearts of Tie1 mutant compared to that of the littermate controls. The biological significance of this change requires further investigation. We have included this discussion in the revised manuscript.

Reviewer #3 (Recommendations for the authors):

Unfortunately, line numbers are not provided, which makes it difficult to refer to specific passages.

We thank the reviewer for pointing this out. We have added line numbers to the revised manuscript t.

While it is appropriate to mention in the Discussion that Tie1 variants have been associated with lymphatic defects, this information should be removed from the abstract, as the manuscript does not address lymphatic biology, and this statement distracts from the main focus of the study.

We thank the reviewer for the comment. In the abstract background, we stated: “…TIE1 variants or loss of function mutations were reported in a subset of lymphedema patients, and it is unknown whether the patients have also cardiac defects in addition to the lymphatic abnormality.” By framing the introduction this way, we aimed to establish the central question of our study: whether Tie1 variants or loss-of-function mutations lead to cardiac abnormalities, which could ultimately contribute to or exacerbate systemic lymphedema alongside primary lymphatic defects.

Figure 5F should include single-channel images for Tie1 staining to allow proper assessment of signal specificity and localization.

We thank the reviewer for this comment. The co-staining of TIE1 and PECAM1 is intended primarily to demonstrate gene deletion efficiency. As shown In Fig. 5F, TIE1 is efficiently deleted in Tie1ICD^iECKO;Tek^+/- mice compared to control mice. Furthermore, as previously reported by Cao et al. [1] and Shen et al. [2], the mutant mice (Tie1ICD^iECKO;Tek^+/-) exhibit disrupted venous development alongside defective lymphatic formation.

Parts of the manuscript use non-scientific language, for example, "doubly knockout mice" or "atria appear to have some wavy edges." These should be revised for a more formal scientific tone.

We thank the reviewer to point this out. We have modified the text in the revised manuscript.

Typo error in Figure 1D: "ventricle"

We thank the reviewer to point this out. We have made the correction with Fig. 1D.

Reference

(1) Cao, X., et al., Endothelial TIE1 Restricts Angiogenic Sprouting to Coordinate Vein Assembly in Synergy With Its Homologue TIE2. Arterioscler Thromb Vasc Biol, 2023. 43(8): p. e323-e338.

(2) Shen, B., et al., Genetic dissection of tie pathway in mouse lymphatic maturation and valve development. Arterioscler Thromb Vasc Biol, 2014. 34(6): p. 1221-30.

(3) Chu, M. et al., Angiopoietin receptor Tie2 is required for vein specification and maintenance via regulating COUP-TFII. Elife, 2016. 5:e21032.

(4) Rodewald, H.R. and T.N. Sato, Tie1, a receptor tyrosine kinase essential for vascular endothelial cell integrity, is not critical for the development of hematopoietic cells. Oncogene, 1996. 12(2): p. 397-404.

(5) D'Amico, G. et al., Loss of endothelial Tie1 receptor impairs lymphatic vessel development-brief report. Arterioscler Thromb Vasc Biol, 2010. 30(2): p. 207-9.

(6) Qu, X., et al., Abnormal embryonic lymphatic vessel development in Tie1 hypomorphic mice. Development, 2010. 137(8): p. 1285-95.

(7) Dumont, D.J., et al., Dominant-negative and targeted null mutations in the endothelial receptor tyrosine kinase, tek, reveal a critical role in vasculogenesis of the embryo. Genes Dev, 1994. 8(16): p. 1897-909.

(8) You, L.R., et al., Suppression of Notch signalling by the COUP-TFII transcription factor regulates vein identity. Nature, 2005. 435(7038): p. 98-104.

(9) Wu, S.P., et al., Atrial identity is determined by a COUP-TFII regulatory network. Dev Cell, 2013. 25(4): p. 417-26.

(10) Kim, K.H., et al., Myocardial Angiopoietin-1 Controls Atrial Chamber Morphogenesis by Spatiotemporal Degradation of Cardiac Jelly. Cell Rep, 2018. 23(8): p. 2455-2466.

(11) Albu, M., et al., Distinct mechanisms regulate ventricular and atrial chamber wall formation. Nat Commun, 2024. 15(1): p. 8159.

eLife Assessment

This valuable study relating brain anatomy to reading ability in children studied longitudinally finds that a difference in sulcal anatomy is more predictive of reading ability than cognitive measures. The evidence that this anatomical difference is predictive and confers some benefit in the connectivity between brain areas is based on careful analyses and is solid. The findings would be of interest to researchers studying brain and cognitive development.

Reviewer #1 (Public review):

[Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

Summary:

There is prior literature showing a robust relationship between sulcal interruptions in the posterior occipital temporal sulcus (pOTS) and reading ability. The goals of this study were to extend these findings to children examined longitudinally as they become better readers, and to examine the underlying white matter properties in individuals with and without pOTS sulcal interruptions. To do this, the authors collected longitudinal structural, diffusion, and behavioral data in 51 children (TP1 age 5.5, TP3 age 8.2 years).

First, the authors found that the gyral gap was consistent across time within the subject. This is expected, as they state in the introduction that sulcal patterns are typically established in utero. Next, they found that children with an interrupted pOTS have higher reading scores (across a variety of measures) at timepoint (TP) 3 than children with continuous pOTS, and this was specific to the pOTS, as no associations emerged for the anterior OTS or MFS; this is again expected from prior literature. They then found that the binary presence of this gap, but not anterior OTS or MFS predicted T3 reading performance. Further, they found that a subsample of the lowest readers at TP1 did not have differences in reading score by gyral gap, but that this difference emerged at TP3. Additionally, the gyral gap at TP1 is similar to variance TOWRE 3 reading skills as some behavioral measures at TP1. Examining underlying white matter in a smaller subset of children, the authors found higher MD in children with an interrupted pOTS vs. those with a continuous pOTS, which was contrary to their hypothesis, and higher local connectivity for interrupted, aligning with their hypothesis, but this difference was no longer present when accounting for TP3 reading scores. The authors conclude that structural properties, in this case, the gyral gap, may guide neural plasticity for reading.

Strengths:

This paper has an interesting set of longitudinal data to examine the perhaps changing relationship between sulcal interruptions in the pOTS with reading scores. I commend the authors on data collection and attention to detail in the anatomical analyses.

Reviewer #2 (Public review):

Summary:

This manuscript examined the impact of sulcal morphology on reading development. A very specific feature on the ventral surface of the brain was identified, namely the presence of an interruption in the posterior portion of the left occipitotemporal sulcus (pOTS). Compared to children with a continuous pOTS, children with an interruption at age 5 years had better reading ability at age 8. This was a large effect measured in 43 children. Surprisingly, this morphological feature was a better predictor of reading ability than measures of pre-literacy cognitive skills, such as phonological awareness. The effect was tested and reproduced across several different measures of reading ability. The authors hypothesised that the mechanism underlying this benefit related to greater local connectivity, which confers a computational advantage. This was demonstrated using analysis of diffusion-weighted imaging data in 29 of the children obtained at age 8.

Strengths:

The novelty of the manuscript is threefold: (i) the measure was made in children who were pre-literate (previous work was in older children and adults); (ii) longitudinal brain imaging and behavioural data were analysed; and (iii) diffusion data were analysed to test a hypothesis about the underlying mechanism.

The manuscript is exceptionally well written. The methods are detailed and easily reproduced. The approach is thoughtful and meticulous. All possible alternatives appear to have been considered. Where possible, further analyses have been done to address these alternatives. For example, the testing of the specificity of the sulcal interruption to left pOTS was an important addition. None predicted reading skills.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

There is prior literature showing a robust relationship between sulcal interruptions in the posterior occipital temporal sulcus (pOTS) and reading ability. The goals of this study were to extend these findings to children examined longitudinally as they become better readers, and to examine the underlying white matter properties in individuals with and without pOTS sulcal interruptions. To do this, the authors collected longitudinal structural, diffusion, and behavioral data in 51 children (TP1 age 5.5, TP3 age 8.2 years).

First, the authors found that the gyral gap was consistent across time within the subject. This is expected, as they state in the introduction that sulcal patterns are typically established in utero. Next, they found that children with an interrupted pOTS have higher reading scores (across a variety of measures) at timepoint (TP) 3 than children with continuous pOTS, and this was specific to the pOTS, as no associations emerged for the anterior OTS or MFS; this is again expected from prior literature. They then found that the binary presence of this gap, but not anterior OTS or MFS predicted T3 reading performance. Further, they found that a subsample of the lowest readers at TP1 did not have differences in reading score by gyral gap, but that this difference emerged at TP3. Additionally, the gyral gap at TP1 is similar to variance TOWRE 3 reading skills as some behavioral measures at TP1. Examining underlying white matter in a smaller subset of children, the authors found higher MD in children with an interrupted pOTS vs. those with a continuous pOTS, which was contrary to their hypothesis, and higher local connectivity for interrupted, aligning with their hypothesis, but this difference was no longer present when accounting for TP3 reading scores. The authors conclude that structural properties, in this case, the gyral gap, may guide neural plasticity for reading.

Strengths:

This paper has an interesting set of longitudinal data to examine the perhaps changing relationship between sulcal interruptions in the pOTS with reading scores. I commend the authors on data collection and attention to detail in the anatomical analyses.

Weaknesses:

However, my enthusiasm was somewhat dampened after finding numerous prior publications on this very topic and I'm unclear as to how much more this paper adds to the current literature. Would we expect the existence of sulcal interruptions to be aligned with reading skills in older kids but not younger kids? Is the point to see if the interruptions exist prior to reading (but these children are not really prereaders)? What is the alternative- why would these interruptions not exist? After all, this anatomy is determined prenatally. Children who have pOTS interruptions at T1 should also have these interruptions at T3 (and indeed that is what the authors find). So how can this be the mechanism that drives plasticity? The authors also talk about the neuronal recycling hypothesis but their data cannot speak to this because they do not have fMRI data nor does their sample include only prereaders with no reading experience. The conclusions are overall overstated and not supported by the results. I think this paper could add interesting knowledge for the specific subfield of reading and the brain. However, the current state of the results, especially with the inclusion of so many trending results and the comparison of so many different processing pipelines and models, in addition to a conclusion that is not motivated by the work makes it difficult to appreciate the paper.

We thank Reviewer 1 for their detailed feedback. The reviewer raises several related concerns regarding the novelty and framing of the study, the interpretation of stable sulcal anatomy, early reading abilities, statistical robustness, and the proposed link between pOTS morphology, white matter, and reading development. In response, we have revised the manuscript to make clearer that the main contribution is not the stability of pOTS morphology itself, nor simply the replication of previously reported cross-sectional associations with reading, but rather how this stable anatomical feature relates prospectively to reading acquisition from the onset of literacy instruction. In particular, we now make more explicit that left pOTS interruption contributes to later reading beyond associations already present at age 5, including through an updated serial mediation analysis, and we better highlight how this relationship develops over time.

The reviewer’s comments also prompted us to explain more clearly how the complementary analyses in the full sample and in the subgroup with very limited initial reading skills converge to support our conclusion that left pOTS interruption is a particularly strong longitudinal predictor of later reading relative to the benchmark behavioral and familial predictors considered, despite the collinearity between pOTS interruption and emerging literacy skills at T1. We have renamed this subgroup more accurately from “pre-readers” to “initial minimal readers” and clarified its characteristics and range of reading abilities.

We also moved key analyses from the Supplement to Figure 3 to make the distinction between separate and unique longitudinal contributions easier to follow. The initial-minimal-reader analyses provide a complementary assessment in which confounding with early reading is substantially reduced. At the same time, as the reviewer accurately pointed out, some of our previous wording implied a more categorical superiority than the analyses warrant; we have therefore softened these claims throughout the manuscript while preserving the conclusion that pOTS interruption is among the strongest longitudinal contributors considered. We also added, in response to Reviewer 2, a new systematic analysis supporting the relative cognitive specificity of the pOTS effect.

Finally, we made several interpretations more precise, notably by presenting the white-matter findings as a candidate pathway rather than a definitive mechanism and by removing implications that our data directly test the neuronal recycling hypothesis. We believe these revisions make the longitudinal contribution, scope, and significance of the findings considerably clearer, and we are grateful to the reviewer for prompting these improvements.

Reviewer #2 (Public review):

Summary:

This manuscript examined the impact of sulcal morphology on reading development. A very specific feature on the ventral surface of the brain was identified, namely the presence of an interruption in the posterior portion of the left occipitotemporal sulcus (pOTS). Compared to children with a continuous pOTS, children with an interruption at age 5 years had better reading ability at age 8. This was a large effect measured in 43 children. Surprisingly, this morphological feature was a better predictor of reading ability than measures of pre-literacy cognitive skills, such as phonological awareness. The effect was tested and reproduced across several different measures of reading ability. The authors hypothesised that the mechanism underlying this benefit related to greater local connectivity, which confers a computational advantage. This was demonstrated using analysis of diffusion-weighted imaging data in 29 of the children obtained at age 8.

Strengths:

The novelty of the manuscript is threefold: (i) the measure was made in children who were pre-literate (previous work was in older children and adults); (ii) longitudinal brain imaging and behavioural data were analysed; and (iii) diffusion data were analysed to test a hypothesis about the underlying mechanism.

The manuscript is exceptionally well written. The methods are detailed and easily reproduced. The approach is thoughtful and meticulous. All possible alternatives appear to have been considered. Where possible, further analyses have been done to address these alternatives. For example, the testing of the specificity of the sulcal interruption to left pOTS was an important addition. None predicted reading skills.

Weaknesses:

The correlation of the interruption with all kinds of literacy measures and in particular reading comprehension and then PIQ suggest this interruption might confer a more general cognitive advantage rather than specifically a reading one.

It would be interesting to know if the anatomical difference predicts any other cognitive ability or if there might be any cognitive cost (a negative correlation) of this sulcal interruption.

The location of the interruption in the sulcus is quite variable and in some cases, there is more than one interruption. The sample size is probably not big enough to compare these different patterns or to evaluate the influence of the location of the sulcal interruption.

The sample is quite high-functioning and the generalisability of the findings outside of this specific population is inevitably limited.

We would like to thank Reviewer 2 for their positive feedback and for their insightful suggestions.

Cognitive specificity: We agree with the reviewer that the association between left pOTS interruption and performance IQ, as measured in the concept formation learning task, on top of its association with multiple measures of literacy, raises the possibility that this sulcal feature relates to general cognition more broadly. The fact that reading and general cognitive abilities are highly correlated and mutually influence one another over development complicates any attempt to interpret the effect of left pOTS interruption as purely reading-specific. Our results however support a relative specificity of this sulcal feature to reading: left pOTS interruption remains a strong longitudinal predictor of later reading, even when controlling for the moderate association with performance IQ, and its impact on reading skills is larger than on pIQ. Importantly, a new systematic analysis of associations with all the cognitive and academic skills obtained at T3 further supports a relative cognitive specificity to reading: We show that while left pOTS interruption is also significantly associated with several assessments of math, language and general cognition at age 8 (possibly due to the inherent correlations across these domains), hierarchical clustering indicates that pOTS interruption aligns most closely with literacy-related outcomes (cf. See rev 2 recommendation comment (1)). We observe no cognitive costs of left pOTS interruption across the many linguistic, pre-literacy and cognitive domains tested.

Neuroanatomical precision: We also fully agree that our sulcal marker of interest, i.e., whether the posterior portion of the left pOTS is interrupted or not, is relatively coarse and may lack in neuroanatomical precision (though this robustness may alternatively be seen as a strength). Beyond this first approximation, several questions remain unanswered, including whether additional interruptions confer an additional reading advantage, or whether interruptions in more specific locations of the left posterior OTS would be optimal. Unfortunately, multiple interruptions were too rare and gap locations too unevenly distributed in our cohort to support adequately powered comparisons. These questions can only be answered on larger sample sizes, which we now mention in our revised Discussion. Please refer to See rev 2 recommendation comment (3) for more details and for the results of tentative explorations which did not support clear trends.

Sample limitations: The characteristics of our sample (being both from a high socio-economic background, as often seen in pediatric research, and high-functioning) indeed limit the generalizability of our results, as we now acknowledge in the Discussion. However, we replicate results by Cachia and colleagues (2018) which were obtained in a sample where half of the participants were from a rural and (very) low socio-economic background, in order to match a sample of illiterate adults, and this association has been now found in multiple samples with various SES and cognitive characteristics. This suggests that overall such associations may be generalizable across extremely varied socio-economic backgrounds and also possibly along the spectrum of general cognitive skills (cf. R2 comment-toaddress #5).

Additional correction made during revision:

During the revision process, we identified one subject for whom an anterior OTS gap had been incorrectly included in the diffusion analyses (participant 6007 on previous Supplementary Figure S6). We removed this gap, reran all diffusion analyses, and updated the relevant results and figures. The results and conclusions remained essentially unchanged.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Comment (1) The results are currently framed in a manner that makes it difficult to figure out what the main findings of the paper are, and to see if those findings support their conclusions.

The beginning of the results is focused on showing that the longitudinal pipeline has a better consistency across time points, yet the authors still report follow-up analyses with the cross-sectional pipeline, even though those results are only trending. I think it would make more sense to only focus on the longitudinal pipeline because that is the novelty of this manuscript.

We fully agree with Reviewer 1 that the longitudinal pipeline shows better consistency and likely constitutes a more accurate, less noisy measure of sulcal interruptions, as we report. For this reason, all the results reported in the main figures concern the longitudinally processed data only.

Yet, we choose to also include and report supplementary cross-sectional analyses since these analyses, while less reliable, are important for several conceptual reasons:

(1) Since we wanted to demonstrate that T1 brain anatomy predicted T3 behavior longitudinally, it seemed important as a proof of principle to show that this prediction held with T3 brain data completely unseen, i.e., with T1 cross-sectional processing only. We have now clarified the text referring to this analysis (lines 192-196):

“This importantly implies that left pOTS morphology at T1 was still a significant longitudinal predictor of T3 reading behavior when T3 brain images were not used to improve T1 anatomical segmentations, i.e., with T1 cross-sectional processing alone, despite noisier tissue segmentations with cross-sectional processing (see also Reuter et al., 2012).”

(2) Since most studies are cross-sectional, reporting cross-sectional effects may constitute a more relevant reference for other studies, and for the generalizability of our results.

(3) Larger samples were available for the cross-sectional processing, allowing to test a maximum number of children, despite the higher noise in the sulcal measurements.

Please also note that across the 4 types of processing (2 timepoints x cross-sectional & longitudinal pipelines), only pOTS interruptions measured cross-sectionally at T3 showed a marginal association with reading scores, while the other 3 processings showed a large significant effect. Having established that cross-sectional measures are less reliable, and since our main goal was to demonstrate that T1 morphology predicted T3 reading, such trending results are found in the least reliable and least relevant condition, with otherwise large and significant effects, which seems quite acceptable. If anything, we believe that transparently reporting these consistent results altogether support, rather than challenge, the robustness of the association.

Our main findings are summarized in the last paragraph of the Introduction as well as in the first paragraph of our Discussion.

Comment (2) Also, as the authors note, it is known that sulcal patterns develop in utero; therefore, we would expect a perfect alignment between TP1 and TP3- there should be a more in-depth analysis of the lack of alignment in the main results. And further, this is more reason that the results should focus on the longitudinal pipeline which is closer to that expectation. However, it is unclear whether T1 data processed through the longitudinal pipeline preserves anatomical nuances of that time point (since the images are averaged across the timepoints). Maybe the cross-sectional pipeline is a better depiction of how anatomy changes as a child develops.

We thank Reviewer 1 for raising these relevant questions.

First, we would like to clarify that T1 longitudinal processing does preserve anatomical nuances of this time point. In fact, the averaging of T1 and T3 MRIs into a subject base is only an intermediate step of the longitudinal pipeline, which then crucially allows to re-initialize T1 MRI and T3 MRI tissue segmentations separately from their respective scans, but using this time the common tissue priors provided by the subject base (hence the reference to distinct “T1 longitudinal” and “T3 longitudinal” processing). We describe this in the Methods (lines 764-769):

“When images were available at both Time 1 and Time 3 for a child (N=43), the longitudinal Freesurfer pipeline was used, which first creates an unbiased subject average across time points (“base”) and then restarts multiple processing steps at each time point separately from the base template in order to refine tissue segmentations on each image (respectively T1 and T3 longitudinally processed images, Reuter et al., 2012; see Figure 1A).”

Therefore, a perfect match between T1 and T3 anatomy is not expected, even if we could expect (and indeed find) pOTS morphology, and interruption patterns in particular, to be extremely stable. Actually, our left pOTS interruption classification only changes between the T1 and T3 longitudinal processing for one child, with two left pOTS sulcal components separated by a distance estimated as 1.34mm at T1 and 0.99mm at T3, leading to a different classification given our 1 mm minimum-distance interruption criterion. We added the following sentence accordingly (lines 146-149):

“Only one child changed left pOTS interruption classification between T1 and T3 under longitudinal processing, due to the measured inter-component distance crossing the predefined 1-mm threshold, consistent with high stability of this macro-anatomical feature.”

Our results are therefore consistent with the view that sulcal interruption patterns are set in utero and stable after birth. Given that there actually is limited direct evidence to this effect, we chose to confirm this stability as a first step, which we refer to in the Discussion (lines 606-610):

“In our study, we observed high reliability of the pOTS interruption from age 5 (T1) to age 8 (T3), which is consistent with the longitudinal stability of interruptions of (i) the intraparietal sulcus and postcentral sulcus (Schwizer Ashkenazi et al., 2024), (ii) sulci in inferior frontal cortex (Tissier et al., 2018), and (iii) patterns of the anterior cingulate sulcus (Cachia et al., 2016) observed in late childhood (approximately 8 years onwards) to early adulthood.”

As discussed above, the longitudinal processing pipeline leads to better precision in assessing neuroanatomy, as well as its change across time (see the seminal paper by Reuter et al., 2012), which is why we mostly use measures derived from the longitudinal processing.

Comment (3) The analyses starting on p.4 ("Left pOTS interruption is prospectively associated with better reading skills") seem overly fixated on cross-sectional vs. longitudinal processing pipelines. The key comparisons should be on anatomical specificity (posterior or anterior OTS, etc) as well as including T1 reading skills as a regressor to see if the interruptions at T1 explain additional variance in T3 skills.

We thank Reviewer 1 for highlighting that we had started our result paragraph by a verification that was less essential. We have now reordered our results to highlight those concerning the anatomical specificity to the left pOTS, in the analyses reported under “The association between left pOTS interruption and reading skills was robust to different measurements of reading skills and sulcal interruption in four ways” (line 175 onward).

Testing whether the interruptions at T1 explain additional variance in T3 skills over and above T1 reading skills is so essential that it is actually the object of the entire next subsection “The effect of left pOTS interruption on reading accumulates over time” (lines 235-274). This claim is supported by the following findings:

(1) In the subsample of minimal readers with little to no reading experience at T1, we see that despite the absence of associations between left pOTS interruptions and reading at T1, the association is present and significant at T2 and T3 and appears to grow (lines 243-253 and Supplementary Figure 4);

(2) In the serial mediation analysis across the whole sample that follows, we observe important and significant additional contributions of left pOTS on later reading skills beyond the contributions at T1. We have now clarified this essential point and rephrased these results, notably adding this particular sentence:

“Importantly, there were also significant contributions of left pOTS interruption onto T3 reading on top of those mediated through T1, with large cumulative effects through T2 and directly onto T3 (d*c+f: ꞵ = 4.580, SE = 1.757, p = .009), while indirect effect through T1 TOWRE scores alone were non-significant (a*b*c+a*e: ꞵ = 0.971, SE = 0.839, p = .25).” '

We thank the reviewer for bringing to our attention that this was lacking in the previous version and needed to be made explicit. (Please note that p-values for the serial mediation analysis slightly changed, as the bootstrap was rerun to estimate this additional path.)

The analyses in the following section “Left pOTS interruption makes a strong longitudinal contribution relative to established predictors of reading” also support the same conclusion that left pOTS interruptions explain variance in T3 reading over and above T1 reading, despite the concern of collinearity with early literacy skills in our entire sample, as we explain below in details.

Comment (4) The authors do a version of this on p. 6 lines 269-273 and show that indeed, pOTS binary classification contributes a very similar proportion of the variance in T3 reading scores as the next best predictor (letter knowledge). They do not seem to be significantly different- are they?

When longitudinal contributions are tested separately, left pOTS interruption accounts for 20.5% of T3 reading variance while T1 letter knowledge accounts for 19.6% of this variance. While these values appear numerically close, one should first bear in mind that, as an emerging literacy measure rather than a pre-literacy predictor from a different domain, letter knowledge is expected to be a strong predictor of future literacy skills. The fact that left pOTS interruption, a much less proximal neuroanatomical marker, explains slightly more variance in reading stands out as a more interesting observation.

More importantly, while proportions of explained variance cannot be compared statistically in a straightforward fashion, the collinearity between these two predictors across the full sample anyway completely biases this comparison: because left pOTS interruption already accounts for significant variance in letter knowledge (cf. Figure 3B and analyses reported in lines 320-326), the estimation of the longitudinal contributions of letter knowledge to reading skills is inflated by this collinearity, picking up in addition some of the true effects of sulcal interruption on T3 reading (and vice versa).

For that reason, our claim that left pOTS interruption is a strong longitudinal predictor of reading relative to pre-literacy skills actually rests on the integration of all of the results reported in the corresponding section (lines 275-367), including: the contributions of each predictor separately under discussion, in addition also their unique contributions above all other predictors, and the analysis of the underlying collinearity and how it reduces the power in testing the true effects of pOTS interruption, as well as the same analyses repeated in the subsample of minimal initial readers where the collinearity between pOTS and emerging literacy is removed. (Please note that in this subsample, letter knowledge has minimal nonsignificant contributions to T3 reading, contrarily to sulcal interruptions - cf. Figure 3D and F and Comment-to-address #6).

To clarify this entire section and how all these results converge to support our conclusion, we have now moved several supplementary figure panels from Supplementary Figure 5 to Figure 3 (also addressing See rev 2 recommendation comment (9)) and largely edited the text, including:

“These significant associations between left pOTS interruption, performance IQ and pre-literacy skills (Figure 3B and D), as well as potential associations between pre-literacy skills themselves (cf. correlation matrices in Supplementary Figure 5D and H), make it challenging to fairly assess the separate unique contributions of each of these variables to the longitudinal prediction of reading skills. For instance, because left pOTS interruption explains a significant amount of variance at T1 in letter knowledge, early TOWRE scores and performance IQ, some of its true direct effect on future T3 reading skills is already captured in a baseline model including these cognitive predictors only, even before left pOTS interruption is explicitly entered as a predictor. Yet, even when controlling in a such a stringent way for all pre-literacy skills, early reading abilities, family history of dyslexia and their shared variance with left pOTS interruption in such nested model comparison, left pOTS interruption still had a marginal additional contribution to predicting T3 TOWRE reading scores (ΔR² = 5.0%, F(1, 31) = 3.74, p = .062) inspite of this important collinearity. Rapid naming and family history were the only two other predictors that outperformed left pOTS interruption in terms of unique contributions to T3 reading (ΔR² = 7.7% and 6.6% for RAN and mother ARHQ respectively, see Figure 3C), possibly because these predictors were among the least confounded with others (see Figure S5D for correlations among predictors) and thus among the least affected by collinearity. Letter knowledge, in contrast, had the smallest, non-significant unique contributions to T3 reading (ΔR² = 0.61%, p = .50).”

Thus, while the separate contributions of letter knowledge to T3 reading was almost as large as that of left pOTS interruption, their unique contributions above each other and other predictors strikingly differed: ΔR² = 5.0% for left pOTS interruption vs. 0.61% for letter knowledge.

Please also note that our statistical approach is particularly stringent: our nested model comparisons evaluated the unique contributions of each variable above all other predictors, i.e. after accounting for the variance in the outcome jointly explained by the covariance between predictors. In such cases of highly collinear variables, statistical power is importantly reduced.

Comment (5) And again even if the T1 presence of pOTS interruption explains this additional variance, what does that tell us? The brain (as measured by this pOTS measure) does not change from T1 to T3, as evidenced by this result: Lines 178-180: "left pOTS morphology at T1 remained a significant longitudinal predictor of T3 reading without taking into account brain information at T3, i.e., with T1 cross-sectional processing alone." T3 anatomical data is therefore redundant to this fit. This is further confirmed by the results in lines 220-226, showing differences in T1 TOWRE scores between children with continuous vs. interrupted pOTS. Additilly, the finding that "the effect of pOTS interruption accumulates over time" may instead suggest that TOWRE skills are more accurately measured at T3, especially for the lowest scores. It had alreaonady been shown that pOTS interruptions correlate with reading skills in readers- paired with assessment issues/lack of norms at T1 of these reading measures and the lack of change in pOTS over time, these results are not surprising.

We fully agree with the reviewer that given the previously observed associations between left OTS interruptions and reading skills in 10-year-olds and adults (Borst et al., 2016; Cachia et al., 2018) and the general understanding that sulcal patterns are stable after birth, it was natural to expect that left OTS patterns observed before or at the onset of literacy acquisition would predict future reading skills. Yet, validating hypotheses -however natural they may seem- is essential for moving science forward.

Beyond this expected longitudinal association, the fact that left pOTS interruption at T1 predicts additional variance in T3 reading above early literacy skills and all other cognitive, linguistic and familial predictors of literacy also tells us new and important things. Notably:

(1) It shows that this sulcal marker is a large and important predictor of future reading skills, with a predictive performance comparable to or exceeding the conservative benchmark of pre-literacy skills and familial predictors used for the identification of children at risk of dyslexia. This highlights the relevance of this neuroanatomical marker at a theoretical level for reading acquisition, with possible practical implications for early identification and intervention further down the line.

(2) We learn that the association between left pOTS morphology and reading is not merely a byproduct of pre-existing associations between sulcal morphology and cognitive or familial factors before literacy learning. Instead, our results clarify the neurocognitive origins of the relationship between reading and OTS morphology by showing a prospective and sustained association with reading acquisition between sulcal morphology and reading acquisition during literacy learning.

Concerning the relevance of T1 vs. T3 brain data, please refer to our response to Comment-to-address #2, in which we clarify that the longitudinal pipeline serves to improve the reliability of the assessment of sulcal interruptions at T1, but that our results are indeed also confirmed using only T1 brain data.

Concerning the matter of measurement accuracy of the TOWRE at T1, we agree that reading is harder to measure at the very beginning of literacy acquisition, and that the TOWRE is not normed below age 6. For this reason, our analyses controlled for early literacy skills at T1 using both scores of letter knowledge, as measured by the age-appropriate and normed WMRT letter identification subtest, and the principal component of the two TOWRE raw scores (scaled). This ensured a strict control procedure for all measurable emerging literacy skills: letter knowledge, but also word reading and phonological decoding (if already acquired in precocious/early readers, or to confirm this was not acquired yet).

In the absence of TOWRE norms for all children at T1, we used this TOWRE principal component to derive a T1 TOWRE population-scaled index. This scaling mostly served to ensure that this early reading score would be of a similar nature and scale as the T2 and T3 TOWRE standard scores, notably for statistical and plotting purposes (Supplementary Figure 4A and D).

Interestingly, we observed in the full longitudinal sample that left pOTS interruption was already associated around age 5 with this T1 TOWRE score, in addition to the letter identification scores. In the context of stronger associations observed at T2 and T3, this suggests that the T1 TOWRE score, even if not normed nor very accurate, does still capture relevant variance in early literacy skills.

Comment (6) Further, while one main conclusion is that the gyral gap predicts T3 reading ability better than typical behavioral TP1 measures (but it really does not appear to), this result is mixed in with a series of nonsignificant results (labeled as marginal or trending, but p>0.05) and difficult to parse sections. Some, but not all, analyses are performed in a set of 'pre-readers' (defined as scoring <95th percentile on a reading test which is not particularly stringent), and at some points, the authors get a significant result in the whole sample, but not the prereaders and interpret these results as only trending because of a smaller sample size (line 287), but at other times they get a trending result in the main group, and a significant result in the pre-reader group (line 303), and choose to focus on the result of the prereaders. This inconsistency makes me skeptical of why some analyses were done with different samples. The definition of a pre-reader is not really a pre-reader but rather just below an average reader. This also leads back to my question above on the underlying hypotheses, which critically needs to focus on prereaders. Can you further explain the decision to run some analysis in pre-readers (example line 285/286) and further explain what it means to be a pre-reader in this sample? Perhaps try to think of a more direct way to examine the impact of the gyral gap at TP1 predicting behavioral performance at TP3, as a trending result in the whole sample followed by significance in just a subset is not convincing (line 298, 303).

We thank Reviewer 1 for this opportunity to clarify our results and some misunderstandings.

We acknowledge, and are sorry for, the lack of clarity of our previous explanations, especially as some analyses are complex and some conclusions require integrating results across multiple analyses to understand the overall implications. We believe that our statistical approach was justified, though perhaps not well explained in the original version. We have now clarified and addressed these issues in the revision by renaming pre-readers to ‘minimal initial readers’ - which is indeed more accurate (thank you for raising this concern) -, better explaining the composition of this subsample and what analyses restricted to this subsample bring (notably lines 243-253, and in the descriptive demographic Table 1), rewriting the entire section “Left pOTS interruption Left pOTS interruption makes a strong longitudinal contribution relative to established predictors of reading” (lines 275-367), and adding a number of figure panels to Figure 3 (previously in supplementary figure 5), to aid the comparison and integration of results across both samples to support our conclusion. To answer directly the reviewer’s questions and remarks:

Why two groups? We carry out analyses both in our entire available sample and in a subsample with little to no early literacy skills, since these samples are both relevant for our hypotheses and offer different advantages:

The entire sample allows to maximize sample size, and is more representative of the general population. This naturally implies that some kindergarteners have emerging literacy skills, including letter knowledge, knowledge of a few sight words and minimal decoding. Interestingly, we found these early literacy measures to already be significantly associated with our sulcal marker, but two problems also derive from this: (1) These emerging reading skills preclude from directly testing in this sample our strong hypothesis that pOTS interruption before literary learning predicts future reading scores. (2) The collinearity between the T1 behavioral predictors and left pOTS interruption is a statistical obstacle to testing reliably how each predictor truly contributes to predicting future reading skills separately (as explained in our answer to Reviewer 1’s Comment-to-address #4 above, taking the comparison with letter knowledge as a key example).

To remove the association between pOTS and T1 reading and bypass both these problems, we perform additional analyses in a decently-sized subsample of children with little to no reading abilities at T1 - that we have now renamed more accurately as ‘initial minimal readers’.

Definition and composition of our subsample of “initial minimal readers”:

Kindergarteners are expected to know letters and how to write their names, they may have understood the basics of the alphabetic principle and phonological decoding, and may have learned to recognize a few highly frequent sight words through repeated exposure in their books and environments (e.g., “the”, “and”). From our entire sample, including a number of children accordingly able to read some sight words and decode some syllables and pseudowords (cf. Table 1), we defined a subsample of lowest readers by using a cut-off of a TOWRE population-scaled index at T1 under 95 (due to the lack of norms, cf. see rev 1 recommendation comment (5)). For reference, the equivalent standard score of 95 would correspond to a 37th percentile in a normal distribution.

This cut-off was chosen as it translated to very little actual reading abilities overall and crucially removed the statistical association between pOTS and T1 reading needed to address our two problems, while being still decently sized to allow for our statistics (n = 24, i.e. 56% of the sample, cf. distributions on Supplementary Figure 4A). We now specify more clearly notably in the results lines 244-253:

“We conducted a supplementary analysis in the subset of 24 children with the lowest reading scores at T1 (defined as those with a TOWRE T1 population-derived composite score equivalent below 95). Hereafter, we refer to these children as initial minimal readers: 79% could not decode a single syllable or pseudoword in 45 s, with a maximum of four pseudowords decoded, and 63% could not recognize a single frequent sight word, with a maximum of seven words recognized. Although these children showed no difference in T1 reading scores as a function of whether their left pOTS was interrupted or continuous (F(1, 19) = 0.21, p =.66, ΔR² = 0.82%, Cohen’s d = 0.034), [...] “

and methods - lines 988-991:

“We therefore selected a subsample of children whose TOWRE standard score equivalent at Time 1 was below 95, resulting in a subset of 24 initial pre-readers (median of 0 words read in 45s in SWE, range: 0-7; median of 0 syllables in 45s in PDE, range: 0-4; see Table 1).” as well as Table 1.”

Expecting a larger proportion of children of this age with no word reading or decoding at all would be both unnecessary for our analyses and developmentally unrealistic, especially in this high-SES sample, selecting an atypical population at especially high risk of dyslexia.

Discrepancies between statistics in the full sample and the subsample:

Because this subsample was chosen to remove the associations at T1 between pOTS and early literacy which affected some analyses (cf. Comment #4), some discrepancies in the results across the two samples are actually expected. Conducting statistics on 56% of the full sample may also affect the statistical power for other analyses in the subsample.

First, concerning issue #1 of the simple longitudinal prediction of T2 and T3 reading, we find that left pOTS interruption also reliably predicts future reading in the subsample of initial minimal readers, even if the association is absent at age 5, confirming analyses in the main sample.

Concerning the issue #2 of the comparison with other benchmark predictors of literacy, our results in the entire sample show that despite a large amount of collinearity between our T1 predictors (documented, explained and quantified in lines 320-334, Figure 3B and Supplementary Figure 5D, including strong significant associations of left pOTS morphology with T1 reading, T1 letter knowledge and T1 performance IQ) and despite using the most stringent statistical approach to control for it (nested model comparison), left pOTS interruption still has a marginal contribution to predicting T3 reading on top of all other early literacy, pre-literacy, cognitive and familial predictors of reading (Figure 3C), in line with its large separate contributions (Figure 3A). In the subsample of initial minimal readers, where there is only partial collinearity with performance IQ but not early literacy, we find the specific longitudinal contributions of left pOTS morphology to be strong and significant over and above all of the other T1 behavioral and familial predictors (see Figure 3F and corresponding text). In other words, the collinearity present in the full sample weakens statistical power when controlling for other collinear predictors, with only marginal unique longitudinal pOTS contributions in our stringent analyses, while removing this collinearity in the subsample reveals a large and highly significant unique contribution, in line with the large contributions of pOTS observed overall (Figure 3A, D, and F).We believe that detailing these analyses and finding coherent results on both samples, with potential differences in statistical strength explained by specific factors, overall brings reliability to our results and conclusions, rather than challenges them.

The reviewer refers also to one interesting incidental finding when examining the T1 cognitive associations of the pOTS interruption in the full sample, consisting in a significant association with the concept formation subtest, measuring a form of performance IQ (though a weaker association than with reading skills, see Figure 3B). When restricting to the subsample of initial minimal readers, this association was then indeed found marginal, though of a similar amplitude as in the full sample. We therefore speculated that the combination of a significant effect in the full sample, and of a marginal finding in the subsample, revealed a possibly true underlying effect, while likely affected by a decrease of statistical power associated with using only 56% of the sample. This effect and speculation are in no way fundamental to our study and claims. The new additional analysis revealing a significant association with performance IQ at T3 again in the full sample (Supplementary Figure 6) would however further support this idea.

Comment (7) Given that there are these potentially major differences in the pOTS, how are the results of structural connectivity mapped to a common template for Figure 4? There is likely an error in registration at the exact location of the observed differences.

Individual structural connectivity results were mapped to an MNI template space using standard procedures (lines 919-922 of our Methods:):

“The obtained individual probabilistic fiber density maps were then transformed into template MNI space for statistical analysis using non-linear registration in ANTs, from individual diffusion space (individual FA map) to the HCP1065 MNI FA template, and smoothed with a 4-mm-FWHM Gaussian kernel.”

As the reviewer points out, group analyses performed in atlas space are inherently affected by variations in neuroanatomy across individuals and by the quality of registrations. Mapping individual connectivity results to template space was however the only approach that made sense to us to evaluate which connectivity benefits would be associated with the presence of such gaps, given the variability in the number and location of gyral gaps in children with interrupted sulci, and the need to find a common comparison ground with children who had a continuous sulcus. High neuroanatomical variability around the pOTS in particular likely plays against us in investigating such connectivity differences. In spite of these difficulties, we do find differences that seem to explain the reading advantage associated with the presence of gyral gaps.

Comment (8) While I appreciate the control analyses for the white matter connectivity, I think more robust analyses are more appropriate. Are there MD differences between white matter underneath sulci and gyri of neighboring or even distant regions? I'm not sure what the white matter analyses are really adding to the story. The white matter differences do not explain the differences in T3 reading ability.

Our rationale for choosing and conducting the reported white matter analyses was the following: Having provided strong evidence that left pOTS sulcal interruptions facilitate reading acquisition, a naturally following question was that of its underlying mechanisms. It had been hypothesized - though never tested directly - that white-matter differences associated with the presence of gyral gaps may improve the connectivity around sulcal interruptions and constitute a potential mechanism linking interruptions to behavioral advantages (cf. lines 399-424). We therefore set out to investigate the underlying white matter, how it differed between children with and without interruptions, and its associations with reading skills.

Since long-distance projection connections tend to terminate more in gyri than sulci, one potential hypothesis was that gyral gaps in pOTS could be a hot-spot for such connections, facilitating the information flow between regions of the OTS and other functional hubs of the reading network. Because a higher density of such projection fiber endings is expected to result increased intra-cortical myelination, we test whether MD values in the gaps vs. sulcus are consistent with a higher myelination in the gaps than in sulcus, but we instead find the opposite pattern of lower myelination (cf. lines 435-457). This invalidated our first hypothesis.

In this context, it is thus unclear which questions would warrant additional analyses on MD values in other anatomical regions, or what these would bring in our specific search for mechanisms underlying the reading advantage of pOTS interruptions.

A second potential mechanism was that left pOTS gaps could be associated with increased local connectivity surrounding the left pOTS, through the packing of more U-shaped fibers traveling superficially, resulting somehow in a computational advantage for reading. This time, we found evidence in favor of this hypothesis: children with an interrupted left pOTS had more local white-matter connectivity around the pOTS travelling superficially under the left pOTS and it gaps (cf. lines 477-493). Moreover, when accounting in the same model for T3 reading on top of the presence of interruptions, this increase in local U-shape fibers was instead captured by T3 reading scores rather than the presence of a pOTS interruption. As such, these white matter differences did indeed explain differences in T3 reading, suggesting a candidate local-connectivity pathway underlying the reading advantage.

Comment (9) Were any of the analyses performed on the right pOTS? What are the differences in the laterality of gaps? Connectivity?

The right pOTS had an interruption rate of 58.3%, which was similar to that of the left pOTS (56.5%, Fisher test: p = .75, cf. Supplementary Figure 1 and lines 134-139).

The right pOTS does not show associations with reading skills, which we report in Figure 2C and lines 184-186:

“[A]ssociations between sulcal interruption and reading skills were specific to the left pOTS within VTC. No significant associations were observed for anterior OTS (aOTS) or MFS in either hemispheres, or for the right pOTS (see Figure 2c, all p-values > 0.58, all ΔR² < 0.7%).”

Due to the absence of an association between the right pOTS and reading abilities, additional connectivity analyses in the right hemisphere did not seem warranted.

Comment (10) Finally, the authors' conclusion, and particularly the last section of their discussion on the neuronal recycling hypothesis, does not seem related to their data, analyses, or questions. The authors do not motivate the link between a hypothesis focused on changes in functional activation during skill acquisition, to their current work which only has structural and diffusion connectivity data. Either remove or further motivate this discussion.

While there has been a heated debate on the specific nature and underlying mechanisms of “neuronal recycling”, our intention was merely to frame our results within the general well-established and widely accepted premises that (i) cultural skills such as reading and symbolic number processing are too evolutionarily recent to rely on dedicated neural circuits and therefore must recruit pre-existing cortical territories, and (ii) reading acquisition is accompanied by the development of functional responses to written words in left posterior vOT, around the pOTS, free from dedicated evolutionary constraints.

Keeping in mind that the left pOTS and surrounding cortex see a major functional development of functional responses to words literacy acquisition is indeed fundamental in order to understand the relevance and implications of our structural results in this particular patch of cortex that hosts the VWFA, and their relative cognitive specificity to reading. Specifically, we claim that strong structure-function associations around the left pOTS for reading may mediate the behavioral benefits of left pOTS gyral gaps.

We therefore refer to these questions and hypotheses throughout our text. We first mention the existence and importance of structure-function relationships in the case of the VWFA early on in our introduction (lines 57-66). Throughout the paper, we state our hypotheses on how sulcal morphology and the associated structure of white-matter fiber tracts may result in functional and behavioral reading benefits, motivating our white-matter connectivity analyses. The discussion paragraph “Local connectivity differences at gyral gaps: a computational advantage?” (lines 629-648) then explicits in depth our final interpretation linking our sulcal marker and the computational efficiency of the functional reading network, as supported by an extensive literature.

To avoid any implication that our findings speak directly to the neuronal recycling framework, we have revised the relevant sentence (lines 651-659) and conclusion to focus solely on its broad, non-specific premises and have removed explicit mention of neuronal recycling.

“Interestingly, associations between sulcal interruptions and behavior have also been reported for symbolic, but not non-symbolic, number processing (Roell et al., 2021; Schwizer Ashkenazi et al., 2024), highlighting common processes by which cultural acquisitions, too recent on the scale of human evolution to rely on dedicated neural circuits (Dehaene & Cohen, 2007), may be boosted by the presence pre-existing gyral gaps. In the case of the left pOTS, various other aspects of early brain structure have been proposed to constrain the development of reading skills and of functional activations for reading in the VWFA and broader neural network involved in reading such as anatomical connectivity (Bouhali et al., 2014; Grotheer et al., 2021; Li et al., 2020; Moulton et al., 2019; Saygin et al., 2016) and cortical morphology. In terms of the latter, longitudinal studies following pre-reading children at familial risk of dyslexia showed for instance that the surface area of the left fusiform gyrus (Beelen et al., 2019) and patterns of grey-matter volume around the VWFA (Skeide et al., 2016) could be predictive of future dyslexia diagnosis.”

We have also updated the subsection header to “Common cortical constraints on cultural learning” to reflect this more neutral framing, and adjusted the first discussion paragraph and the conclusion to avoid confusion with the neuronal recycling hypothesis. These changes clarify that our discussion situates the findings within a general conceptual background, rather than interpreting them as evidence for a specific theoretical model.

Comment (11) I think this paper could add interesting knowledge for the specific subfield of reading and the brain. However, the current state of the results, especially with the inclusion of so many trending results and the comparison of so many different processing pipelines and models, in addition to a conclusion that is not motivated by the work makes it difficult to appreciate the paper.

We thank Reviewer 1 for their suggestions to improve the manuscript. We have revised the text to focus more clearly on the main longitudinal and anatomically specific findings, and have clarified the role of the initial minimal-reader subgroup, the significant longitudinal contribution of pOTS interruption beyond associations with emerging literacy at age 5, and its unique contribution relative to other established predictors of reading. We hope these changes make the central contributions of the study clearer.

Although the revision adds some content to an already detailed article, we believe that, taken together, these analyses provide a thorough and integrated account of the association between OTS morphology and reading. In particular, they help clarify how macro-anatomical differences in sulcal morphology may relate to cognition, how these associations unfold across development, and how a candidate white matter pathway may contribute to this relationship.

Minor points:

Comment (12) Figure 1 needs more explanation for those who have not hand-labeled these regions. Can you clarify what differentiates the continuous left pOTS in the top row and then interrupted pOTS in the bottom row? The top row all has a gap between dark green and light green. Also, what does the color scale mean? Is it a posterior to anterior gradient?

We thank Reviewer 1 for highlighting the need for clarification. Our legend now stipulates:

“Left hemispheric views from exemplar children are shown, either with a continuous left posterior OTS on the top row (i.e., no more than one connex OTS component in the space posterior to the white dotted line) or an interrupted left pOTS on the bottom row (two or more sulcal components present in posterior space, with a total Euclidean distance between components ≥1mm). Green and orange gradient colors are respectively used to highlight the different sulcal components of the OTS and MFS.”

Comment (13) The discussion on sulcal interruption associations with symbolic but not non-symbolic number processing is interesting! Are these sulcal interruptions present in non-human primates?

To our knowledge, pOTS interruptions directly analogous to the human left pOTS gyral gaps studied here have not been systematically characterized in non-human primates. Macaques do have an occipitotemporal sulcus, and comparative work shows that sulcal landmarks in macaque and human VTC can relate to functional organization. However, the human OTS is laterally displaced relative to the macaque OTS, and the specific left pOTS interruptions examined here are defined in relation to human VTC and literacy. We therefore view the comparative anatomy of these interruptions as an open question.

Comment (14) Your T1 voxel size (line 629 0.859 x 0.859 x 1.2mm) does not match the recommendation for the FreeSurfer pipeline you stated, being both anisotropic and having a dimension with > 1mm. "The method works well for voxel sizes 0.75 mm3. It should work with voxel between 1mm3 and 0.75mm3 ... We haven't tested the stream for inputs with anisotropic voxels." (https://surfer.nmr.mgh.harvard.edu/fswiki/SubmillimeterRecon). Further, you note in the methods section that small interruptions can be a result of bad segmentation (line 718/719). Given that segmentation is affected by voxel size, and potential partial volume effects. How do the voxel parameters impact the results and reliability of the analyses?

Our T1 voxel volume was 0.885 mm³, within the voxel volume range for which the FreeSurfer submillimeter stream is expected to work, although we acknowledge the caveat that the stream was not specifically validated for anisotropic voxels. Yet, Freesurfer pipelines resample anisotropic voxels to isotropic ones by default, resulting here in a final voxel size of 0.859mm isotropic on which the submillimeter pipeline was run. To address possible segmentation concerns, we visually inspected all segmentations in bilateral ventral occipitotemporal cortex thoroughly, excluded scans with excessive motion, imposed a conservative> 1 mm criterion for classifying interruptions, and verified high T1–T3 reliability. Thus, while voxel anisotropy may affect very fine morphometric estimates, our macroanatomical feature of interest - the binary pOTS interruption classification - appears robust.

Comment (15) On line 417, you should refer to the panel of the figure (I think 4c bottom panel), like you did for many other results. It was helpful when you linked a particular result to its matching figure.

We thank Reviewer 1 for pointing out this omission, which has now been corrected.

Comment (16) The percentage values on the bars should be larger for Figure 2, they are hard to read.

We had chosen to represent the p-value significance on the bars, rather than the percentages themselves, which are indicated by the x-axis and reported numerically in the main text (except from the least relevant ones). We have now increased the font size of these significance stars. Thank you.

Reviewer #2 (Recommendations for the authors):

The authors may want to consider the points below.

Comment (1) The relationship with Performance IQ is interesting. It does not predict much variance in reading scores at T3 but it is strongly predicted by the pOTS interruption at T3. There is quite a significant correlation among all the measures. The authors acknowledge the difficulty in assessing the unique contribution of each measure. I was unclear how they "over" controlled for pre-literacy skills etc. (line 296) and would like more elaboration on that in the main text.

It would be interesting to know whether there were any cognitive measures where better performance was not predicted by the presence of the interruption? Were there any measures where worse performance was predicted? (i.e. a potential cost of this feature).

A relative specificity of left pOTS interruption on reading:

We agree that the association between left pOTS interruption and performance IQ at T1 raised the question of the cognitive specificity of a reading advantage associated with left pOTS interruption, as opposed to a possibly more general cognitive advantage as suggested by Reviewer #2 in their public review above.

This question is of course complicated by the fact that reading skills are by nature highly correlated with general cognitive abilities, as they are also with math abilities (see e.g., the meta-analyses by Ünal et al., 2023) and language. General cognitive skills and reading are in fact so intertwined that they positively and symbiotically influence each other over development (Ferrer et al., 2007; Peng and Kievit, 2020). Yet, bearing these strong associations in mind, our results suggested a certain degree of cognitive specificity of left pOTS interruption benefits on reading. Indeed, while left pOTS interruption predicted about 14.3% of variance in T1 concept formation (measuring performance IQ), it predicts about 20.5% of T3 TOWRE reading scores. In addition, left pOTS interruption significantly and strongly predicts T3 reading scores, while T1 performance IQ did not (cf. Figure 3A, D and C, except in the subsample of initial minimal readers where pIQ has a significant contribution but about twice smaller than that of pOTS interruption - Figure 3F). Therefore, the prospective association between left pOTS interruption and reading skills could not merely be driven by an association with general cognitive abilities at age 5. Finally, left pOTS interruption showed no other significant association with the pre-literacy skills considered (other than early literacy itself), or family history of dyslexia. This points towards a strong and relatively specific contribution of left pOTS on reading acquisition beyond its moderate association with T1 general cognitive skills.

We also conducted a new analysis to directly address this question, by systematically testing the associations between left pOTS interruption and all the cognitive and linguistic scores that were obtained at T3. Our new results also support the conclusion of a relatively specific cognitive advantage on reading, and are shown in Supplementary Figure 6 (with Supplementary Table 2 providing a description of the different subtests). The results are exposed in the main text as follows, in a new result subsection “Left pOTS interruption has a relatively specific cognitive impact on reading abilities” (lines 369-397):

“The incidental observation of a significant association between left pOTS interruption and performance IQ at T1, although smaller in magnitude and not sufficient to account for the association with literacy skills, raises the possibility that left pOTS interruption may confer a broader cognitive advantage rather than a specific benefit for reading acquisition. We therefore conducted a supplementary analysis to examine the relative domain-generality versus cognitive specificity of this association. This question is inherently difficult to address because reading and general cognitive skills are strongly correlated and mutually influence each other across development (Ferrer et al., 2007; Peng & Kievit, 2020; Ünal et al., 2023).

“We gathered all cognitive and linguistic scores obtained at T3, including measures of mathematics, general cognition, executive functioning, language, and reading (see Table S2 for a full list and description of the included subtests), and tested how each measure was associated with left pOTS interruption, residualized for demographic variables and total brain volume. Left pOTS interruption showed strong and significant associations with several measures of language, reading, mathematics, and general cognitive ability (see Supplementary Figure 6). As expected, we also observed strong and significant correlations among the behavioral measures themselves.

“However, the strongest correlations with left pOTS interruption residuals were observed for reading scores. Moreover, hierarchical clustering grouped left pOTS interruption with a subcluster that included most reading assessments - spanning phonological decoding, word reading, and sentence reading fluency -, as well as spelling and more surprisingly an expressive language index. This reading-related cluster was most closely related to another cluster comprising measures of written and oral language comprehension (i.e., receptive language), followed by a cluster of phonological awareness. In contrast, measures of mathematics, general cognitive ability, backward working memory and rapid naming formed a separate cluster. Thus, although left pOTS interruption is associated with several cognitive, linguistic, and academic measures, its pattern of associations is most closely aligned with children’s reading outcomes, supporting a relatively specific cognitive advantage for reading.”

We have therefore modified the Discussion paragraph that refers to the question of cognitive specificity vs. domain-generality as follows (lines 564-580):

“Interestingly, left pOTS interruption also showed associations with other cognitive skills: it was associated with performance IQ at age 5 (see Figure 3B and E), and with performance IQ, math and language skills more broadly at age 8 (Supplementary Figure 6). However, left pOTS interruption was most strongly associated and clustered with assessments of reading at age 8 (as well as spelling and expressive language). In addition, the prospective association between left pOTS interruption and reading skills was not merely driven by an association with general cognitive abilities at age 5 (Figure 3C and F) and no association with any other pre-literacy skill was found (Figure 3B and E). Math, reading, language and cognitive skills are intrinsically highly intercorrelated, as we also observed at age 8, and are known to mutually and positively influence each other over time (Ünal et al., 2023; Peng & Kievit, 2020; Ferrer et al., 2007; Huettig & Hulstijn, 2025; Dehaene et al., 2015). The weaker pOTS cognitive associations may therefore result from the mere combination of a specific pOTS influence reading and of associations between reading and other cognitive domains, although more research is needed to better understand these associations and how they unfold during development. Together, these findings highlight a relative degree of specificity of left pOTS influences to reading abilities, in line with the welldocumented functional specificity of the left pOTS for orthographic processing and reading (e.g., Dehaene & Cohen, 2011; Gaillard et al., 2006; Pillet et al., 2024; Wilson et al., 2013; Woolnough & Tandon, 2024; Zemmoura et al., 2015; Zhan et al., 2023).”

No observed cognitive cost to left pOTS interruption:

These different analyses show that there is no negative cost to the presence of left pOTS interruption, or at least not across the wide variety of cognitive skills we considered. While left pOTS interruption was significantly and positively associated with many assessments of general cognition, math and language (perhaps due to inherent associations and interactions between these domains), several domains were not significantly associated with the morphology of the left pOTS: rapid naming, backward working memory, as well as receptive vocabulary and phonological awareness (at least at T1). Except for a nonsignificant correlation of -0.01 with CTOPP Blending Words at T3, no subtest showed negative associations with left pOTS interruption.

“Over-controlling” and assessing unique contributions:

We had used the term “over-controlling” to refer to the fact that testing the longitudinal effects of left pOTS interruption on T3 reading over and above a number of strongly collinear covariates (such as letter knowledge, T1 reading scores and performance IQ) by nested model comparison necessarily decreased the estimation of the true effect of left pOTS interruption on reading. Indeed, the baseline model with collinear variables then already accounts for a part of the variability in pOTS morphology and of its effect reading, through its joint effects with other co-variates, even before the sulcal variable is explicitly added in the model. The statistical power to test the true sulcal longitudinal effect by nested model comparison over such a strong baseline is therefore decreased, and the effect is underestimated.

Since this non-canonical term and our previous explanations appeared confusing, we clarified the corresponding paragraph as follows (lines 336-361):

“These significant associations between left pOTS interruption, performance IQ and pre-literacy skills (Figure 3B and D), as well as potential associations between pre-literacy skills themselves (cf. correlation matrices in Supplementary Figure 5D and H), make it challenging to fairly assess the separate unique contributions of each of these variables to the longitudinal prediction of reading skills. For instance, because left pOTS interruption explains a significant amount of variance at T1 in letter knowledge, early TOWRE scores and performance IQ, some of its true direct effect on future T3 reading skills is already captured in a baseline model including these cognitive predictors only, even before left pOTS interruption is explicitly entered as a predictor. Yet, even when controlling in a such a stringent way for all pre-literacy skills, early reading abilities, family history of dyslexia and their shared variance with left pOTS interruption in such nested model comparison, left pOTS interruption still had a marginal additional contribution to predicting T3 TOWRE reading scores (ΔR² = 5.0%, F(1, 31) = 3.74, p = .062) in spite of this important collinearity. Rapid naming and family history were the only two other predictors that outperformed left pOTS interruption in terms of unique contributions to T3 reading (ΔR² = 7.7% and 6.6% for RAN and mother ARHQ respectively, see Figure 3C), possibly because these predictors were among the least confounded with others (see Figure S5D for correlations among predictors) and thus among the least affected by collinearity. Letter knowledge, in contrast, had the smallest, non-significant unique contributions to T3 reading (ΔR² = 0.61%, p = .50). In order to reduce the collinearity confounds resulting from the associations between pOTS interruption and T1 early reading skills, we also reran the same analysis in the subsample of initial minimal readers. Left pOTS interruption in this subsample significantly contributed to predicting future T3 reading over and above all pre-literacy skills and family history (ΔR² = 27.8%, F(1, 12) = 13.80, p = .003), making it the first independent longitudinal contributor of reading (see Figure 3F). The second largest longitudinal predictor in terms of unique contributions was performance IQ (ΔR² = 12.8%, F(1, 12) = 6.34, p = .027). No other predictor had a significant unique contribution to the longitudinal prediction of T3 reading in the subset initial minimal readers, including rapid naming even though its separate contribution was larger than that of pOTS.”

We also moved several figure panels from Supplementary Figure 5 to Figure 3 so that readers would have the necessary material in the main text to document this collinearity and see how it impacted the separate vs. unique contributions of each predictor to T3 reading.

Comment (2) It was surprising that no other anatomical feature of the pOTS (like sulcal depth) predicted reading skills. I would have expected these to be quite closely correlated with interruption distance that did predict reading. Do the authors have any thoughts as to why this might be?

We thank Reviewer 2 for this question. While it seems intuitive that gyral gaps would affect sulcal depth, this mostly depends on whether we consider these gaps to be part of the sulcus -which here we do not- and on how the depth measure is actually calculated. With a definition of the sulcus as exclusive of gyral gaps, sulcal interruptions affect sulcal depth profiles only locally around the interruptions, where the fundus has to emerge to the convex hull surface. Because we measure sulcal depth based on the depth of the 8% deepest vertices (using an approach adjusted from that of Madan, 2019, see Methods lines 845-852), our measure of sulcal depth is essentially unaffected by interruptions.

As a confirmation and for more details, we provide Author response image 1 for the correlation matrix between the various morphological measures under consideration. (Using the same conventions as other correlation matrices displayed in the article and supplements: only significant correlations are displayed, according to FDR-corrected p-values above the diagonal.)

Author response image 1.

Comment (3) I hadn't quite appreciated on first reading the quite large variability in the location of the interruption (see Figure 4a, where three clusters seem to emerge). Nor did I appreciate that more than one interruption was often present (perhaps Figure 1 could be updated to show examples with more than one interruption). Some measures of distance along the OTS could be helpful and a possible comparison of the "types" based on location. I would have thought that proximity to the location of the VWFA might have meant a stronger effect. I presume this could be tested or discussed.

We entirely agree with Reviewer 2 that these are important questions. Unfortunately, it also appears that these cannot be reliably tested in our sample.

Specifically, while 25 children out of 43 had at least one interruption in the left pOTS, only 3 of those had two or more interruptions larger than 1mm each (or 6 children with two interruptions of any size). These incidence rates, combined with the fact that the locations of pOTS gaps do not distribute uniformly along the postero-anterior axis (Figure 4A), do not allow us to reliably assess in this sample either whether having more interruptions may confer an additional reading behavioral advantage, or whether interruptions in particular locations of the pOTS could be more beneficial than others.

We explored - for speculative purposes - whether a relationship may exist between reading scores and the location of left pOTS interruptions, taking the center of mass of the cortical gyral gap volumes as a coarse reference. We found no evidence for any relationship between reading and the locations of the left pOTS gap along the antero-posterior axis, nor between reading and the distance of the gap centroids to the VWFA (taking the Neurosynth reading meta-analysis peak coordinates, MNI -44 -58 -12), or between reading and the distance to the VWFA standard coordinates along the Y axis only (please see Author response image 2).

We choose not to report these analyses in the main text as they seem largely underpowered.

However, we added an example participant with two left pOTS interruptions in Figure 1 as suggested, and now added reference to these important questions in the Discussion, for future studies to address (lines 596-600):

“Such large-N approaches are needed to answer a number of outstanding questions, such as, in our case, whether there is an additional behavioral reading benefit to having two or more left pOTS interruptions - since these were too infrequent in our sample -, or whether gyral gaps placed in specific locations along the left pOTS may be optimal to for facilitating reading acquisition.”

Author response image 2.

Comment (4) It wasn't clear to me how the analysis of the diffusion data proceeded in the case of more than one interruption - it appears that each one was measured but this could be clarified. Is the location of relevance to the findings?

Regarding diffusion analyses, the diffusion tensor analysis averaged values across all pOTS interruptions together in the few children who had several. The tractography analysis used the full pOTS and all its interruptions, if present, together as a seed volume.

The precise location of the gyral gaps would indeed be very relevant to the tractography findings as one expects the gyral gaps to relate in a local fashion to the neuroanatomy of the underlying white matter. However, using the whole pOTS with its gaps as a seed seemed to us the only reasonable way to investigate the white-matter correlates of the effect of gyral gaps on reading at a group level, given (i) the large variability in gap location among the 15 of children with interruptions and diffusion data, (ii) the fact some children had multiple gaps, and most crucially (iii) the need to find a common ground for comparison with the group with no left pOTS interruption (continuous sulcus). The fact that our analyses average along the whole pOTS and across children with different pOTS gap locations likely washes away some of this neuroanatomical specificity, thus reducing the strength of any expected effect.

Comment (5) The authors could include a discussion of the sample's characteristics. They are not necessarily representative given the high SES and the average PIQ being more than 1SD above the mean.

We thank Reviewer 2 for pointing out this limitation we had overseen.

High-SES and high-functioning population samples indeed constitute a common bias in pediatric research samples, since study participation requires parents to have the time, interest and access to necessary information. This bias was likely increased here by the fact the data was acquired at Stanford University, in a very privileged area (Palo Alto, CA). The median performance IQ was indeed about one standard deviation above the population mean. Moreover, our sample is enriched for family history of dyslexia. We agree with the reviewer that this limits the generalizability of our results.

Yet, we replicate the association between left pOTS interruption and reading skills observed by Cachia and colleagues (2018) in a sample where half of the participants were from a rural and (very) low socioeconomic background (so that they would match a sample of illiterate adults), and with limited higher education. This suggests that, overall, such associations may generalize across extremely varied socioeconomic backgrounds and also possibly along the spectrum of general cognitive skills.

Since our reviewed preprint came out, two other studies further replicated associations between the left pOTS and reading skills (Cachia et al., 2025; Yao et al., 2026), in populations of different ages, countries and across the reading spectrum, including many children with dyslexia (Yao et al., 2026). While SES and cognitive skills are not systematically reported in those studies, one could assume that these studies and ours pooled together would be more representative.

Accordingly, we added this paragraph to the discussion (lines 548-562):

“Our study further extends previous results by Borst, Cachia and colleagues (2016; 2018) to a population of children with above-average cognitive and reading scores, from relatively high-SES families, and enriched for family history of dyslexia (cf. Table 1) - although no child met criteria for dyslexia by T3. While our sample is not representative of the general population, thereby limiting the direct generalizability of our results, half of the adult sample from Cachia et al. (2018) came from low socio-economic and rural backgrounds (matched to a sample of illiterate adults), suggesting that the association between left pOTS interruption and reading skills is robust across socio-economic contexts. Since our reviewed preprint came out, the association between left pOTS interruption and reading was further replicated (Cachia et al., 2025; Yao et al., 2026), notably in a larger child sample along the full reading continuum and oversampling children with dyslexia (Yao et al., 2026). (Of note, the finding that pOTS morphology correlated in adults with reading fluency rather than accuracy (Cachia et al., 2025) may also explain one of two prior negative findings concerning the association between pOTS and reading (Del Mauro et al., 2022)). Thus, our results together with the extant literature overall underline a common effect of left pOTS interruption observed across the continuum of reading abilities, and possibly of cognitive abilities.”

Minor

Comment (6) Line 342-44: A visualisation of what is meant by superficial vs deep plis de passage would help.

While an illustration of deep vs. superficial plis de passage would greatly help in clarifying this distinction for the readers, we feel that providing such a visualization in our figures would be equally confusing, if not more, as it may prime readers to wrongfully assume that we differentiate between superficial and deep plis de passage in our own analyses. Instead, we have tried to further clarify the distinction between deep and superficial plis de passage in the text as follows (lines 418-424):

“Indeed, Bodin and colleagues (2021) showed that, in the superior temporal sulcus (STS), the presence of plis de passage was associated with a higher density of U-shaped fibers connecting adjacent gyri. Superficial plis de passage, i.e., complete sulcal interruptions with a gyral gap visible on the cortical surface, showed an even higher local connectivity advantage than deep plis de passage, the complete gyral gaps or partial interruptions with sulcal pinching which are buried within the depth of the sulcus. (Please note that in our study, we focus on superficial gyral gaps visible on the cortical surface only, and do not classify deep plis de passage.)”

Comment (7) Line 96 - replace "blindly from" with "without awareness of" or "while .... were concealed".

The text was edited accordingly (lines 102-103 and 781) - thank you.

“[...] the OTS and MFS were delineated manually in each hemisphere, without awareness of the reading scores [...] Sulcal labeling was conducted without awareness of the children’s reading scores [...]”

(8) Line 547 - it is my impression that the findings of morphological differences in dyslexia are quite mixed. This gives the impression that the studies reported are representative of quite a vast number of studies and I don't think that is the case. Perhaps a sentence here with some more scepticism or alluding to the diverse findings in dyslexia?

It was indeed not our intention here to simplify or summarize a broad literature on morphological differences in dyslexia, but rather to cite a few example results in support of the preceding assertion “various other aspects of early brain structure have been proposed to constrain the development of reading skills”.

We clarified this sentence as follows (lines 659-662):

“[L]ongitudinal studies following pre-reading children at familial risk of dyslexia showed for instance that the surface area of the left fusiform gyrus (Beelen et al., 2019) and patterns of grey matter volume around the VWFA (Skeide et al., 2016) could be predictive of future dyslexia diagnosis.”

Comment (9) Finally, there is quite a lot of supplemental information, some of which is very interesting. I would recommend that the authors check what could be usefully transferred to the main body of the article. eLife uses figures and figure supplements followed by Appendices and the authors might want to check this before deciding on the final version of record.

We thank Reviewer 2 again for this suggestion to improve our manuscript. We have chosen to move several figure panels for Supplementary Figure 5 to Figure 3, following this suggestion, and given that our previous explanations in the paragraph concerning reading predictors appeared confusing to both reviewers (cf. see Rev 2 recommendation comment (1) and rev 1 recommendation comment (4)).

While figure supplements seem handy for reading the article directly on the eLife website, we chose to stick with the more standard practice of one appendix or supplementary file gathering all the additional information, which seems more accommodating of different reading practices overall.

References for Response to Reviews:

Borst G, Cachia A, Tissier C, Ahr E, Simon G, Houdé O (2016) Early Cerebral Constraints on Reading Skills in School-Age Children: An MRI Study. Mind, Brain, and Education 10:47–54.

Cachia A, Dupont C, Palmis S, Menu I, Auzias G, Sein J, Coulon O, Longcamp M (2025) The sulcal patterns of the occipito-temporal and anterior cingulate cortices influence reading and writing in children and adults. Cereb Cortex 35:bhaf145.

Cachia A, Roell M, Mangin J-F, Sun ZY, Jobert A, Braga L, Houde O, Dehaene S, Borst G (2018) How interindividual differences in brain anatomy shape reading accuracy. Brain Structure and Function 223:701–712.

Del Mauro G, Del Maschio N, Abutalebi J (2022) The relationship between reading abilities and the left occipitotemporal sulcus: A dual perspective study. Brain and Language 235:105189.

Ferrer E, McArdle JJ, Shaywitz BA, Holahan JM, Marchione K, Shaywitz SE (2007) Longitudinal Models of Developmental Dynamics Between Reading and Cognition From Childhood to Adolescence. Developmental Psychology 43:1460–1473.

Peng P, Kievit RA (2020) The Development of Academic Achievement and Cognitive Abilities: A Bidirectional Perspective. Child Development Perspectives 14:15–20.

Reuter M, Schmansky NJ, Rosas HD, Fischl B (2012) Within-subject template estimation for unbiased longitudinal image analysis. NeuroImage 61:1402–1418.

Roell M, Bellon E, Polspoel B, Declercq M, De Smedt B (2022) Investigating the association between variability in sulcal pattern and academic achievement. Sci Rep 12:12323.

Ünal ZE, Greene NR, Lin X, Geary DC (2023) What Is the Source of the Correlation Between Reading and Mathematics Achievement? Two Meta-analytic Studies. Educ Psychol Rev 35:4.

Yao JK, Mitchell JL, Davison A, Yeatman JD (2026) Sulcal anatomy of ventral temporal cortex and reading development. :2026.04.06.716640 Available at: https://www.biorxiv.org/content/10.64898/2026.04.06.716640v1.

eLife Assessment

This study shows that the compound Yoda potentiates the function of the PIEZO2 channel, challenging the assumption that they selectively activate PIEZO1. This is an important finding and the evidence supporting this claim is solid. The authors have addressed the methodological and conceptual issues raised during review. Overall, this work will be of broad interest to researchers working with PIEZO channels across various biological scales.

Reviewer #1 (Public review):

[Editor's note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

Summary:

In this work, T. Wijerathne et al. investigated and reported the agonistic effect of Yoda1 and Yoda2 over PIEZO2 function using patch clamp electrophysiology, Ca2+ imaging, and molecular dynamics. They find that Yoda1 sensitizes PIEZO2 to membrane tension, can induce Ca2+ influx, and decreases its inactivation to a lesser degree than it does to PIEZO1 channels. Additionally, their data shows that Yoda2 sensitizes PIEZO2 channels to membrane indentation to a greater extent, but it has a weaker effect on channel inactivation than Yoda1. Interestingly, they report that a mutation in a conserved arginine between PIEZO channels can be used to abolish PIEZO1-mediated Ca2+ flux in response to Yoda molecules. As a whole, the results presented here should be put into perspective against previous and future works involving systems where both PIEZO1 and PIEZO2 might be expressed. This is especially true for works where Yoda1 has been used as a basis for determining the absence of PIEZO2.

Strengths:

The authors use multiple techniques to investigate how Yoda molecules affect the three most important biophysical aspects of PIEZO channels that, when changed, result in pathophysiological responses: a) sensitivity to mechanical stimuli, b) Ca2+ entry, and c) channel inactivation. Lastly, they find a specific amino acid/region that could be exploited for drug design and/or development.

Reviewer #2 (Public review):

Summary:

This manuscript challenges the long-standing assumption that Yoda1 and Yoda2 are PIEZO1-selective activators. Using patch-clamp electrophysiology and calcium imaging in HEK293TΔPZ1 cells overexpressing PIEZO2, the authors demonstrate that Yoda1 potentiates PIEZO2 stretch-activated currents to a similar extent as PIEZO1 and slows PIEZO2 poking-current inactivation (albeit with lower efficacy). They further show that the more potent analog Yoda2 affects PIEZO2 at nanomolar concentrations and use mutagenesis and molecular dynamics simulations to propose that Yoda2's benzoic acid group forms a transient salt bridge with R1724 in the putative Yoda binding pocket, explaining its enhanced potency.

Strengths:

The authors are established Piezo/biophysics experts; the study is highly important, technically competent, and carries significant implications for the reinterpretation of prior work that used Yoda compounds as PIEZO1-selective probes.

The core finding that Yoda1 modulates PIEZO2 stretch currents is convincing and important.

Reviewer #3 (Public review):

Summary:

The manuscript reports that Yoda1 and Yoda2 agonize PIEZO2 in a manner similar to PIEZO1, increasing open probability and stretch sensitivity, but the mechanism underlying this sensitivity is incomplete. Mutagenesis was shown exclusively in PIEZO1, with no corresponding mutagenesis in PIEZO2, so the proposed mechanism in PIEZO2 is inferred by homology rather than directly tested. All experiments use mouse PIEZO2, and the human ortholog should be used before generalizing the proposed reinterpretation of the field.

Strengths:

The pressure-clamp electrophysiology demonstrating a shift in half-activation pressure for PIEZO2 is compelling evidence in support of the central claim.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this work, T. Wijerathne et al. investigated and reported the agonistic effect of Yoda1 and Yoda2 over PIEZO2 function using patch clamp electrophysiology, Ca2+ imaging, and molecular dynamics. They find that Yoda1 sensitizes PIEZO2 to membrane tension, can induce Ca2+ influx, and decreases its inactivation to a lesser degree than it does to PIEZO1 channels. Additionally, their data shows that Yoda2 sensitizes PIEZO2 channels to membrane indentation to a greater extent, but it has a weaker effect on channel inactivation than Yoda1. Interestingly, they report that a mutation in a conserved arginine between PIEZO channels can be used to abolish PIEZO1mediated Ca2+ flux in response to Yoda molecules. As a whole, the results presented here should be put into perspective against previous and future works involving systems where both PIEZO1 and PIEZO2 might be expressed. This is especially true for works where Yoda1 has been used as a basis for determining the absence of PIEZO2.

Strengths:

The authors use multiple techniques to investigate how Yoda molecules affect the three most important biophysical aspects of PIEZO channels that, when changed, result in pathophysiological responses: a) sensitivity to mechanical stimuli, b) Ca2+ entry, and c) channel inactivation. Lastly, they find a specific amino acid/region that could be exploited for drug design and/or development.

Weaknesses:

The methods and discussion sections are lacking enough detail to fully evaluate the findings and put them into perspective, respectively.

We have included new results, expanded our discussion, and added additional details in our Methods section.

Reviewer #2 (Public review):

Summary:

This manuscript challenges the long-standing assumption that Yoda1 and Yoda2 are PIEZO1selective activators. Using patch-clamp electrophysiology and calcium imaging in HEK293TΔPZ1 cells overexpressing PIEZO2, the authors demonstrate that Yoda1 potentiates PIEZO2 stretch activated currents to a similar extent as PIEZO1 and slows PIEZO2 poking-current inactivation (albeit with lower efficacy). They further show that the more potent analog Yoda2 affects PIEZO2 at nanomolar concentrations and use mutagenesis and molecular dynamics simulations to propose that Yoda2's benzoic acid group forms a transient salt bridge with R1724 in the putative Yoda binding pocket, explaining its enhanced potency.

Strengths:

The authors are established Piezo/biophysics experts; the study is highly important, technically competent, and carries significant implications for the reinterpretation of prior work that used Yoda compounds as PIEZO1-selective probes.

The core finding that Yoda1 modulates PIEZO2 stretch currents is convincing and important. However, several conceptual, methodological, and presentational issues need to be addressed before acceptance, as detailed below.

Weaknesses:

(1) The abstract states that Yoda1 potentiates PIEZO2 "as efficaciously as PIEZO1." This claim is accurate only for stretch currents and single-channel open probability, but the paper itself demonstrates important asymmetries: i) Yoda molecules slow PIEZO2 poking-current inactivation ~2-fold, versus ~5-10 fold for PIEZO1 (Figure 3b and ref #60). ii) Spontaneous Ca²⁺ entry via PIEZO2 requires non-physiological conditions (high extracellular Ca²⁺, hypertonic solutions) that are unlikely to occur in native cells. The abstract should be revised to clearly qualify where equivalence holds and where efficacy differences exist. IMO, the current wording risks overcorrecting the historical bias (PIEZO1-only) by going too far in the other direction.

We have changed our language throughout the manuscript to convey more accurately differences of Yoda agonism between the two Piezo homologs.

(2) Related concern: the PIEZO2 Ca²⁺ signal in Figure 2 is only detectable using a Ca²⁺-boosted solution (CBS ie 30 mM Ca²⁺). Physiological extracellular Ca²⁺ and cells normally do not experience sustained hypertonicity at these magnitudes. The authors should explicitly clarify that the practical implication of their findings is primarily for electrophysiological (patch-clamp) experiments and that the Ca²⁺ imaging caveat applies only under amplified conditions. Specifically, the authors should state that in standard Ca²⁺ imaging assays with physiological buffers, PIEZO2 is unlikely to confound Yoda1 results.

Agreed. As mentioned above, we have modified our wording to convey this point more clearly.

Related point: Can cytochalasin D (CytoD) restore a Yoda1-dependent Ca²⁺ signal in physiological saline? This would help determine whether the weak PIEZO2 response is primarily a membrane tension issue (cytoskeletal tethering) versus intrinsically lower channel expression or permeability. The authors already have tagged PIEZO1/2 constructs and could, in principle, normalize by surface expression.

We have repeated our acute pharmacological treatment in cells pre-treated with CytoD (or 0.5% DMSO vehicle control). These experiments were done in normal saline solution. We found that Yoda1 produces a very small (but significant) calcium influx in cells treated with CytoD. This result strengthens our conclusion that Yoda1 is capable of producing an activatory effect on Piezo2 in absence of external mechanical stimulation - albeit this effect is very small and only detectable in non-physiological (calcium-rich or destabilized cytoskeleton) conditions. It remains difficult to definitely conclude about why Yoda molecules exert different agonistic effects. We can only discuss possibilities. Differences in Ca2+ permeability is unlikely to contribute to this difference, as pore properties and selectivity for monovalent ions are similar between the two channels (see PMID: 39543876). Slight differences in how Yoda bind the Yoda binding region could explain why Yoda1 robustly modulates Piezo2 stretch currents but fails to produce robust calcium influx, as recent work suggests that these two agonist effects may originate from two distinct Yoda binding poses in this region (PMID: 38060566; PMID: 41402296). Piezo2’s intrinsically high threshold for stretch-activation could also represent a contributing factor, as figures 1e-f show that Piezo2’s midpoint activation pressure in the presence of Yoda1 (about -30 mmHg vs. -60 mmHg) might still be too high to enable robust Yoda-induced activation at the resting membrane tension of the cell. Since Piezo1 has been consistently shown (by us and others) to be expressed at a higher level than Piezo2 in transfected HEK293 cells (e.g., PMID: 20813920; PMID: 32227462, PMID: 37102984; PMCID: PMC13423835), we have not replicated these experiments with Piezo1. We have modified our manuscript to include these new results and discussion.

(3) The mean inactivation tau values for wild-type PIEZO2 poking currents in both DMSO and Yoda1 conditions (Figure 3b, approximately 15-40 ms range) appear substantially higher than values reported in published literature (typically 5-10 ms; eg, PMID: 20813920). This discrepancy needs to be addressed.

This discrepancy likely arises because inactivation kinetics of Piezo1 and/or 2 poking currents can be modulated by many biological and experimental variables that are difficult to control. Biological variables include membrane potential (PMID: 20813920), extracellular pH (PMID: 25561736), temperature (PMID: 35858335; PMID: 31413193), the presence of modulatory proteins (PMID: 37590348) and extracellular matrix (PMID: 28164706). Experimental variables include the mechanics of probe displacement (velocity and impact angle) and the precise positioning of the time-cursor near the peak current for exponential fitting. In addition, the decaying portion of Piezo poking currents is likely contributed by both inactivation and deactivation, as a recent study demonstrates relaxation of the poke stimulus during mechanical indentation in N2a cell (PMID: 42649207). Cellular viscoelastic properties and cytoskeletal organization are thus anticipated to dictate how fast this relaxation occurs, influencing the proportion of inactivating and deactivating currents in poking recordings. Since this point deviates from our main goal - to test an effect of Yoda on channel kinetics measured side-by-side – we prefer not to discuss it.

(4) The authors perform all MD simulations on a truncated PIEZO1 model and justify this choice by noting that the Yoda binding region is highly conserved between homologs. This is a reasonable and defensible starting point given the availability of well-validated PIEZO1 simulation set ups in their lab. A few points are nonetheless worth addressing: While PIEZO2 simulations are not strictly required, the authors are encouraged to briefly discuss whether any long-range structural differences between PIEZO1 and PIEZO2 (outside the binding site itself) could influence Yoda2 binding dynamics, particularly in light of the chimera data showing that PIEZO2 sequence in repeat A abolishes Yoda1 sensitivity. This reviewer still doesn't understand the reason behind this discrepancy despite it being acknowledged in the text.

We have now conducted three replicas of all-atom MD simulations of the Yoda2-bound PIEZO2 system. The new results show that the equivalent Arg residue in PIEZO2 (R2029) also forms a salt bridge with Yoda2, consistent with our observations from the Yoda2-bound PIEZO1 simulations. Taken together, these simulation results seem to point to the fact that Yoda molecules bind to a similar region in both PIEZO1 and PIEZO2. The lack of Yoda1-sensitivity in the chimera is quite puzzling, and we can only speculate about why. The swapped chimeric region includes 103 residues, 21 of which are non-conserved between PIEZO1 and PIEZO2 and distributed around the binding region (PMID: 29795280). These mutations could collectively and allosterically alter the binding free energy landscape in a manner that either 1) strongly reduce Yoda binding affinity or 2) equalize Yoda binding affinity for both curved and flat conformations, effectively rendering Yoda molecules silent binders (similar to how we previously described Dooku1 binding to Piezo1 PMID: 38060566).

Another MD-related comment is that three simulation replicas (which is impressive for such a big system) show markedly different salt bridge occupancy (82.6%, 49.7%, 99.8%; stated in the text). This wide variation suggests incomplete sampling in at least one replica. The authors should provide RMSD plots for ligand and protein backbone to assess convergence and possibly discuss whether the 49.7% replica represents a genuinely distinct binding mode or incomplete equilibration.

We have extended PIEZO1-Yoda2 replica #2 from 600 to 1400 nanoseconds. This replica shows that the salt-bridge turned ON and OFF during the simulation because the Arg side chain tends to flip. This flipping behavior is evident in the 20 µs Replica #1 (Figure 6c). In Figure 6-figure supplement 1, we analyzed our previous five replicas of MD simulations using the PACE force field. We show that the angle distributions indeed vary between replicas and subunits at the timescales currently affordable by MD technique for this large system. Therefore, it is not too surprising that the salt bridge occupancy differs between three all-atom simulations. We now provide RMSD of ligand and protein backbone for each Replica as Figure 6 supplement 1. The time evolution of Yoda2 binding poses across all simulation replicas are now illustrated in Figure 6-figure supplement 2, which shows that Yoda2 is mobile, yet remain stable in this overall hydrophobic binding site, consistent with our previous report (PNAS2023 PMID: 38060566). We also added all corresponding analysis of Yoda2-PIEZO2 in the same figures.

(5) The Discussion proposes that PIEZO2's weaker Ca²⁺ response to Yoda1 could partly reflect lower membrane expression. Since the authors already have fluorescently tagged PIEZO1 and 2 constructs, a simple fluorescence intensity comparison between the two (acknowledging it would reflect total rather than surface expression) could provide at least indirect support for this claim. Alternatively, if such a comparison is not feasible, the authors may consider removing membrane expression from the list of proposed explanations or explicitly acknowledging that this remains unsubstantiated speculation. The max poking currents may somewhat and roughly indicate the level expression difference too, if done exactly side by side.

PIEZO2 expression is known to be lower than PIEZO1 in heterologous systems (PMID: 20813920; PMID: 32227462, PMID: 37102984; PMCID: PMC13423835), hence we have not attempted to test if this was the case with our tagged constructs.

(6) The abstract or concluding remarks should highlight that Dooku1 is not PIEZO1-selective in its agonist-like action on PIEZO2, and that Cmpd15/Cmpd64 appear to be better PIEZO1-selective tools. This nuance is buried in the Results section.

We now list these results in the abstract.

(7) The authors should not cite PMID 31015490. Clearly, any work on MCC13 is confounded by the overwhelming expression of PIEZO1 (PMID: 42084270). Instead, the authors should also cite the literature from others who have clearly recorded stretch currents from PIEZO2 before the cited studies (eg, PMID: 37590348).

Agreed!

Reviewer #3 (Public review):

Summary:

The manuscript reports that Yoda1 and Yoda2 agonize PIEZO2 in a manner similar to PIEZO1, increasing open probability and stretch sensitivity, but the mechanism underlying this sensitivity is incomplete. Mutagenesis was shown exclusively in PIEZO1, with no corresponding mutagenesis in PIEZO2, so the proposed mechanism in PIEZO2 is inferred by homology rather than directly tested. All experiments use mouse PIEZO2, and the human ortholog should be used before generalizing the proposed reinterpretation of the field.

Strengths:

The pressure-clamp electrophysiology demonstrating a shift in half-activation pressure for PIEZO2 is compelling evidence in support of the central claim.

Weaknesses:

(1) In the single-channel recordings (Figure 1a), it's unclear how many channels were present in those patches. After applying -60 mmHg pressure, multiple channels would be activated (as seen in Figure 1e). The number of channels in the patch and their inactivation rate could significantly influence the open probability in such experiments. To overcome this, in the original Yoda1 article (Syeda, Ruhma, et al. eLife 2015), no additional pressure was used.

To clarify the confusion, we record single-channel PIEZO2 currents less than 24 hrs after transfection and record macroscopic PIEZO2 currents 2-3 days after transfection. Hence, unlike the cited study, our single and macro currents are obtained from different patches. When patching cells for single recordings, most patches do not show any mechano-currents and those that do typically show only one current level of single-channel conductance through the recording, even when the patch is highly pressurized.

Additionally, the reported open probability comparison (n=7 Yoda1 vs n=17 DMSO patches) has an SEM nearly as large as the effect itself (0.30 {plus minus} 0.11), consistent with a small number of outliers driving this. The underlying mean open and shut times are reported without any statistical test; only the derived open probability receives a p-value. Additionally, in Figure 1a, the Yoda1 condition noise is different from the control. This should be stated if noise filtering was applied and how, given that this could affect open probability analysis.

This is a fair point. We have removed two outliers and have performed statistical analysis on mean dwell times. The effect of Yoda1 is still highly significant. We have replaced current traces with two of similar noise levels. We acquired all single-channel current traces at 100kHz, filtered them at 5kHz, and displayed them at 500 Hz.

(2) The calcium imaging data in Figure 2 raise significant concerns regarding the chemical activation claim. The calcium-boosted solution (30 mM Ca2+) is not physiological and appears to be generally stressing cells rather than specifically activating PIEZO2: the control condition under CBS already shows an elevated signal, consistent with cells being unwell at this calcium concentration, and adding Yoda1 on top of this shifted baseline raises further questions about specificity rather than confirming it. Separately, it is unclear why DMSO alone produces measurable PIEZO2-associated calcium influx in HBSS, a result that is not addressed in the text. Figure 2 should clearly indicate when DMSO/Yoda1 perfusion was initiated, and y-axis labels are missing from panels A and B.

We agree and have toned down our chemical activation claim, insisting that this activation is very weak in absence of external mechanical stimulus, being only detectable under extreme, non-physiological conditions. We also re-analysed our data (and added fluorescence scale bars, thank you) showing time-course and scatter plots from the same cells in Figure 2a. We think the “bump” seen upon DMSO perfusion in PIEZO2 cells could be amplified by the low expression of GC6 in these cells. We now discuss this point in our revisions.

(3) In the poke experiments, an activation threshold should be calculated and reported, and amplitude data (e.g., peak current versus indentation depth) should be shown rather than only inactivation tau values. It is also unclear why mClover3- and N-GFP-tagged constructs were used in these experiments, since electrophysiological recording already confirms channel expression without requiring a fluorescent tag.

Valid point, we are now comparing peak currents for different poking distances. We have removed our Tau analysis for our tagged-PIEZO2 clones.

(4) For inactivation kinetics (Figure 3b), the authors use unpaired comparisons across separate cells, whereas the deactivation experiments (Figure 3c) use paired; it should be applied to the inactivation experiments as well. Deactivation kinetics for PIEZO2 itself should be shown.

These experiments have different goals which do not require using the same protocol. Deactivation kinetics experiments were used for the chimera because this parameter was shown by us to be the most sensitive to Yoda1. Hence, if a Yoda modulation exists, but is hard to detect in the chimera, using this protocol with paired samples provide us with the highest detection sensitivity. This level of sensitivity was not needed to detect an effect of Yoda1 on PIEZO2.

If the claim is that Yoda1 acts on PIEZO2 through the same mechanism proposed for PIEZO1, then a PIEZO1/2 chimera should be expected to show a corresponding effect on deactivation tau; instead, this chimera is reported as completely Yoda1-insensitive despite both parental channels being Yoda1-sensitive, as shown in this study.

This discrepancy is acknowledged in the text. We do not know why the chimera is Yoda insensitive and can only speculate at this point. Of note, one must acknowledge that this chimeric swapping introduces 21 amino acid changes (vs. WT PIEZO1) which may have profound allosteric effect on binding affinity and/or efficacy. MD studies would be useful to explore this possibility in a future study.

(5) Given that this reflects a different experimental paradigm for Yoda EC50, PIEZO1 should be included within Figure 4b. Additionally, EC50 bar plots should be present on this figure.

Showing EC50 plots here is not possible because each cell was exposed to only one dose of Yoda1 (unpaired measurements). We have included a Yoda1 dose-response curve against PIEZO1 inactivation Tau from our prior study (PMID: 35858335) for comparison as a Figure Supplement.

The inactivation time constant for PIEZO2 without Yoda1 is inconsistent across figures, below 20 ms in Figure 3b but above 20 ms in Figure 4c.

As explained above in our answer to Reviewer #2, these inconsistencies likely occur because PIEZO inactivation kinetics are susceptible to modulation by a number of biological and experimental factors that are not easily traceable, making it difficult to define absolute channel kinetics.

(6) Finally, the modeling is performed exclusively on PIEZO1, whereas the manuscript's central focus is PIEZO2. It is therefore unclear whether the proposed structural mechanism, including the basis for Yoda2's reduced efficacy on PIEZO2, can be directly extrapolated to PIEZO2.

We have now conducted three replicate all-atom MD simulations of the Yoda2-bound PIEZO2 system. The new results show that the equivalent Arg residue in PIEZO2 (R2029) also forms a salt bridge with Yoda2, consistent with our observations from the Yoda2-bound PIEZO1 simulations. However, the current simulations do not allow us to distinguish differential efficacy of Yoda2 toward PIEZO1 versus PIEZO2. The reduced efficacy of Yoda2 on PIEZO2 may instead arise from differences in its binding affinity for the open versus closed conformations, or from contributions of cellular components that are not represented in our MD simulations, such as the cytoskeleton.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Main:

(1) Adding more depth and detail to the methods section will help others replicate the findings.

Agreed.

(2) Methods: line 325. Describe the way each compound was prepared for each concentration without undissolved particles. Avoid using the phrase "as described before" in the methods section.

Compounds were dissolved in DMSO and simply diluted in recording solution at the time of experiments with vortex mixing.

(3) Figure 1B (comparison of open probability): only three points out of 7 (with Yoda1) seem outside the data range of the control (DMSO). It seems very unreliable for an agonist. Moreover, in Figure 1F, the data seem very noisy. I would suggest adding more n.

We have eliminated two outliers and reanalyzed our data, showing statistics for mean dwell times and open probability.

(4) The authors should add to the "Yoda1 activates PIEZO2 in the absence of external mechanical stimulus" section (or the methods) how perfusion of Yoda1 was applied.

We used manual pipetting.

(5) Lines 165 - 171. This part should be moved to the discussion and expanded on. Particularly since the reference they allude to comes from the same group.

This discussion part is needed to introduce the chimera experiments and thus should remain in Results.

(6) Figure 3B. The noise range of the time constants of inactivation is a lot. Specially for transfected cells. I would suggest a couple of sentences explaining why the authors think this happens.

This discussion is now included.

(7) Figure 4C. Tau inac of DMSO control seems very high when compared to others. I would suggest a couple of sentences explaining why the authors think this happens.

One difference is that DMSO concentration here is higher (0.25% vs 0.075%) to match the higher concentration of Yoda1 tested. Inactivation kinetics are subject to modulation by many parameters; hence we focus on side-by-side comparisons instead of absolute values.

(8) Line 213. Which standard binding equation? Include in the methods.

Done.

Rethink calculating the EC50 with dose responses that saturate. Figure5C. Are these absolute or relative EC50s?

EC50 of mutants are normalized to WT.

(9) Discussion: "Our discovery that Yoda1 modulates PIEZO2 stretch currents as effectively as PIEZO1's strongly suggests that some of the reported biological roles attributed to PIEZO1 using this agonist could be (at least partly) transduced by PIEZO2. For instance, our discovery seamlessly reconciles the puzzling fact that reducing PIEZO2 expression in cells that co-express PIEZO1 also reduces the effects of Yoda1 (49, 76, 77). Along." In the referenced works (49, 76, and 77) its not checked if PIEZO1 levels remain the same when PIEZO2 was knocked out. Here, the authors should discuss this possibility.

Good point!

Minor:

(1) Indicate in the figure legends, result or methods section, what the error bars represent.

Done.

(2) Indicate the p-value for Figure 1B.

Done.

(3) Figure 2. Include the ionomycin value that each group is normalized against.

We provide now vertical scale bars for all samples.

(4) Why not use PIEZO2 for the MD simulations?

Done.

Reviewer #2 (Recommendations for the authors):

Minor comments

(1) In the Abstract: The sentence "Although a structural framework for PIEZO1 activation by Yoda1 exists" may be confusing for readers outside the field. The authors may mean "mechanistic" or "computational" framework. Please clarify or reword.

We have reworded the abstract.

(2) Residue number in abstract: The abstract states "a conserved arginine in the Yoda binding site" without identifying it. Please add the residue number (R1724 in mouse PIEZO1) for clarity.

Done

(3) Figure 1 legend or in the method: recording mode and filter settings: The figure legend should explicitly state that single-channel recordings were obtained in cell-attached mode, and should report the filter frequency (kHz) and sampling rate used for idealization. This is essential for readers to evaluate dwell-time analysis.

Done, but details are in the method.

(4) Several instances throughout the text use "native" where "naïve" is clearly intended.

Corrected, thank you.

(5) The statement that PIEZO2 requires more mechanical force than PIEZO1 (~−60 vs ~−40 mmHg P1/2) should be interpreted cautiously without this information, and the authors should clarify whether pipette geometry was controlled between PIEZO1 and PIEZO2 experiments.

We have shown that PIEZO2 requires more force to be opened by membrane tension relative to PIEZO1 using MscL gating as a tension-calibration system (PMID: 41703062). Pipette geometry is consistent for each type of experiment (macroscopic or single-channel recordings). We have expanded our Methods.

(6) The finding that Dooku1slows PIEZO2 inactivation similarly to Yoda1 is striking and is currently underplayed. This raises important questions: Does Dooku1 elicit Ca²⁺ entry in PIEZO2-expressing cells? It could at least be discussed further in the text.

Likely not, since Yoda-induced calcium influx is very weak and requires non-physiological conditions to be detectable. But this is worth trying at some point.

eLife Assessment

The study investigates, from multiple angles, the still-debated function of insect rhodopsin-7 (Rh7). The authors present compelling results for its ancient phylogenetic origin across pan-arthropods, a non-visual role based on expression analyses in the fly brain, an unusual G-protein signalling pathway, and - using behavioural genetics - that Rh7 affects how Drosophila melanogaster interprets and responds to light-dark transitions. Through this, the work provides fundamental new insights into the evolution and function of non-visual opsins.

Reviewer #1 (Public review):

Summary:

The study investigates the Drosophila non-visual light receptor rhodopsin7 with regard to its role in light information processing and resulting consequences for behavioral patterns and circadian clock function. Using behavioral, in situ staining, and receptor activation assays together with different fly mutants, the authors show that rhodopsin7 is an important determinant of activity under and response to darkness, which likely signals via a pathway distinct from other, visual Drosophila rhodopsins. Based on phylogenetic analysis, the authors further discuss a potentially conserved functional role of non-visual photoreceptors like rhodopsin7 and the mammalian melanopsin light information processing and circadian clock modulation.

Strengths:

The manuscript follows a very clear structure with all investigations logically building onto each other. Background information and methodology are provided in appropriate detail so that readers can fully understand why and how experiments were conducted. It is further praiseworthy that the authors provide the details that allow also non-experts in the field to fully understand their approaches. Experimental work was conducted in a highly standardized manner, and also considered potential "side-aspects" like the consequences of temperature cycles and changed photoperiods. The detailed and clear description of the obtained results makes them very convincing, with (almost) all observable patterns being addressed.

By highlighting the evolutionary old phylogenetic position of rhodopsin7 and its conservation across numerous clades, the authors provide strong reasoning for the relevance of their work, also pointing out the similarities to the mammalian melanopsin. The postulated hypothesis regarding protein structure and functioning, as well as the role in light information processing and behavioral and circadian clock modulation are well based on the authors' observations, and speculative aspects are correctly pointed out.

Weaknesses:

Where the manuscript still has potential for improvement is the discussion, which in its current form does seem slightly self-contained and does not fully integrate the findings of previous studies on Drosophila rhodopsin7. As the introduction specifically points out that previous findings have been contradictory, this seems like a missed opportunity. Further details on this are provided in the recommendations below.

Similarly, the manuscript currently lacks a discussion of the possible relevance of rhodopsin7 (and other non-visual light receptors in other organisms) in the context of a species' environment and lifestyle, i.e., what is the relevance/benefit of having rhodopsin7 in the fly's everyday life? While this clearly involves speculation, when done carefully, it can elevate the paper's relevance from a primarily academic to a societal one.

An additional point concerns the title and abstract, which postulate rhodopsin7 roles in contrast vision as well as motion and brightness perception. Contrast remains poorly defined in the text, leaving it ambiguous whether it refers to bright/dark contrasts, e.g., along edges, or the temporal contrast that results from dark pulses (startle response). While the latter seems to apply here, the former is likely more intuitive. Thus, this aspect should be rephrased (also in the title) or properly clarified early on. Regarding motion detection, this is backed up by the optomotor response results, but the findings stand somewhat isolated from the other results, lacking a clear connection aside from general visual processing. Lastly, brightness perception is mentioned in the abstract, but never again, possibly due to inconsistent phrasing throughout the manuscript.

Reviewer #2 (Public review):

Summary:

This is a very interesting paper bringing new and important information about the poorly understood rhodopsin 7 photoreceptive molecule. The very ancient origin of the gene is revealed in addition to data supporting a signaling pathway that is different from the one known for the canonical rhodopsins. Precise expression data, particularly in the optic lobe of the fly, as well as clear behavioral phenotypes in responses to light changes, make this study a strong contribution to the understanding of the still-debated function of rhodopsin 7.

Specific comments

(1) Title and abstract: Contribution of Rh7 to circadian clock regulation

a) It is not that clear to me what rhodopsin does in terms of circadian regulation (even though its function might be circadianly regulated). The clear role in the light/dark distribution of activity might not be circadian per se, but mostly light/dark- driven and there is no evidence here for a role in the entrainment of the clock.

b) The authors should cite Lazopulo, which nicely shows that Rh7 has an important role in peripheral neurons to allow flies to escape from blue light (see below).

(2) Figure 2 C

The finding showing that Galphaz but not Galphaq can trigger signaling from light-excited Rh7 is a very intriguing finding to better understand Rh7 function. Since Galphaz is related to Gi/o, it would be interesting to test those, for example, by expressing RNAi with Rh7-gal4 and testing the Light-dark or light-off response behavior.

(3) Figures 3-4

The change in the locomotor activity distribution between light and dark in LD conditions provides a nice assay for Rh7 function. Since Lazopulo et al. (2019) have shown that wild-type but not Rh7 mutants do escape from blue light, it would be important to compare and discuss these LD behavior data with the Lazopulo results. Precisely, is this nighttime preference linked to blue light?

The expression data are really nice and show that Rh7 is mostly a non-retinal photoreceptor. However, the paper would be strongly reinforced by correlating this with the LD behavior. The LD phenotype should be tested in flies with Rh7 expression rescued under Rh7gal4 control (as done for the startle response). This is important to show whether the expression pattern is likely responsible for the described Rh7 function in LD. If L5 and or M11 drivers are available, they should be used to rescue Rh7? Since expression in some clock neurons is shown, the rescue experiment should also be done with a clock neuron driver.

In the same line, can the LD phenotype (or startle response phenotype of Figure 4) be restored by expressing Rh7 under ppk control, as shown for the blue light avoidance phenotype by Lazopulo et al?

Finally, the Rh7 "darkfly" rescued flies should be tested in LD.

Reviewer #3 (Public review):

Summary:

While our knowledge regarding visual opsins is largely very good, a lot more uncertainty exists around the role of non-visual opsins. Using the power of the Drosophila melanogaster model system, Kirsh et al. investigate the role of the non-visual opsin Rhodopsin7 (Rh7). Expression analysis, based on Rh7-Gal4>UAS-GFP and HRC in situ staining, reveals strong expression in the optic lobes and somewhat weaker, but nevertheless extensive expression in the brain. An investigation of motor activity reveals that loss of function leads to an altered day and night rhythm, specifically decreasing activity during the dark phase. These flies were also less sensitive, but still responsive to a light-induced startle response and showed deficiencies in the optomotor response. To further investigate how Rh7 may modulate these responses, inspired by the Dark line of flies (which were kept in the dark for ~1400 generations) and which has accumulated C-terminal related losses, the authors conducted rescues with an intact and a C-terminal-deficient Rh7 and were able to pinpoint that region as an important driver of related behavioral shifts. These findings are particularly intriguing as Rh7 represents an ancient opsin with phylogenetic and mechanistic parallels to mammalian melanopsin.

Strengths:

The paper is well-written and contains high-quality data with appropriate sample sizes, and the conclusions are well supported.

Weaknesses:

No weaknesses were identified by this reviewer, but the following recommendations are made:

(1) The authors should clarify exactly what tissues were taken for the comparative qPCR. This is particularly interesting in terms of the retina. Since Rh7 appears not to be expressed within the photoreceptor cells of the retina, this raises the important question as to which cells it is expressed in. To address this important question, it would also be helpful to include an expression analysis of the retina itself (by extending the RH7-GFP expression patterns and/or adding HCR in situ of the ommatidia array). The cell types of the retina are very well classified, and some evidence already exists for Rh7 expression in support cells (e.g., Charlton-Perkins et al., (2017); PMID: 28562601). This study has a unique opportunity to investigate this further by adding these critical data for a more complete picture of Rh7.

(2) Mammalian opsins should be included in the phylogenetic analysis illustrated in Figure 2A and indicate their position on the tree. This will allow readers to better put the authors' statements regarding the intermediate position of Rh7 into perspective. In addition, note that the distinction between red and deep red is easy to miss regarding the Rh7 cluster. Perhaps the authors could use a more distinct colour scheme, for example, orange and deep red.

(3) More details should be provided on the optomotor response experiments. Specifically, specifications of the frequencies used for the optomotor response are needed. Results show a relatively large level of variation, which may be due to different angular perspectives that flies may have had while viewing the stimulus. If possible, provide videos as examples, as they will make it clearer to viewers how much flies could move around in the setup (from the methods, it seems they could move within the 2.2 of the 3 cm diameter of the arena, which would lead to substantial differences in the visual angle of the viewed grating.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

(1) Where the manuscript still has potential for improvement is the discussion, which in its current form does seem slightly self-contained and does not fully integrate the findings of previous studies on Drosophila rhodopsin7. As the introduction specifically points out that previous findings have been contradictory, this seems like a missed opportunity.

We thank the reviewer for this helpful suggestion and agree that the previous version of the Discussion did not sufficiently integrate earlier studies on Drosophila RH7. In the revised manuscript, we have expanded and restructured the discussion to explicitly incorporate key studies, including those by Ni et al. (2017), Senthilan et al. (2019), Lazopulo et al. (2019), and Meyerhof et al. (2024). In Discussion 5.4, we directly address similarities and discrepancies, such as differences in reported expression patterns, behavioral phenotypes, and experimental conditions. We use these inconsistencies to better position our findings and propose that RH7 acts as a modulatory photoreceptor with context dependent effects, which may explain the variability observed across studies. These changes provide a more integrated and coherent view of RH7 function.

(2) Similarly, the manuscript currently lacks a discussion of the possible relevance of rhodopsin7 (and other non-visual light receptors in other organisms) in the context of a species' environment and lifestyle, i.e., what is the relevance/benefit of having rhodopsin7 in the fly's everyday life?

We agree that the ecological relevance of RH7 deserved a more detailed discussion. According to the suggestion, we have now expanded the discussion to address the ecological and functional relevance of RH7 in the context of environmental light conditions and species-specific lifestyles (see Section 5.5). In particular, we now highlight that RH7 is frequently retained in species inhabiting sun-exposed environments and that its presence correlates with conditions of high light exposure, especially during early developmental stages. Together with our functional data in Drosophila, showing a role of RH7 in light dependent behavioral flexibility and circadian regulation, this supports the idea that RH7 contributes to the detection and integration of environmental light cues under natural conditions.

(3) An additional point concerns the title and abstract, which postulate rhodopsin7 roles in contrast vision as well as motion and brightness perception. Contrast remains poorly defined in the text, leaving it ambiguous whether it refers to bright/dark contrasts, e.g., along edges, or the temporal contrast that results from dark pulses (startle response). While the latter seems to apply here, the former is likely more intuitive. Thus, this aspect should be rephrased (also in the title) or properly clarified early on.

Thank you for this comment. We agree that the term “contrast” was ambiguous, as it can refer to spatial contrast (e.g., edge detection) or temporal contrast (changes in light intensity). Our study focuses on temporal contrast, such as light-dark transitions. Accordingly, we have revised the manuscript, replacing “contrast” with the more precise term “temporal changes in light intensity” throughout the relevant sections. In addition, to clarify this further, we conducted additional experiments with graded changes in illumination. These showed that Rh7⁰ mutants have reduced responses compared to wild type in both abrupt (e.g. 30→0 µW/cm²) and gradual (e.g. 30→5 µW/cm²) transitions, indicating that RH7 contributes to detecting temporal changes in light intensity.

(4) Regarding motion detection, this is backed up by the optomotor response results, but the findings stand somewhat isolated from the other results, lacking a clear connection aside from general visual processing.

We have clarified the connection between motion detection and the other visual assays. In the Results section (4.7), motion detection is now described as “relying on the processing of spatiotemporal changes in luminance across the visual field,” explicitly linking it to general luminance-based visual processing. In the Discussion (5.3), we added the concise statement “These computations rely on changes in light intensity over space and time” to further emphasize this connection. We believe these changes make the relationship between the optomotor results and the other assays clearer and easier to follow.

(5) Lastly, brightness perception is mentioned in the abstract, but never again, possibly due to inconsistent phrasing throughout the manuscript.

We agree that the terminology was not used consistently throughout the manuscript. To improve clarity and consistency, we replaced the term "brightness perception" in the Abstract with "temporal changes in luminance", which more accurately reflects the focus of our study.

Reviewer #2 (Public review):

(1) It is not that clear to me what rhodopsin does in terms of circadian regulation (even though its function might be circadianly regulated). The clear role in the light/dark distribution of activity might not be circadian per se, but mostly light/dark- driven and there is no evidence here for a role in the entrainment of the clock.

We thank the reviewer for this important point. We agree that our current data does not provide sufficient evidence for a direct role of RH7 in circadian regulation or clock entrainment. In this revision, we have examined LD (12:12) behavior in more detail and report changes in several activity parameters (Figure 3A). While these differences could in principle involve circadian mechanisms, our experiments do not allow us to distinguish between circadian effects and acute light-driven modulation of locomotor activity. We therefore present these results without interpreting them as evidence for a circadian function of RH7, and instead focus on light-driven behavioral effects under LD conditions.

(2) The authors should cite Lazopulo, which nicely shows that Rh7 has an important role in peripheral neurons to allow flies to escape from blue light (see below).

We thank the reviewer for pointing this out. We have now cited Lazopulo et al. (2019) at multiple relevant points in the manuscript, specifically regarding atypical RH7 expression in peripheral neurons, the functional effects of RH7 outside the LNvs, and its protective role in light-dependent behavior (discussion 5.1, 5.4, and 5.5). We believe these additions ensure that the manuscript accurately reflects the existing literature and appropriately acknowledges the contribution of Lazopulo et al. to understanding RH7 function.

(3) The finding showing that Galphaz but not Galphaq can trigger signaling from light-excited Rh7 is a very intriguing finding to better understand Rh7 function. Since Galphaz is related to Gi/o, it would be interesting to test those, for example, by expressing RNAi with Rh7gal4 and testing the Light-dark or light-off response behavior.

We agree that testing additional Gi/o-type Gα subunits would be highly informative for further dissecting RH7 signaling. However, in our preliminary experiments, RNAi-mediated knockdown of Rh7 itself did not result in a robust or statistically significant behavioral phenotype under our conditions. This is consistent with previous reports that RNAi in Drosophila can show variable efficiency and often results in incomplete knockdown, leading to weak or undetectable phenotypes (Dietzl et al., 2007; Green et al., 2014; Perkins et al., 2015). Given these limitations, we did not further pursue RNAi-based approaches targeting additional Gα subunits in the current study. Instead, we complemented our experimental findings with additional comparative and in silico analyses using FlyPredictome, which support a potential association between RH7 and Gi/o-type signaling components. These results have now been incorporated into the revised manuscript (Discussion 5.2). We agree that further functional validation of Gi/o-type G protein involvement, for example using genetic mutants, interaction assays or alternative approaches, would be an important direction for future studies.

(4) The change in the locomotor activity distribution between light and dark in LD conditions provides a nice assay for Rh7 function. Since Lazopulo et al. (2019) have shown that wildtype but not Rh7 mutants do escape from blue light, it would be important to compare and discuss these LD behavior data with the Lazopulo results. Precisely, is this nighttime preference linked to blue light?

To address this point, we conducted additional experiments, which are now included in the supplementary materials. Instead of the usual white light–dark (LD) cycles, we tested blue light–dark and green light–dark conditions. However, we observed only minor differences and no significant shift in light preference, as reported by Lazopulo et al. (2019). A likely explanation for this discrepancy is the difference in experimental light intensity. In our assays, we used moderate light intensities, approximately 30 µW/cm² for white light and 5–6 µW/cm² for color lights. Lazopulo et al. used substantially higher light intensities (2480 µW/cm² for white light and approximately 600 µW/cm² for colored lights). Even the higher intensities used in our arrhythmia experiments (~120 µW/cm²) remained well below those applied in their study.

Importantly, the behavioral paradigms differ: Lazopulo et al. assessed acute light preference/avoidance, whereas our experiments quantify locomotor activity distribution across the day under LD conditions. The latter is strongly shaped by circadian regulation and integrates multiple internal and external factors. Therefore, wavelength-dependent avoidance behavior observed under acute conditions may not directly translate into changes in daily activity patterns under moderate LD conditions. Although we cannot directly replicate the strong avoidance of blue light or preference for green light described by Lazopulo et al., our results do not contradict those of Lazopulo. The reduced activity during the midday phase in the wild-type flies, which is less pronounced in Rh7⁰, may suggest a potential protective function or preventive behavior in the WT flies. Since the moderate light intensities used in our LD assays are much less aversive, no strong wavelength-dependent preference was observed.

(5) However, the paper would be strongly reinforced by correlating this with the LD behavior. The LD phenotype should be tested in flies with Rh7 expression rescued under Rh7gal4 control (as done for the startle response). This is important to show whether the expression pattern is likely responsible for the described Rh7 function in LD.

We agree that assessing LD behavior in flies with Rh7 expression rescued under Rh7-Gal4 control is important to link expression pattern to function. We have now performed these experiments. The Rh7-Gal4–driven rescue results in a substantial recovery of the LD locomotor phenotype, bringing it closer to control levels than the UAS-empty condition. While the rescue does not fully reach wild-type levels, the observed trend and significant differences across specific parameters support a functional contribution of Rh7-expressing cells. These data are now included in Figure 5C and described in section 4.6.

(6) If L5 and or M11 drivers are available, they should be used to rescue Rh7? Since expression in some clock neurons is shown, the rescue experiment should also be done with a clock neuron driver.

Thank you for this helpful suggestion. We agree that cell type-specific rescue using L5- and/or Mi1 drivers, as well as clock neuron drivers, would be highly informative.

However, implementing these experiments is currently not straightforward. Targeting these neuronal populations requires the use of split-Gal4 driver lines, which can differ in expression strength and are therefore not quantitatively comparable across combinations. In addition, the required genotype would be highly complex: the two split-Gal4 components will be located on separate chromosomes, the UAS-Rh7 transgene is on another chromosome, and the Rh7⁰ mutation resides on the third chromosome. Combining all necessary elements in a single line would therefore require extensive genetic recombination. Importantly, we aim to avoid introducing mutations or markers on the X chromosome, as these are known to influence behavioral readouts. However, many of the available lines are in a white^- background, and assembling the full rescue genotype while maintaining a controlled genetic background across all chromosomes is therefore particularly challenging. For these reasons, we have not pursued these rescue experiments in the current study but consider them an important direction for future work.

(7) In the same line, can the LD phenotype (or startle response phenotype of Figure 4) be restored by expressing Rh7 under ppk control, as shown for the blue light avoidance phenotype by Lazopulo et al?

We thank the reviewer for this suggestion. However, in our experiments, we did not observe a blue-light avoidance phenotype comparable to that described by Lazopulo et al. (2019). Since this behavioral effect did not occur in our assay, we did not perform rescue experiments with ppk-regulated Rh7 expression. As previously discussed, the absence of pronounced blue light avoidance in our experimental setup is likely due to the moderate light intensities we used, which are significantly lower than those applied in Lazopulo et al.

Therefore, a ppk-mediated rescue would not be meaningful in our context.

(8) Finally, the Rh7 "darkfly" rescued flies should be tested in LD.

We agree that testing Rh7Darkfly rescue flies under LD conditions provides important insight into Rh7 function. We have now performed these experiments. The Rh7Darkfly rescue leads to a marked shift of the LD locomotor phenotype toward the Darkfly phenotype, with a stronger effect than observed in the Rh7 rescue lines. While the rescue does not fully recapitulate either wild-type or Darkfly controls, the observed trend supports a contribution of RH7 to the modulation of LD behavior in this genetic background. These data are now included in Figure 5C.

Reviewer #3 (Public review):

(1) The authors should clarify exactly what tissues were taken for the comparative qPCR. This is particularly interesting in terms of the retina. Since Rh7 appears not to be expressed within the photoreceptor cells of the retina, this raises the important question as to which cells it is expressed in. To address this important question, it would also be helpful to include an expression analysis of the retina itself (by extending the RH7-GFP expression patterns and/or adding HCR in situ of the ommatidia array). The cell types of the retina are very well classified, and some evidence already exists for Rh7 expression in support cells (e.g., Charlton-Perkins et al., (2017); PMID: 28562601). This study has a unique opportunity to investigate this further by adding these critical data for a more complete picture of Rh7.

We thank the reviewer for this helpful comment. We have now clarified in the Materials and Methods exactly which tissues were collected and analyzed for the comparative qPCR experiments. Regarding Rh7 expression in the retina, in our hands, we did not detect signals in retinal support cells, likely due to low expression and masking by autofluorescence. However, low Rh7 mRNA levels detected by qPCR in retinal tissue, together with the faint staining reported by Kistenpfennig et al. (2017), suggest that expression in these cells is possible. We have added this clarification, along with the observation by Lazopulo et al. (2019), who reported Rh7 expression in multidendritic neurons, to the Discussion (section 5.1) to provide a more complete overview of RH7 localization.

(2) Mammalian opsins should be included in the phylogenetic analysis illustrated in Figure 2A and indicate their position on the tree. This will allow readers to better put the authors' statements regarding the intermediate position of Rh7 into perspective. In addition, note that the distinction between red and deep red is easy to miss regarding the Rh7 cluster. Perhaps the authors could use a more distinct colour scheme, for example, orange and deep red.

In the revised Figure 2A, we have included both mammalian melanopsins and lophotrochozoan opsins, which support our interpretation of RH7 as a primitive opsin. The color scheme has been updated to more clearly distinguish the different clades, and tardigrades are now represented with a red triangle rather than a separate color. These changes improve clarity and better highlight the intermediate phylogenetic position of RH7.

(3) More details should be provided on the optomotor response experiments. Specifically, specifications of the frequencies used for the optomotor response are needed. Results show a relatively large level of variation, which may be due to different angular perspectives that flies may have had while viewing the stimulus. If possible, provide videos as examples, as they will make it clearer to viewers how much flies could move around in the setup (from the methods, it seems they could move within the 2.2 of the 3 cm diameter of the arena, which would lead to substantial differences in the visual angle of the viewed grating.

We would like to thank the reviewer for this valid point. We agree that, due to the relatively simple design of our optomotor apparatus, the flies may have different viewing angles of the rotating grid. To illustrate this point, we have added representative video footage and further photographs of the experimental setup to the supplementary material. A more sophisticated optomotor apparatus that maintains a fixed viewing angle for the animals would certainly reduce experimental variability. Nevertheless, our current, simple setup yielded robust and statistically significant optomotor responses. To further strengthen the robustness of these results, we have increased the sample size from n = 20 to n = 30, thereby confirming the observed effects.

Recommendations for the authors:

Reviewing Editor Comments:

(1) Claimed impact of rh7 on circadian regulation.

As the data are presented at the moment, there is no evidence for a role in circadian regulation. At present, the data can be fully explained by the acute effects of light on behavior. Either remove the circadian from the text or consider specific entrainment and rescue experiments (see suggestions from Reviewer 2).

We thank the editor for this important comment. We agree that the original version of the manuscript overinterpreted the role of RH7 in circadian regulation, and that the observed phenotypes can be explained by acute effects of light on locomotor behavior under LD conditions. In the revised manuscript, we have substantially restructured and clarified the interpretation of our findings. We now explicitly frame our results as light-driven behavioral effects under LD (12:12) conditions and have removed claims suggesting a direct role of RH7 in circadian entrainment. Furthermore, we have also added new rescue experiments, including Rh7-Gal4–driven Rh7 and Rh7Darkfly rescue, which provide further functional insight. These data show partial restoration of the locomotor phenotype and genotype-dependent shifts in activity patterns, supporting a modulatory role of RH7 in light-dependent behavior, while not directly addressing circadian clock function.

We have further revised the text to ensure that conclusions remain strictly aligned with the presented data. In addition, we now emphasize that our data provide a basis for future studies exploring the potential role of RH7 in circadian entrainment.

(2) Improved discussion section

Both Reviewers 1 and 2 point out several aspects that should be better taken care of, especially embedding the new results into the context of the existing literature on rh7- Lazopulo et al 2019, Ni et al 2017, Meyerhof et al 2024 were particularly mentioned.

We have addressed the reviewers’ suggestions by revising the Discussion section. We now refer to Meyerhof et al. (2024), Lazopulo et al. (2019), and Ni et al. (2017) at multiple points throughout the Discussion to better place our findings in the context of the existing literature on RH7 function. These references are now included in Sections 5.1, 5.4, and 5.5 of the Discussion.

(3) The addition of vertebrate melanopsins, as pointed out by Reviewer 3 (they suggest mammalian, but consider adding other vertebrate and also lophotrochzoan proteins), in the phylogenetic tree will help to put the evolutionary ancestrality claim into a better perspective.

We performed new bioinformatic analyses using RefSeq proteomes and carefully selected representative species across relevant clades to construct a balanced phylogenetic tree. These additions support our interpretation of RH7 as a primitive opsin and help place its intermediate phylogenetic position into better perspective. The results are presented in the revised Figure 2A, which now also includes mammalian melanopsins and lophotrochozoan opsins.

(4) There were several questions raised on the qPCR and optomotor response experiments (Reviewer 3), as well as the definition of contrast (Reviewer 1).

We have addressed the reviewers’ comments by substantially improving the clarity and presentation of the qPCR and optomotor response experiments, as well as the definition of “contrast.” We now provide a more detailed description of the qPCR analysis in the Methods section. For the optomotor response, we have added representative images and videos of the experimental setup to the Supplementary Material to increase transparency. We further clarified that the term “contrast” was previously used ambiguously and may refer to different concepts. In the revised manuscript, we now consistently refer to “temporal changes in light intensity” to specify that our study focuses on light-dark transitions. In addition, we included additional experiments with graded changes in illumination, which show reduced responses in Rh7 mutants compared to wild type under both abrupt and gradual transitions.

(5) Supplementary Material needs sorting

Accession numbers are in Supplementary File 2, not Supplementary File 1. Supplementary Material S2B is in Supplementary File 3 - this is really confusing. Supplementary Materials S3 is probably Supplementary File 1, and the phylogenetic tree shown there actually looks identical to the tree shown in Figure 2A of the main manuscript.

We sincerely apologize for the confusion regarding the Supplementary Material numbering and file assignments. This issue likely arose during the upload process or transfer from the bioRxiv version of the manuscript. We have now carefully checked and corrected the Supplementary Files to ensure that all accession numbers, figures, and files are correctly assigned and consistently labeled throughout the revised submission.

Reviewer #1 (Recommendations for the authors):

(1) Here, I will address specific aspects of the manuscript, elaborating on the points mentioned in the public review and also listing detailed suggestions.

As mentioned in the public review, the discussion could do a better job of relating the obtained findings to previous publications. For example, the paper by Ni et al. (2017) investigates the role of Drosophila rhodopsin7 in circadian light entrainment and its interplay with dcry, which seems relevant in the context of rhodopsin7 apparently being only present in CRY-negative clock neurons (except DN1a).

We thank the reviewer for this helpful suggestion. We have expanded the manuscript to further explore the potential relationship between RH7 and DCRY. Specifically, we performed additional bioinformatic analyses, which are now presented in Results Section 4.1, and in discussion sections 5.4 and 5.5 addressing this aspect. While previous observations by Ni et al. (2017) and Kistenpfennig et al. (2017) have already pointed to a possible relationship between RH7 and DCRY, our phylogenetic analyses provide additional support for such a connection from an evolutionary perspective. Although our data do not yet allow us to infer the precise functional nature of this relationship, they further strengthen the evidence that Rh7 and DCRY are evolutionarily and functionally linked.

(2) Another relevant paper by Meyerhof et al 2024 (doi:10.1073/pnas.2400964121) investigates the role of rhodopsin7 in shade preference and startle responses to short light pulses. The lack of shade preference and the altered startle response in rh7 mutants could be related to the authors' own findings on altered behavior in darkness. Further, the paper shows rh7 mutants to more rapidly return to baseline activity after a light pulse, which is somewhat reminiscent of the abrupt activity decrease in rh7 mutants at the start of the dark phase (18h). Generally, I recommend using the contradictory findings mentioned in the introduction as a starting point to describe where the new findings expand previous knowledge and possibly resolve contradictions.

We thank the reviewer for this suggestion. We have revised the Discussion to integrate our findings with the available literature, including the recent study by Meyerhof et al. (2024), and explicitly compare similarities and differences between the reported phenotypes and our own observations.

These aspects are now addressed in the newly added Section 5.5, where we discuss the evolutionary retention of RH7 in sun-exposed species and relate our phylogenetic findings to the emerging evidence that RH7 may play a conserved role in mediating responses to sun-exposed and shaded environments.

(3) One specific aspect that does not fully convince me is the interpretation of the behavioral patterns concerning the short photoperiods (Figure 3C). The authors state that rh7 mutants lose the flexibility to shift their activity into the dark phase. However, drawing a vertical line through panels A, B, and C, it rather seems that activity peak times in wildtypes are fixed to 6-7h and 17-18h, while rh7 mutants do actually shift their peak activities with the decreasing photoperiod. A way of quantifying this would be to determine the phase differences between M- and E-peaks. It may also help to align plots in a way so that lightson of long and short photoperiod plots align. There is clearly a difference between WTs and mutants, but I would be more careful in calling this a reduction in flexibility.

We thank the reviewer for this comment. We also noticed that the behavioral patterns under short photoperiods can be interpreted in different ways. Based on this comment, as well as the comments from the other reviewers and the handling editor, we decided not to further discuss the potential circadian component in this manuscript. We agree that additional and more targeted experiments will be necessary to better separate circadian effects from light-induced behavioral responses. These experiments are beyond the scope of the present study but will be important to further investigate the potential role of RH7 in circadian regulation. Therefore, we revised the manuscript to focus on the light-induced behavioral effects of RH7 and avoided drawing conclusions about circadian flexibility that are not yet fully supported by our data. In this process, we removed the data showing actvity under short days from the manuscript.

Below, I list individual points according to their line number in the manuscript:

(1) Line 1: Consider rephrasing the title based on the above comment regarding contrast.

We have revised the title accordingly.

The new title is: RHODOPSIN 7: An ancestral non-canonical photoreceptor shaping light-responsive behavior

(2) Line 10: Vision is also used to adapt to the environment, so maybe change to "...also to otherwise/indirectly adapt behavior...".

We have revised the sentence as suggested to clarify the non-image-forming role of light in behavioral adaptation.

“Animals rely on light not only for vision but also to adapt behavior to their environment through non-image-forming pathways.”

(3) Line 15: Consider rephrasing the title based on the above comment regarding contrast, motion, and brightness perception.

We thank the reviewer for this suggestion. The title has been revised accordingly. The new title is: RHODOPSIN 7: An ancestral non-canonical photoreceptor shaping light-responsive behavior.

(4) Lines 23-25: This reads as if light perception and vision are the same. Please clearly distinguish to make clear that vision is only one way in which light can be perceived and influence behavior.

We have revised the text to clarify that vision represents only one modality of light perception and that light can influence behavior via both visual and non-visual pathways. “The ability to detect and respond to light is fundamental for many organisms and can influence behavior through both visual and non-visual pathways. Vision relies on….”

(5) Lines 26-27: Here and also later in the introduction, it is hard to follow how different groups of opsin light receptors relate to each other (e.g., opsins (in general,) rhodopsins, r-opsins, c-opsins, non-visual opsins), etc.) This is also related to the fact that opsins can be differentially grouped based on phylogeny/structure (e.g., r- vs. c-opsins) or role/localization (e.g., visual vs. non-visual). Please revisit the introduction with a non-expert reader in mind, and try to address the groups hierarchically. In fact, a small tree-like figure of the different groups may be a simple solution.

We thank the reviewer for this helpful suggestion. We have revised the Introduction to simplify the presentation of opsin classification. c-Opsins and r-opsins are now introduced only briefly to place RH7 within the rhabdomeric opsin family. Since Drosophila contains only rhabdomeric opsins and these are the focus of our study, we did not further expand the Introduction with a detailed discussion of c-opsins or add a classification figure.

(6) Line 33: Please expand to "...image formation in the eye, some..." to be more precise. We have revised the sentence to clarify that rhodopsin-mediated image formation refers to visual systems.

“While the primary function of rhodopsins is image formation in visual systems, some are also known to participate in non-image-forming functions.”

(7) Lines 46-55: This detailed information about the different Drosophila rhodopsins is not referred to anywhere else in the manuscript. I suggest shortening and focusing on rh7.

We thank the reviewer for the suggestion. We have shortened the section to focus on RH7, as the other rhodopsins are not discussed elsewhere in the manuscript. The revised text now highlights the debated expression and potential function of RH7.

“Drosophila melanogaster (Dmel) has seven rhodopsins (RH1-RH7), all belonging to the ropsin group. While RH1–RH6 are expressed in well-characterized photoreceptors, the expression pattern and function of RH7 remain less clear and are still debated.”

(8) Line 55: The statement that the expression of Rh7 remains debated also occurs later in the introduction (line 88). If in line 55, it is about the expression sites specifically, references should be added. Otherwise, it may be sufficient to only address this aspect in line 88.

The statement in line 55 specifically refers to the expression sites of RH7. References to Ni et al., 2017 and Senthilan et al., 2019 have been added to support this point.

(9) Line 61: Which of these features distinguishes rhodopsins from other opsin groups?

We thank the reviewer for this comment. The RH7 section has been thoroughly revised to clarify both the features shared with other rhodopsins and those that distinguish RH7 from other opsin groups. Key distinguishing features, including the extended N- and C-terminal tails and the absence of the QAKKMNV motif in ICL3, are now highlighted at the beginning of the paragraph, while additional Drosophila-specific features, such as the palindromic RCSI, are discussed subsequently.

“RH7 displays the canonical rhodopsin features, including seven transmembrane helices, a chromophore-binding lysine, and the LRTPXN motif located in the first intracellular loop (ICL1)….”

(10) Figure 1: For panel B (left), please increase label sizes. For panel B (right), the color scale makes it very hard to see low-level expression (e.g., in Mi15 and below). Please try adjusting the color scale and/or consider only showing the cell types also indicated in the left part of the panel. For ninaE, please add "aka Rh1" or similar for the less informed reader.

We have revised Figure 1 accordingly. Specifically, we increased the label sizes in panel B (left) to improve readability. For panel B (right), we adjusted the color scale to better visualize low-level expression and improve overall clarity. In addition, we updated the labeling of ninaE to “Rh1 (ninaE)” to aid accessibility for a broader audience.

(11) Lines 88-90: What exactly were the contradictions that emerged from previous studies? Also see the comment in the public review regarding the discussion of the results.

Thank you for pointing this out. We have clarified this in the revised manuscript. “Previous studies on RH7 have yielded conflicting results regarding its function and expression, including discrepancies in the identified neuronal populations and associated behavioral phenotypes”

(12) Line 91: Briefly explain the characteristics of white and yellow mutations so the reader can assess why these are relevant here.

We have clarified the role of the white and yellow genes in the genetic background. The revised text now reads:

“Both genes are expressed in the optic system (Davis et al., 2020; Supplementary Material) and are known to affect neuronal function and behavior (Bond et al., 2024; Drapeau et al., 2003; Rickle et al., 2025).”

In addition, we have visualized the expression data for yellow and white from Davis et al. in the Supplementary Material for more clarity.

(13) Lines 100-101: Circadian entrainment is not a typical visual function. Please rephrase.

The sentence has been revised for clarity and now reads:

“As an ancestral rhodopsin, RH7 appears to contribute to basic visual functions, such as contrast processing, as well as to non-visual functions like circadian synchronization, similar to melanopsins in mammals.”

(14) Line 105: The exact address is not necessary, mention BLASTP is enough.

Thank you. We have now removed the exact address and we now simply refer to BLASTP.

(15) Line 108-109: If default settings were used, the details are not needed.

We have omitted the details, as they are not necessary.

(16) Line 113: These settings seem to repeat line 108. Needed?

We thank the reviewer for pointing this out. In the original version, the settings in line 108 referred to the BLAST search, while those in line 113 pertained to the phylogenetic tree, where the parameters are essential, because phylogenetic inference depends on dataset- and method-specific choices. However, this entire section has now been revised, as the phylogenetic tree has been completely rebuilt with a much broader dataset, including mammalian melanopsins and lophotrochoan r-opsins.

(17) Line 122: Without further context, these abbreviations are non-informative and confusing to many readers. Provide details or rephrase. The same applies to DMEM, FBS, and Glo22F in the same paragraph. While many readers will understand instantly, keep the diverse audience of eLife in mind.

We have now added explanations for all abbreviations in the paragraph to improve clarity for a broad audience. Specifically, the Gα subunits (GNAS, GNAL, GNAQ, GNA11, GNA12, GNA13, GNAZ) are now clearly defined, and DMEM, FBS, and Glo22F are also explained as Dulbecco’s Modified Eagle Medium, fetal bovine serum, and a cAMP-sensitive luciferase reporter, respectively. These changes should make the methods clearer and more accessible to all readers.

(18) Line 144: "constant conditions" is not ideal as it could be misinterpreted as DD or LL. Please rephrase.

We agree with the reviewer and have removed the phrase “constant conditions” to avoid potential misinterpretation.

“All recordings were performed under 60% relative humidity and 25 °C…”

(19) Line 148: Stating Lux values for Drosophila is not informative as the unit is based on the spectral sensitivity of the human eye. Ideally, provide a spectrum of the used light source as a supplement.

We thank the reviewer for this suggestion. We agree that lux values alone are not informative for Drosophila, and we have now included the spectrum of the light source in Figure 3–figure supplement 1A.

“A standard day was defined as 12 h light (30 µW/cm² ~ 100 lux, the spectrum of the light is shown in Figure 3–figure supplement 1A) and 12 h darkness”. In addition, to ensure complete consistency of the experimental conditions, we repeated some earlier experiments using the same light source as in the rest of the study, since the exact light source used in the initial experiments could no longer be verified. This ensures that all behavioral experiments presented in the revised manuscript were performed using the same light source.

(20) Lines 149-153: The startle response protocol is hard to understand when only reading it. It could help to state when (relative to lights-on) the first dark pulse was given and how many dark pulses were given per light phase.

We have clarified the timing and frequency of the dark pulses in the Methods section. We now specify that the first dark pulse was applied 1.5 h after lights-on and that six dark pulses were delivered during each 12 h light phase.

(21) Lines 159ff: The 4-hour startle response assay is not clearly distinguished from the previously mentioned assay (30 min D + 1.5h L). I could help to start with e.g., "In a second assay..." and to clearly state why this additional assay was performed. That is, what information can the 4-hour assay provide that the previous assay could not?

We have revised the text to clearly introduce the 4-hour startle response assay as a second, distinct paradigm (“In a second startle response assay…”).

We also clarified its purpose: in contrast to the initial assay, this paradigm was designed to measure the acute startle response independently of circadian influences. To this end, flies were rendered arrhythmic and exposed to repeated 4-hour light–dark cycles, preventing re-entrainment and isolating the startle response.

In addition, we now refer to a schematic of the light regime in Figure 3–figure supplement 1C.

(22) Lines 161-163: As flies were kept in LL aside from the dark pulses, it could be more intuitive to talk about 72h or three 24h intervals rather than 3 days. Otherwise, one might assume that the initial constant light was followed by 3 days of LD plus dark pulses.

We have revised the text to emphasize the total duration (“72 h”) to avoid confusion with a 24-hour LD regime. In addition, we now provide a schematic of the light regime in the Supplementary Material (Figure 3–figure supplement 1C).

(23) Line 172: Please write out CantonS once, as otherwise it only occurs in Table 1. Also, clarify that y+ and w+ refer to the (absence of the) yellow and white mutation, respectively.

We have revised the text to write out Canton-S (CS) upon first mention and clarified that y⁺ and w⁺ denote wild-type alleles of the yellow and white genes, respectively, meaning any mutations in these genes are absent.

(24) Table 1: Consider adding "(truncated RHSOPSIN7)" or similar to the "Darkfly" cell.

We have updated Table 1 to indicate that the Darkfly line carries a truncated RH7. The description now reads: “This line has been maintained for decades in constant darkness and carries numerous mutations, including a truncated RH7 lacking the last 21 amino acids.”

(25) Table 2: In the formatted paper, Table 2 should appear after section 3.5 so the reader can properly interpret the applications and names.

We have placed Table 2 immediately after section 3.5 as suggested.

(26) Table 3: Consider adding the product IDs of the used antibodies for convenience.

We have added the product IDs of all antibodies used to Table 3 for convenience.

(27) Lines 180ff: This section should initially state the purpose of establishing the rh7-Gal4 line, i.e., the expression of selected gene in rh7-expressing cells. Again, keep your (lee-informed) readers in mind.

We have revised the section to clarify the purpose and design of the Rh7-Gal4 line. It now states that the line was generated to identify Rh7-expressing cells and explains the use of a relatively large promoter region including exon 1 and 2.

(28) Line 190 w1118 flies do not occur in Table 1. While used by a commercial company here, please still include them for completeness.

We have added the w¹¹¹⁸ line to Table 1 for completeness, as well as the BDSC Stock #8621 line, which was used as the background for the UAS-construct flies.

(29) Lines 192ff and 203ff: As for lines 180ff, start with a brief description of what the purpose of the respective works was.

We have revised the section to start with a brief description of the purpose of the rescue experiments. The text now clearly states that UAS constructs were generated to assess the ability of different Rh7 variants to rescue the phenotype, with UAS-Rh7 as the wild-type rescue, UAS-Rh7^Darkfly carrying the Darkfly variant, and UAS-empty serving as a control. We have added an introductory sentence to the optomotor section in Materials & Methods to clarify the purpose of the experiments: “To investigate the contribution of Rh7 in the lamina and medulla to visual motion detection, male flies were individually tested for their optomotor walking behavior.” This makes clear why the assay was performed before describing the setup and measurements.

(30) Line 225: Just to be safe, make sure to not mix up Δ-CT and 2-ΔΔCT (see Livak & Schmittgen 2001, doi: 10.1006/meth.2001.1262).

We confirm that ΔCT was indeed used to calculate relative mRNA levels. To clarify this, we have now added the formula to the text: ΔCT = CT_α-tubulin − CT_Rh7.

(31) Line 229: Please state the light (cycle) conditions under which the flies were raised, as well as the sampling times at which they were fixed.

We have now specified the light cycle conditions under which the flies were raised (12 h light:12 h dark at 25 °C) in the Methods section. For qPCR experiments, tissues (retina, lamina, and brain) were carefully dissected from adult flies without prior fixation to preserve RNA integrity. Dissections were performed during the light phase and were not restricted to specific time points. These details have been added to the revised manuscript (section 3.10).

(33) Lines 296-297: This brief description could go at the start of section 3.2.

The sentence describing the GsX assay (“To test the functionality of RH7, we used a GsX assay designed to determine which Gα proteins can bind to a given GPCR, in this case RH7 (Ballister et al., 2018)”) has been moved to the Methods section (now 3.3) as a brief introduction to the approach.

(34) Line 305: Clarify that 2 different retinals were tested.

We clarified that two different retinals were tested. The text now reads: “We observed robust activation with the Gαz chimera, with both 9-cis-retinal and all-trans-retinal (Figure 2C).”

(35) Figure 2A, B: I suggest using a different color for the RH3,4,5, branch for better distinction from RH7.

We have generated a new tree and adjusted the branch colors to improve clarity between RH3,4,5, branch and RH7.

(36) See comment above regarding the shifting/flexibility of WTs vs. rh7 mutants.

Following the reviewers and editor's recommendations, we revised the manuscript to focus on the light-dependent effects of RH7 and no longer interpret our data in terms of shifting ability or circadian flexibility. The corresponding statements have been removed or revised accordingly.

(37) Line 348: Please change "occur" to "be triggered".

The wording has been updated as suggested, replacing “occur” with “be triggered” in the text.

(38) Line 355: I would mention the double mutants only after the results from the rh7 single mutants were described.

We have adjusted the order of the text accordingly and now introduce the double mutants only after describing the Rh7 single mutants.

(39) Line 362: See comment on line 159ff.

We have added a sentence at the beginning of the section to clearly introduce the 4-hour startle response assay as a second, distinct paradigm: “In a second startle response assay (4-hour assay), designed to measure the acute response independently of circadian influences…” This clarifies its purpose and distinguishes it from the initial assay.

(40) Line 368: Write "rh7 mRNA" to be precise.

We have updated the text to use “Rh7 mRNA” for precision, as suggested.

(41) Figure 5: If possible, please sort the panels in the order in which they are mentioned in the text. For panel E (right), the orientation of the lower schematic, relative to the upper one and the left panel, is unclear. Is it possible to add orientation indicators? Alternatively, could L5 and Mi1 neurons be indicated in the upper schematic instead?

We thank the reviewer’s helpful suggestion. Since reorganizing the figure panels was challenging and would have affected the overall layout, we instead revised the corresponding text to ensure that the order of panel descriptions matches the figure arrangement. To address the concern regarding panel E, we have added orientation indicators to the schematic to clarify the relative positioning.

(42) 419 Add "A)".

We have added an A).

(43) Line 432: Later on (line 598), it is mentioned that the other accumulated mutations in Darkfly may also have affected the observed patterns. Are there any specific mutations worth highlighting in this context aside from rh7?

We thank the reviewer for this comment. In response, we have expanded the Discussion to highlight additional mutations in the Darkfly genome that may contribute to the observed phenotype. Nonsense mutations in HisCl1 and Vha100-1 could affect visual processing and neuronal signaling, potentially altering sensory input and synaptic function. Missense mutations in DopR, bgm, and developmental regulators such as Sox21b may further modulate activity, sleep, and neural development. We have updated the Discussion accordingly to make these points clear.

(44) Lines 435-436: Unchanged mRNA levels do not necessarily equal the presence of the translated protein. Is there another way to properly confirm this? E.g., rh7 HCR in Darkfly?

We agree that unchanged mRNA levels alone do not confirm the presence of translated protein. While in situ HCR does not directly measure protein levels, it provides additional information on the cellular localization and presence of Rh7 transcripts beyond qPCR measurements. We have therefore performed Rh7 in situ in Darkfly and included this data in Figure 5—figure supplement 1B.

(45) Line 439: "activity" is too vague based on only the altered dark pulse response. I would rather suggest factors like increased protein concentrations/stability, increased sensitivity, constitutive activation, etc.

By “activity,” we refer to the molecular activity of RHODOPSIN 7 as reflected in the behavioral response. The Results have been revised to:

“When we exposed the Darkfly to dark pulses in a 12:12 light-dark cycle, they exhibited the opposite phenotype to Rh7⁰ mutants, showing a stronger startle response (Figure 6A), possibly reflecting RHODOPSIN 7 adopting a constitutively active form in the Darkfly strain, driving the enhanced behavioral response.”

(46) Lines 442-446: The no-rescue control does not fully resemble WTs as they lack the gradual activity decrease after lights-off. The abrupt decrease (also in the 2 rescues) is more reminiscent of the rh7 mutants. This could be discussed in the context of what was mentioned at the start of this section (Meyerhof et al 2024 paper).

We agree that the rescue lines do not fully restore all aspects of the WT control locomotor pattern. One possible explanation is that the Rh7-Gal4 driver does not fully represent the endogenous Rh7 expression pattern. As described in the Results section, our in situ hybridization analysis identified additional Rh7-expressing cells, particularly in the optic system, that were not detected by the Rh7-Gal4 line (results, section 4.5). We have added this point to the Discussion to address how the broader Rh7 expression pattern may contribute to the observed behavioral effects (discussion 5.4).

(47) Line 472: "restored the reduced startle response" reads a bit convoluted. Better write e.g., "restored the enhanced startle response that was absent in the no-rescue mutants".

We have revised the text as suggested to clarify the phrasing.

(48) Lines 533-535: Please provide a reference for the statements about L5 and Mi1 functional roles. Also, this is the only part where there is some connection made between rh7 (expressed in Mi1) and contrast processing (which involves Mi1). As this is a rather weak connection, please consider the last paragraph of the public review section.

We have added references to support the functional roles of L5 and Mi1 neurons. L5 neurons relay signals from outer photoreceptors to the medulla and show strong contrast adaptation (Matulis et al., 2020). Mi1 neurons contribute to motion and contrast processing (Behnia et al., 2014; Matulis et al., 2020). To address the connection between Rh7 expression and contrast processing, we have carefully phrased the text to indicate a potential role of Rh7 in modulating contrast-related signals in Mi1, while noting that the functional link remains to be directly tested. We believe this wording, together with the added references, provides proper support for the statements and maintains a cautious interpretation.

If contrast perception shall indeed be investigated, this may be done by performing the optomotor assay with gray stripes of different contrast, with stripes where the edges are not sharp but show a gradual change from black to white.

We agree that contrast perception can be assessed in more detail using stimuli with defined contrast levels and gradual luminance transitions. However, our aim was not to directly measure contrast sensitivity, but rather to investigate behavioral responses to changes in luminance. We have clarified this distinction in the revised manuscript and now specify that the optomotor response reflects processing of spatial and temporal changes in luminance, while the light-off startle response measures responses to abrupt changes in luminance.”

(49) Lines 546-547: Are there references on how activity or behavior in general can affect external light cues?

We have added appropriate references to support this statement. In particular, we now cite Rieger et al. (2003), which demonstrates direct effects of light on locomotor activity (masking), and Helfrich-Förster (2017), which provides a comprehensive overview of circadian regulation and light-dependent behavioral modulation. These additions clarify how external light cues can directly influence behavior in addition to circadian clock control.

(50) Lines 557-560: See comment above regarding the shifting/flexibility of WTs vs. rh7 mutants.

We thank the reviewer for this comment. We agree that assessing shifting and behavioral flexibility would require dedicated experiments, such as jet lag or entrainment assays. We have therefore revised the focus of the manuscript and now primarily discuss RH7 function in light-dependent behavioral responses rather than broader flexibility of the circadian system. Future studies will further explore the potential contribution of RH7 to behavioral adaptation and flexibility using dedicated experimental approaches.

(47) Lines 564-567: How many species possess dcry but not rh7? This is an important distinction, as otherwise one could assume that both must always co-occur. Alone, this provides only limited support for a potential functional interplay. The findings by Ni et al. (2017) regarding dcry and rh7 in entrainment could, however, help strengthen this point.

We have clarified the evolutionary relationship between DCRY and RH7 in the revised manuscript. We now specify that, although DCRY is widespread across Bilateral and many non-panarthropod species possess DCRY without RH7, RH7 and DCRY show a strong evolutionary association within arthropods, with approximately 98% of DCRY-retaining arthropod species also retaining RH7. At the same time, the presence of RH7 in some lineages lacking DCRY suggests that RH7 may retain functional relevance in the absence of DCRY. Together with previous findings on the functional interaction between RH7 and DCRY in light adaptation (Ni et al., 2017), these evolutionary patterns support a conserved connection between the two light-sensing systems, while the exact nature of this relationship remains to be further investigated. These points are now shown and discussed in section 4.1 and 5.5.

(48) Lines 579-580: Here, one could explicitly mention that this is supported by the lack of diel startle response modulation in the per mutants (Figure 4A).

We have modified the text to explicitly highlight this point. The sentence now reads: “Dark-pulse experiments show that both wild-type controls and per⁰ mutants, which lack a functional circadian clock, exhibit a strong startle response, consistent with the absence of diel modulation of the response in per0 mutants (Figure 4A).”

(49) Lines 596-597: The increased night activity in Darkfly seems largely limited to the first half of the night (compare e.g. Figure 6A and Figure 3A). Again, the loss/alteration of rh7 seems to have its major effect after a transition from light to dark (see lines 442-446, start of the section).

We thank the reviewer for pointing out this issue. The wording in the previous version could have been misleading. By “dark phase,” we meant the 30-minute periods without lighting during the dark-stimulus experiments, and not the “night phase.” We have revised the text to clearly refer to the startle response.

We also clarify that the increased locomotor activity observed in the “Darkfly” strain outside of the startle response is likely caused by additional background mutations that accumulated during long-term rearing in constant darkness, and not exclusively by the RH7 deletion.

(50) Line 622: See comments regarding contrast and motion perception. Do the N- and Cterminal regions of RH7 and melanopsin share similarities in sequence or additional characteristics aside from phosphorylation sites? This could support a conserved role.

We thank the reviewer for this suggestion, which prompted us to perform a more detailed analysis of conserved features between RH7 and melanopsins. In addition to the extended and potentially highly phosphorylatable N- and C-terminal regions, we identified several conserved sequence motifs, including two in the C-terminal region and others within the transmembrane core. One of these motifs extends into the conserved DRY motif in the second intracellular loop (ICL2). These findings are now described in the Discussion (section 5.7) to highlight potential structural similarities beyond phosphorylation sites. The corresponding sequence alignments are provided in the supplementary material.

eLife Assessment

This study provides important information on the role of ErbB4 expression in parvalbumin-positive cells in relation to olfactory behaviour and circuit dynamics in the olfactory bulb. The central question is both timely and novel, and the findings could shed light on the critical role that ErbB4 may play in modulating olfactory bulb cell function and olfactory perception. The authors conducted a comprehensive set of experiments which provide solid evidence in support of their conclusions. However, some concerns remain regarding the technical approaches used, such as spike sorting and control/baseline comparisons.

Reviewer #1 (Public review):

Summary:

The authors have addressed many of the concerns. In sum, this study examined how ErbB4 signalling in parvalbumin-positive (PV) interneurons contributes to olfactory bulb (OB) microcircuit function and olfactory behaviour in mice. Using RNAscope, single-cell RT-PCR, conditional Erbb4 deletion, awake in vivo recordings, slice electrophysiology and optogenetics, the authors report that ErbB4 was expressed in PV interneurons of the EPL and IPL, and that its loss impaired odour discrimination and dishabituation, reduced GABAergic transmission onto mitral cells, and elevated ongoing LFP power. The experiments were well-designed, but several central claims were not yet adequately supported by the data as presented.

Strengths:

Pretty comprehensive set of experiments.

Weaknesses:

(1) Figure 1. The investigation time of the AAV-Cre-GFP group varied considerably across isoamyl acetate concentrations. Were there any differences between the AAV-GFP and AAV-Cre-GFP groups at the 0 (mineral oil) condition? Since the threshold was defined by a comparison against mineral oil, a baseline difference would propagate directly into the threshold estimate and could reduce exploratory drive rather than an olfactory deficit.

(2) Previous papers have reported that ErbB4 is expressed in neuroblasts of the SVZ and RMS, and that disruption of NRG1-ErbB4 signalling may decrease GABAergic neuron number and alter distribution (Anton et al., Nat Neurosci 2004; Ghashghaei et al., PNAS 2006). The authors showed in Figure 2L that ErbB4 expression in the OB was already decreased at P7, suggesting reduced neuron number and disrupted neuronal distribution. The authors should therefore quantify the number and laminar distribution of the different neuronal types in the OB. Cresyl violet staining showed that gross lamination was preserved, but could detect changes in cell density or in the relative proportions of interneuron subtypes.

(3) Figure 4C. First of all, there were no indicators marking the sessions, so we had no idea what state the mice were in. Also, why did the odour-evoked MFRs of PV-Erbb4⁺/⁺ and PV-Erbb4⁻/⁻ mice show opposite trends during the 0-2 s window in the first session - were there any differences in the behaviour of the two groups during this session? As the mice were head-fixed on a floating ball, locomotion and sniffing should have been monitored, since both strongly influence M/TC firing.

(4) Figure 4K-L. The authors did not explain why spontaneous and odour-evoked MFR were so much different among the "no response" M/TCs. Did this mean that ErbB4 knockout in PV interneurons disrupts the function? This would imply a tonic rather than a stimulus-locked disinhibition, which is a different claim from the one made in the Discussion. In addition, the authors stated that panel 4L indicates "the distribution of types of odour-evoked responses was significantly altered in PV-Erbb4⁻/⁻ mice compared with PV-Erbb4⁺/⁺ mice, suggesting compromised odour detection" - what exactly does this mean? A shift in the proportions of response categories is also compatible with a changed baseline firing rate altering the classification itself.

(5) Figure 5. The power of all frequency bands in PV-Erbb4⁻/⁻ mice was higher than in PV-Erbb4⁺/⁺ mice, including the gamma band. However, Tan et al. (2022) reported that inhibition of ErbB4 kinase activity decreases LFP power and GABA release, which is inconsistent with the present data. This discrepancy should be addressed explicitly. A uniform elevation across all four bands was atypical for a PV-dependent mechanism and could reflect a global scaling difference; separating periodic from aperiodic components and reporting normalised power would help.

(6) Photocurrent amplitude and mCherry⁺ PV cell density was not compared across genotypes; Figure 9-figure supplement 1B-C showed photocurrents for only one group. Since the authors themselves reported in Figure 7-figure supplement 1 that PV interneurons in knockout mice reduced mEPSC frequency and amplitude and increased AP frequency, the presynaptic population was clearly not in an equivalent state. The reduced oIPSC could therefore result from differences in ChR2 expression level or in the number of labelled cells. Photocurrent curves and mCherry⁺ cell density must be provided for both groups.

Reviewer #2 (Public review):

The authors aimed to understand how the transcription factor Erbb4 contributes local circuit dynamics and affects performance. This is an interesting question because Erbb4 is largely localized to PV cells in the olfactory bulb (OB) and upregulated following exposure to an olfactory discrimination task.

Strengths:

They have now validated their local, PV-specific knockout strategy, which more directly tests their hypothesis.

The inclusion of optogenetically evoked presynaptic responses from PV cells onto M/T cells also strengthened the conclusion that this synapse was altered.

The authors also now included slice recordings from PV cells, which helps to form a more comprehensive picture of the circuit.

Weaknesses:

(1) The data used to support their spike sorting approach is now included but is difficult to evaluate. In Figure 4 - supp 1B, they show what appears to be mean + sem of concatenated waveforms from one tetrode, but this is not explained anywhere. Moreover, the timescale seems to be off by an order of magnitude. It also seems that the spike widths in panel E are bimodally distributed, with many of the half-widths being more consistent with fast-spiking cells. There is a high density of PV cells in the same layer where they record M/T cells, so is it not possible that some of the units were PV cells (note: this was suggested in the previous review as a potential exciting direction to further validate their findings)? It is still unclear how the authors confidently record from only M/T cells when there is no mention of how other units (e.g., fast-spiking cells) are excluded. The methods state that only M/T cells can be recorded from this layer, which is an oversimplification and appears at odds with the apparent diversity in waveforms and responses.

(2) Upon review, the authors acknowledge that the PV-Erbb4+/+ and PV-Erbb4-/- mice were recorded in separate batches where the control mice were exposed to unwanted artefacts related to a solenoid valve. While we don't know how much that change influenced the results, even with careful analysis, it does draw attention to a major weakness in study design where controls and experimental mice are not run simultaneously. The presence of an artefact from the solenoid valve (perhaps a noise?) immediately before odor presentation for controls, but not experimental mice, should be considered by the reader when it comes to evaluating the odor-evoked responses.

Overall, I feel the manuscript is much improved and more solidly demonstrates that Erbb4 expression in PV cells plays a major role in regulating local OB circuit dynamics and influences odor processing. The remaining weaknesses surrounding the in vivo single unit recordings from PV-Erbb4+/+ versus PV-Erbb4-/- mice reduce my confidence in the validity of the results, but the general agreement of the in vitro work is more re-assuring.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

The study by Hu et al. investigated the role of olfactory ErbB4 in regulating olfactory information processing. The authors demonstrated that ErbB4 deletion impairs odor discrimination, sensitivity, habituation, and dishabituation by using an impressive combination of techniques from morphological to electrophysiology (both slice and in vivo) and from viral injection to cell-type-specific mutation to behavioral analysis. The findings underscore the crucial role of ErbB4 in olfactory PV neurons in modulating mitral cell function and odor perception.

Strengths:

This study contains a pretty comprehensive set of experiments.

Major concerns:

(1) Line 151 page 7, "PV-Erbb4+/+ mice (generated by crossing PV-Cre mice (Wen et al., 2010) with loxP-flanked Erbb4 mice". Does this mean mice carrying PV-Cre and ErbB4 floxed allele? Or with the WT allele? This is confusing. Figures 2B and 2C, ErbB4 expression was evident in many cells that were not positive for PV. What are the identities of those cells? Are they important?

Thank you for your positive comments regarding our study. PV-Erbb4^+/+ mice carried the PV-Cre transgene with the WT allele. In the revised manuscript, we have simplified the description of the genotype for clarity and have provided detailed breeding information in the "Materials and Methods" section (Line 532).

As shown in Figures 2B and 2C, ErbB4 expression was evident in GL and GCL cells that were not positive for PV. The activation of ErbB4-expressing cells in the GL and GCL has been suggested to upregulate theta and gamma oscillations, respectively (PMID:35853717). Mammalian OB are the most abundant and varied interneurons among brain tissues, and their functional importance and diversity remain unclear (PMID:31358754 and PMID:32208366). ErbB4-expressing PV interneurons in the EPL and IPL were relatively less abundant but accounted for the majority of inhibitory inputs to M/TCs. Importantly, one PV interneuron connects with over 1000 M/TCs to mediate recurrent and lateral inhibition, linearly controlling the OB output (PMID:31358754 and PMID:39186804). This study provides evidence that ErbB4 signalling in PV interneurons is required for olfactory discrimination and dishabituation.

(2) In Figure 4, the authors performed tetrode recordings in awake head-fixed animals. Although individual neuron spikes could be obtained by spike-sorting, this is not a "single-unit" experiment due to the nature of this approach.

What is the odor used in Figure 4? How did the authors clean up the odor to limit the stimulation within 2 seconds? In what layer were the tetrodes placed? What is the putative cell type presented in Figure 4C? If Figure 4C is a representative neuron recorded, the odor-induced suppression of spike activity seems to be impaired in PV-ErbB4-/- animals. However, Figure 4D shows that suppressed neurons were similar between the two types of animals. Such comparisons among individual mice are difficult for in vivo electrophysiological experiments because the recorded cell type and placement of electrodes would be different. The authors should apply ErbB4 inhibitors to the same animals and compare the effects before and after. This would ensure the recording of the same population of neurons.

Thank you for your valuable comment and suggestion. We agree that tetrode recording combined with spike-sorting is not a strict "single-unit" approach, although this method is widely accepted and consistently referred to as "single-unit recording" even "single-neuron recording" (PMID:30833706). Based on your comments, we modified the description as follows: “To obtain individual neuron spikes, units were identified by principal-component analysis scan (PCA) clustering of the spikes” (Lines 210-212).

As shown in Figure 4, a standardised panel with eight tubes delivers eight odours: isoamyl acetate, 2-heptanone, phenyl acetate, benzaldehyde, dimethylbutyric acid, n-heptanoic acid, n-pentanol, and 2-pentanone. Odours were delivered via an odour delivery system (Thinkerbiotech), and charcoal-filtered air was supplied to the mouse’s nose at a constant rate of 1 L/min to eliminate airflow interference, which effectively limited the stimulation to 2s. We have added the above description to the revised "Materials and Methods" section (Lines 786-789).

Author response image 1.

The tetrodes were placed within the ventral mitral cell layer based on stereotaxic coordinates. The putative cell type is mitral/tufted (Figure 4C). Despite the presence of diverse cell types in the OB, only the spiking activity of M/TCs can be captured via tetrode recordings (Kay and Laurent, 1999; Doucette et al., 2011). We added the above description in the revised "Materials and Methods" section (Lines 764-767). As you have indicated, Figure 4C shows the impairment of odour-induced spike suppression in PV-Erbb4^-/- mice. Consistently, Figure 4D shows the decreased suppression of odour-induced spike activity in blue (quantified in Figure 4I ΔMFR, ~25%), but the increased number of suppressed neurones in the PV-Erbb4^-/- group (Figure 4L). Considering that applying ErbB4 inhibitors to the same animals has no selectivity for blocking ErbB4 signalling in PV interneurons, in vivo neural signals typically drift over time when performing pre- and post-inhibition comparisons. Therefore, we compared in vivo electrophysiological data at the population level according to previous studies (PMID: 35173173, PMID:28425486, and PMID:28817803).

(3) At a glance in the heatmap in Figure 4D, excited neurons were reduced in PV-ErbB4-/- mice, but not inhibited neurons. This was different from Figure 4L. The authors need to have a criteria or threshold to show how they categorized each population.

In Figure 4L, red denotes excited neurones, whose number decreased from 66 to 16 in the PV-Erbb4^-/- group, and blue denotes inhibitory neurones, whose number increased from 81 to 148 in the PV-Erbb4^-/- group. This trend is consistent with that observed in Figure 4D. The classification criteria for each population were defined as follows:

“To determine whether the odour evoked a significant response, we used the Wilcoxon signed-rank test to compare the spontaneous (baseline) firing rate and odour-evoked firing rate of each unit–odour pair across all trials; unit–odour pairs with P > 0.05 were defined as nonresponsive, those with P < 0.05 as responsive, and responsive pairs were further categorised as excitatory (odour-evoked firing rate > spontaneous firing rate) or inhibitory (odour-evoked firing rate < spontaneous firing rate). Details are provided in the "Off-line spike sorting and statistics of the unit data" section of Materials and Methods (Lines 810-816).

(4) Figure 4D, 4F and 4J seemed to be inconsistent. In Figure 4D before odor, there was no clear increase in the spontaneous activity in PV-ErbB4-/- mice; in Figure 4F-4G and 4J-4K, clearly, there was a high spontaneous activity in PV-ErbB4-/- mice.

Figure 4D, 4F and 4H shows odour-evoked changes in mean firing rate (ΔMFR), calculated by subtracting the baseline firing rate. By contrast, Figures 4G and 4I show raw spontaneous mean firing rate (MFR), which was elevated in PV-Erbb4^-/- mice. We have revised the description of Figure 4D in both the Results section (Line 222) and figure legends (Line 1217).

(5) What are the neurons recorded in Figure 6E-6F? If they were MCs, loss of ErbB4 in PV neurons should not alter their intrinsic electrical properties. Rather GABAergic inputs could be altered. Indeed, the authors presented a reduction of GABAergic inputs from PV neurons to MCs.

The neurones recorded in Figures 6E-6F are indeed mitral cells (MCs). The firing frequency of action potentials reflects not only the intrinsic electrical properties but also the balance between excitatory and inhibitory synaptic inputs. As demonstrated in this study, the loss of ErbB4 in PV neurones leads to reduced GABAergic synaptic inputs from PV interneurons to MCs, which is consistent with our reasoning and is a key factor contributing to the observed changes in the action potential firing of MCs

(6) Figure 8E-8H, a better experiment would be specifically expressing ErbB4 or PV neurons. In Figure 8F and Figure 8I, was it the excitability after the current injection? Why not perform the spontaneous activity recording?

According to your suggestion, we used optogenetics to activate PV interneurons and recorded inhibitory postsynaptic currents (IPSCs) in MCs. As shown in the revised Figure 9 and Supplementary Figure 1F and 1G, deleting ErbB4 of PV interneurons in the OB weakened the inhibitory modulation of MCs by PV interneurons.

Figures 8F and 8I show the excitability after current injection. We did not measure spontaneous activity, primarily because the spontaneous firing frequency varies significantly among different cells, resulting in poor stability. In contrast, current-injection-induced excitability is more stable and quantifiable, making it more suitable for assessing specific changes in MC responsiveness caused by altered PV-MC inhibitory inputs.

Reviewer #2 (Public review):

Summary:

Hu et al investigate the role of PV neurons and their expression of Erbb4 in olfactory performance through a series of behavioral tests, selective knockout experiments, and in vivo and in vitro electrophysiology. Knockout of Erbb4, either in PV cells or the whole OB, resulted in impairment of discriminating complex odors. The authors present data that inhibition is impaired in MCs, which is likely underlying the abnormal odor-evoked responses of MCs in vivo and the impaired behavioral responses.

Strengths:

Overall, a key strength of this manuscript is the breadth of experiments to test the role of PV Erbb4 expression on circuit dynamics and behavior. The behavioral experiments were clear and sufficiently powered.

Weaknesses:

The major drawback of this manuscript is the lack of depth and rigor in experiments. Some experiments are preliminary, underpowered, and not quantified. As a result, many conclusions of the manuscript are weakly supported in its current form and would require significant revisions to address these shortcomings. Major weaknesses that should be addressed are as follows:

(1) AAV-PV-Cre-GFP is not described or validated. Is this the S5E2 enhancer or something else? What is the specificity and efficacy of this approach in selectively knocking out Erbb4 in PV neurons? Reduced Erbb4 expression in the entire OB with PCR does not validate the selectivity of this approach. At a titer of 10^12, it is unlikely to be specific. Even a small amount of off-target Cre expression will knock out the gene in non-PV cells, so the authors should show whether the gene is knocked out at the single-cell level from PV and non-PC cells. Without validation of this approach, this experiment is no different than the AAV-Cre-GFP experiments.

Thank you for your comment and thoughtful suggestion. The AAV-PV-Cre vector used in this study was purchased from BrainVTA (catalogue number: PT-0275; vector name: rAAV-fPV-CRE-bGH polyA), which contains an fPV promoter to ensure the selectivity and a β-globin intron to enhance Cre expression efficiency. Importantly, the specificity and efficacy of this AAV-PV-Cre construct in selectively targeting PV-expressing neurones have been validated in multiple published studies (PMID:40196908 and PMID:31987494). According to your suggestions, we also performed immunofluorescence staining (co-staining for PV and ErbB4) and found that ErbB4 expression was specifically reduced in PV-positive interneurons in OB tissues injected with AAV-PV-Cre, while ErbB4 levels remained unchanged in PV-negative cells (the revised Figures 9B and 9C). This single-cell-level validation directly demonstrated that our AAV-PV-Cre-mediated knockout approach specifically targets ErbB4 in OB PV interneurons, thereby avoiding off-target effects in non-PV cells.

(2) Figure 1D - three mice per group is insufficient. There is no control group error (the same as Figure 9). Why is it a paired t-test when there is a control group? The authors should be comparing go/go vs. go/no-go. The methods for normalization are unclear and are likely to hide the fact that n=3 is insufficient to capture a difference without extra measures to normalize the data.

Thank you for the comment. According to your suggestions, we expanded the sample size to four mice per group and modified the statistical test from a paired t-test to an unpaired t-test in the revised Figure 1C and 1D. Since not all samples were exposed to the same blot for direct comparison, we performed normalisation of each sample relative to its internal control prior to statistical analysis to ensure the reliability of group comparisons. There was an error regarding the control group in the revised Figure 1D. Detailed information on the normalisation method and updated statistical results have been provided in the revised figure legends (Lines 1116-1118).

(3) The analysis of LFP is limited. During what period was this quantified? Are there any differences in task-related LFP changes? Also related to in vivo electrophysiology, the authors should show examples of isolated units, including their waveforms and how units were clustered and assigned to M/TCs.

Thank you for the comment. The spectral power of each frequency band was quantified using a 4-s sliding window during the pre-odour stimulation period (4–0 s before odour onset). Regarding task-related LFP changes, the initial recordings were affected by unavoidable interference, leading to unstable signals that precluded reliable analyses. Based on your suggestions, we have supplemented the revised manuscript with examples of isolated single units, including their representative waveforms, clustering results (e.g. principal component analysis plots), and unit assignment criteria (the revised Figure 4—figures supplement 1A and 1B). As previously noted, the tetrodes were placed in the mitral cell layer based on stereotaxic coordinates. Consistent with established literature (Kay and Laurent, 1999; Doucette et al., 2011), the putative cell type recorded using this approach was mitral/tufted cells (M/TCs). Detailed methodological parameters for spike sorting, unit isolation, and cell type identification are provided in the "Materials and Methods" section (Lines 764-767 and Lines 803-805).

(4) The authors use 80pA and 100pA to elicit equivalent AP spiking in MCs to determine if recurrent inhibition differs, but do not actually show that AP spiking is the same across groups. This should be quantified.

There seems to be a prominent increase in the firing of MCs in PV-Erbb4+/+ mice before odor presentation, but not in PV-Erbb4-/- mice. What is the significance of this?

Thank you for your valuable comment and suggestion. According to your suggestions, we quantified the AP firing frequency of MCs in both groups under the respective stimulation currents, and statistical analysis confirmed that there was no significant difference in the AP spiking levels between the two groups (the revised Figure 8—figure supplement 1A).

We speculate this prominent pre-odour firing increase in PV-Erbb4^+/+ mice (but not in PV-Erbb4^−/− mice) stems from interference caused by the opening of the solenoid valve during odour clearance prior to stimulation—this artefact was present in the initial recordings of control mice but was eliminated during the data acquisition for PV-Erbb4^−/− mice. This was not statistically significant. To avoid this interference, we strictly defined the analysis windows, and the spontaneous firing rate was calculated by averaging the spikes recorded during the 4–2 s period before odour stimulation, excluding the interference-prone interval immediately before odour onset. This rigorous window selection ensured the reliability of our comparisons of spontaneous and odour-evoked firing between the groups.

(5)There is a disconnect between the in vivo firing rates of MCs and ex vivo firing rates. In slice, the authors note that the spontaneous activity of MCs is elevated in the KO, but this is not observed in vivo, where conditions are physiological. Therefore, it is unclear whether the concept of signal-to-noise changes in slice (higher spontaneous, lower evoked), indeed translate to something in vivo. It would be important to know what the PV cells are doing in vivo. Perhaps they have low firing rates prior to odor onset, which may explain the lack of observed difference in baseline FRs in MCs. The authors should have this data in their tetrode recordings, which would offer insight into when inhibition is recruited.

Thank you for your valuable comments, which helped clarify the consistency between in vivo and ex vivo electrophysiological data. In ex vivo slice recordings, we measured both spontaneous and evoked excitatory firing rates of MCs, and found that the spontaneous activity of MCs was elevated in PV-Erbb4^−/− mice (Figures 6A and 6B). Importantly, this observation is consistent with our in vivo data: M/TCs in PV-Erbb4^−/− mice exhibited significantly higher spontaneous firing rates during excitatory responses to odour presentation (Figures 4F and 4G). However, for inhibitory responses during odour presentation, M/TCs in PV-Erbb4^−/− mice showed significantly lower spontaneous firing rates in vivo (Figures 4H and 4I). Thus, the overall in vivo spontaneous activity of MCs appeared unchanged owing to opposing changes in excitatory and inhibitory subpopulations, and there was no true discrepancy between the ex vivo and in vivo results. We agree that in vivo recording of PV interneuron activity would be invaluable for understanding their functional state under physiological conditions. Unfortunately, this is technically unfeasible with the current experimental setup, primarily because of the sparse distribution of PV interneurons in the OB. As a complementary alternative, we characterised the key electrophysiological properties of PV interneurons in ex vivo slices (e.g. action potential and mEPSC, the revised Figure 7—figure supplement 1), which will provide additional insights into the underlying cellular mechanisms.

(6) Since PV neurons are required for gamma oscillations, why is it that KOs have higher gamma oscillations? Is it indeed the case that PV cells have a hypofunctional phenotype in this model? Again, recording from PV cells in vivo would help make sense of this.

Thank you for your valuable comments, which were critical for clarifying the functional consequences of ErbB4 knockout in PV interneurons. The elevated gamma oscillations in the OB of PV-Erbb4^-/- mice appear to be primarily driven by the increased action potential firing frequency of PV interneurons that we observed in OB slice recordings (the revised Figure 7—figure supplement 1). A similar phenomenon has been reported in the hippocampus (PMID:24050403).

Indeed, the PV cells displayed a hypofunctional phenotype in this model. Our optogenetic experiments provided direct evidence for this (the revised Figure 9—figure supplement 1). Specifically, optogenetic activation of PV interneurons in AAV-PV-Cre-injected mice (selective ErbB4 knockout in OB PV interneurons) induced weaker inhibitory postsynaptic currents (IPSCs) in MCs than in controls (the revised Figure 9—figure supplement 1). This indicates that ErbB4 loss induces PV interneurons to be more excitable, but impairs their ability to inhibit neighbouring MCs.

We acknowledge that in vivo recordings of PV interneurons would further clarify their functional state under physiological conditions. As noted earlier, this remains technically challenging using our current setup; however, our ex vivo mechanistic data and optogenetic findings collectively support these conclusions.

(7) A clearer picture of how PV cell inhibition changes with Erbb4 KO would be achieved with optogenetically evoked IPSPs, rather than changes in mini frequency.

Thank you for the comment. Based on your suggestions, we used optogenetics to activate PV interneurons and recorded the optogenetically evoked inhibitory postsynaptic currents (oIPSCs) in MCs. As shown in the revised Figure 9 and Supplementary Figure 1F and 1G, deleting ErbB4 in PV interneurons in the OB weakened the inhibitory modulation of MCs by PV interneurons.

Reviewer #3 (Public review):

Summary:

The authors investigate the role of ErbB4 in parvalbumin (PV) interneurons within the olfactory bulb (OB) and its regulation of odor discrimination behavior in mice. They demonstrate that odor discrimination increases ErbB4 kinase activity and that the loss of ErbB4 in the OB impairs the dishabituation of odor response and discrimination of complex odors. The study also characterizes the expression of ErbB4 in the OB, showing it is enriched in PV neurons. Furthermore, the authors utilize a mouse model in which ErbB4 is knocked out in PV neurons and perform a variety of behavioral, electrophysiological, and local field potential (LFP) recording experiments to characterize alterations in olfactory bulb activity. They then use a model in which ErbB4 is specifically knocked out in PV neurons in the OB and show that this manipulation disrupts odor-related behaviors in mice.

Strengths:

The study's strengths lie in its use of a diverse range of techniques, including RNAscope, IHC, and Western blotting, to assess the presence of ErbB4 in PV neurons within the OB. Additionally, the authors employ various behavioral tests to evaluate the effects of ErbB4 manipulation in different mouse models, alongside comprehensive electrophysiological experiments and LFP recordings to examine the impact of these manipulations on OB physiology.

Weaknesses:

While the data presented in this paper are interesting, several major concerns reduce my enthusiasm for this study, as outlined below:

(1) In reviewing Figure 1C/D, there are several concerns regarding the clarity and interpretation of the data:

(a) While the Western blot for ErbB4 in other figures (Figure 1F, 2I) of the manuscript shows a clear single band, the blot presented in Figure 1C (for both p-ErbB4 and total ErbB4) shows multiple bands, which is unexpected. This discrepancy raises concerns about the consistency of the results.

(b) The data presented in Figure 1D uses only 3 mice per group, and the reported p-value of 0.0492, while technically significant, is very close to the threshold. This raises concerns about the robustness of the finding, especially given the small sample size. Additionally, the p-ErbB4 band intensity in the Go/No-Go condition in Figure 1C does not appear to show a clear increase over the Go/Go condition, which is not congruent with the bar graph in Figure 1D showing a 50% increase in p-ErbB4/ErbB4 levels.

(c) It is a standard practice in many journals to include full, uncropped Western blot images as supplementary material. This transparency helps ensure that no bands are selectively shown or omitted and increases confidence in the presented data.

Thank you for your comment and thoughtful suggestion.

(a) The multiple bands observed for both p-ErbB4 and total ErbB4 in the original Figure 1C are likely attributable to the high concentration of the primary antibody used initially (Santa Cruz Biotechnology), which has since been discontinued. In the revised manuscript, we have replaced the original antibody with an anti-ErbB4 antibody (Cat# 4795, Cell Signalling Technology), which showed a single ErbB4 band (the revised Figure 1C). To detect p-ErbB4, we first performed immunoprecipitation with a phosphorylated tyrosine antibody to enrich the total phosphorylated proteins, followed by immunoblotting for ErbB4.

(b) According to your comments, we have expanded the sample size to four mice per group and modified the statistical test to an unpaired t-test in the revised Figure 1D.

(c) We included all relevant uncropped blot images in the Supplementary Material (the revised Figure 1, 2—figure supplement 1).

(2) In Figure 2, the authors used the anti-ErbB4 antibody sc-283 from Santa Cruz to assess the expression of ErbB4 in PV neurons and the absence of its expression in PV-ErbB4 knock-out mice. However, this particular antibody has been shown to produce non-specific bands in Western blotting and also generate non-specific labeling in IHC. This non-specificity has been demonstrated in Vullhorst et al. (2009, J Neurosci), raising significant concerns about the reliability of the data generated using this antibody.

Thank you for the comment. As you have indicated, the non-specificity of the anti-ErbB4 antibody sc-283 has been reported in J Neurosci (Vullhorst et al, 2009). These antibodies have been widely used in numerous recent studies (PMID:39543371 and PMID:41297670), indicating their utility for assessing ErbB4 expression under appropriate experimental conditions (primary antibody dilution and incubation conditions). As shown in the revised Figure 2I and 2J, the absence of ErbB4 expression in PV-ErbB4 knock-out mice proved the specificity of the antibody for immunofluorescence staining. To address these potential concerns, we conducted a rigorous validation tailored to our model. In the revised manuscript, we complemented the protein-level detection with RT-PCR, which directly validated the specific knockout of Erbb4 mRNA in PV neurones (the revised Figure 2F-H). Additionally, we performed follow-up immunofluorescence staining experiments using a well-validated anti-ErbB4 antibody (19943-1-AP, Proteintech; see PMID: 37060105) to assess the expression of ErbB4 in PV neurones (the revised Figure 9B and 9C).

(3) In reviewing the statistical analysis for the series of odor discrimination tests, there could be a potential issue with the clarity of the significance testing. Although the figure legend reports the F and p values from the two-way ANOVA, it is unclear whether these values represent the main effects or the results of a post hoc test. Additionally, it is not clear whether the asterisk in the figures reflects significance from a post hoc test or from the overall ANOVA. The methods section does not explicitly state whether a post hoc test was performed to assess differences between the knockout and control groups. Given that the tests were conducted across multiple days or conditions, a post hoc test that can adjust for multiple comparisons would be necessary to accurately identify where specific differences between the groups exist.

Thank you for your comment and suggestion. For the odour discrimination tests, two-way ANOVA was followed by Tukey’s post hoc adjustment for multiple pairwise comparisons, which is explicitly stated in the revised Methods section (Lines 832-833). According to your suggestions, we have revised the figure asterisks to indicate significant post hoc differences between the groups.

(4) Throughout the manuscript, the authors use different mouse models, including ErbB4 knockout specifically in the OB (AAV-Cre-GFP), ErbB4 knockout in PV interneurons throughout the brain (PV-ErbB4-/-), and ErbB4 knockout in PV interneurons within the OB (AAV-PV-Cre-GFP). For Figures 4 and 5, the authors use the PV-ErbB4-/- model to examine odor-evoked activity and neural oscillations within the OB. Since the knockout affects PV interneurons across the entire brain, it is difficult to disentangle whether the observed changes in the OB are due to local effects or broader network alterations elsewhere in the brain.

Thank you for the comment. We first clarify the rationale for initially using the whole-brain PV-Erbb4^-/- model: this model is well-validated and widely adopted in in vivo electrophysiological studies targeting brain region-specific phenomena, such as regulating synaptic plasticity in the hippocampus (PMID:24050403) and modulating sensory coding in the cortex (PMID: 37069344).

Additionally, PV interneurons in the OB are local-projection neurones that exert stronger and more extensive inhibitory effects on M/TCs, a key feature that restricts their functional impact primarily to the local OB circuit rather than broader cross-brain network interactions, further supporting the relevance of the global model for OB-focused research.

Furthermore, local AAV-Cre injections have notable experimental limitations in our electrophysiological recordings. Stereotaxic infusion induces mechanical trauma and local inflammation in the OB, altering basal circuit activity independent of ErbB4 deletion. AAV also yields incomplete mosaic gene ablation among PV interneurons, introducing a variable knockout efficiency that complicates the interpretation of local electrophysiological readouts. In comparison, the global PV-Erbb4^-/- line mediates uniform, complete ErbB4 deletion across all PV interneurons, avoiding viral artefacts and rendering our OB electrophysiological measurements more reliable for assessing endogenous local circuit function.

(5) While the electrophysiological experiments shown in Figures 6-8 provide valuable insights into the reduced inhibition to MCs in PV-ErbB4 knockout mice, it appears that the authors did not record from PV interneurons themselves. Since PV interneurons are central to the proposed mechanism, directly recording them would provide critical information on how the ErbB4 knockout affects their intrinsic properties, synaptic inputs, and firing behavior. Without these direct recordings, the conclusions about the specific role of PV neurons in regulating MC activity remain somewhat indirect. Prior studies have established that knockout of ErbB4 in PV interneurons reduces mEPSC frequency in PV neurons (Del Pino et al., 2013).

Thank you for your valuable comment and suggestion. According to your suggestions, we have supplemented the electrophysiological experiments (the revised Figure 7—figure supplement 1) with direct recordings from PV interneurons in ex vivo slices. These recordings revealed that ErbB4 knockout alters some properties of PV interneurons (e.g. action potential frequency), reduces mEPSC frequency (consistent with Del Pino et al., 2013), and modulates their firing behaviour, directly confirming that ErbB4 deletion affects PV interneurons.

(6) In Figure 9, the authors knock out ErbB4 in PV neurons in the OB with AAV-PV-Cre-GFP and show with western blotting that ErbB4 expression is reduced in the mouse injected with AAV-PV-Cre-GFP. However, it is not clear whether ErbB4 was selectively knocked out in PV neurons without the quantification from IHC assays.

Thank you for the comment. To verify the selective knockout of ErbB4 in OB PV neurones, we performed an additional experiment: immunofluorescence co-staining for PV and ErbB4; quantitative analysis confirmed a significant reduction of ErbB4 in PV interneurons in AAV-PV-Cre–injected mice (the revised Figure 9B and 9C).

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Other concerns:

(1) No supplemental Figure 1 was shown.

We apologise for the oversight. Supplemental Figure have been included in the revised manuscript (the revised Figure 2—figure supplemental 2).

(2) In line 59, page 3, abbreviations of mitral/tufted cells (M/TCs) or "MCs" were confusing.

We have clarified the abbreviation by explicitly defining "mitral cells (MCs)", with "M/TCs" retained for mitral/tufted cells to avoid ambiguity (Line 56-59).

(3) The quality of blots in Figure 1C was less satisfactory, with lots of noisy bands. Maybe antibodies from Cell Signaling should be used.

We optimised the experiment as suggested. In the revised Figure 1C, we have used an anti-ErbB4 antibody (Cat# 4795, Cell Signalling Technology) combined with the immunoprecipitation-immunoblotting workflow.

(4) The authors should describe the serotype as well as the promoters of AAV-Cre-GFP used in the experiments.

As you suggested, we have supplemented the details of AAV-Cre-GFP in the Methods section: Viruses driven by the CAG promoter, including pAOV-CAG-EGFP-T2A-Cre (catalogue No. CN396, serotype AAV2/8) and pAOV-CAG-EGFP-2A (catalogue No. CN394, serotype AAV2/8) were purchased from Obio Technology, Shanghai, China (Lines 568-571).

(5) The authors should describe the binary mixtures of isoamyl acetate Vs. Heptanone. Simply 6/4 Vs. 4/6 is not easy to follow.

As you suggested, we have clarified the binary odour mixtures in the Results section: (60%/40% binary mixtures of these odours, "6/4" refers to 60% isoamyl acetate + 40% heptanone, and "4/6" refers to 40% isoamyl acetate + 60% heptanone, in volume) (Lines 111-113).

(6) Figure 1G, "0 limonene" AAV-Cre-GFP group did not show the error bar.

We apologise for the omission—error bars for the "0 limonene" AAV-Cre-GFP group have been added to the revised Figure 1.

(7) Figure 2A, the green signal for ErbB4 was not that strong, it was hard to visualize the expression of ErbB4 in five layers of OB.

Thank you for the comment. The revised Figure 2A shows an enhanced green signal for ErbB4, clearly indicating its distribution in all OB layers.

(8) The authors should clearly label or indicate the cell type of neurons recorded in Figure 6, Figure 7, and Figure 8A-8C.

As you suggested, we have clearly labelled the cell types in the revised Figures 6, 7, and 8A–8C: all recorded neurones are mitral cells (MCs).

(9) line 320-321, page 14, "Figures 8F and 8G, a burst of conditioning 320 stimuli (40 pulses at 100 Hz)", the authors should indicate the stimuli as a tick or scale on Figures 8F and 8I.

As per your suggestion, we have added scales and vertical ticks to the revised Figures 8F and 8I to indicate the timing of the conditioning stimuli (40 pulses at 100 Hz).

Reviewer #2 (Recommendations for the authors):

(1) Figure 1F is not quantified.

As suggested, we have supplemented the quantitative analyses of Figure 1F (see the revised Figure 1G) and the statistical details in the revised legend.

(2) Figures 2G and H are not quantified.

As suggested, we have supplemented the quantitative analyses of Figures 2G and 2H (see the revised Figures 2H and 2J), and statistical details are provided in the revised figure legends.

(3) Bar plots with standard error are not suitable for visualizing the distribution of data. Either show violin plots, histograms, or raw data points throughout.

As suggested, we included the raw data points in the revised manuscript.

(4) Figure 3 legend - (Wen et al., 2010) citations repeated.

Thank you for the comment. We have removed the repeated (Wen et al., 2010) citation in the revised manuscript.

(5) Does Ebbr4 deletion change the waveform shape of PV+ cells? This could influence spike sorting and cell classification.

Thank you for the comment. According to your recommendations, we performed additional ex vivo patch-clamp recordings of PV+ interneurons, which showed that ErbB4 knockout did not alter their action potential waveform shape (e.g. spike amplitude and half-width, the revised Figure 7—figures supplement 1D and 1E). Importantly, this knockout also did not directly affect the spike waveform shapes of M/TCs in vivo (e.g. peak-to-valley ratio and half-valley width, the revised Figure 4—figures supplement 1C-1E). Therefore, we speculate that spike sorting and cell classification were not affected.

(6) Line 327 - is this really a measure of inhibitory circuit plasticity?

Thank you for the comment. We have replaced "inhibitory circuit plasticity" with "inhibitory circuit activity" and comprehensively updated all relevant instances throughout the manuscript to ensure consistent use of the term.

(7) In the discussion: "Another interesting finding is that the changes in spontaneous firing rate differed for odor-evoked excitatory and inhibitory responses in vivo, suggesting that ErbB4 in PV interneurons dynamically regulates the spontaneous activity of M/TCs in response to different odors." The significance of this finding is not discussed and it is not clear what this result means. For example, does this relate to a "winner-take-all scenario"?

Thank you for the comment. According to your recommendations, we have discussed the significance of this finding in the revised Discussion section (Lines 438-449):

This result reveals a dual regulatory role of ErbB4 in PV interneurons in olfactory coding. For odour-evoked excitatory responses, PV interneurons (via ErbB4) constrain M/TC spontaneous firing at a low baseline, minimising background noise and enhancing the SNR. In contrast, the inhibitory responses maintain a higher spontaneous firing baseline to accentuate odour-evoked suppression. This regulation amplifies the "excitation of excited M/TCs and inhibition of inhibited M/TCs," a hallmark of the OB’s "winner-take-all" mechanism that sharpens odour identity coding, which is critical for odour discrimination. These data confirm that ErbB4 in PV interneurons does not exert a global modulation, but dynamically tailors M/TC spontaneous activity to odour-evoked responses, thereby optimising the ability of the OB to resolve olfactory cues.

(8) The discussion on linking AD with odor discrimination impairments, and the potential role of Erbb4 was interesting. However the discussion of "fluctuation status" is not defined. What does this refer to?

According to the comment, we have added the following wording to the Discussion section (Lines 478-497):

As referenced by Iemi et al. (2022), "fluctuation status" herein refers to the spontaneous, endogenous fluctuations in neural activity that occur in the brain prior to sensory stimulation (i.e. pre-stimulus) or in the absence of external sensory input. External cues do not drive this state but arise from intrinsic neural dynamics, and it is primarily characterised by changes in the power of neural oscillations across multiple frequency bands (e.g., alpha, beta, and their neighbouring frequencies, collectively termed "alpha+ band" in Iemi et al. (2022)). Critically, this fluctuation status is distinct from stimulus-evoked neural responses; it reflects a pre-existing "baseline" state of the neural system that modulates subsequent sensory processing and behaviour. Consistent with Iemi et al. (2022), previous work has established that such ongoing neural fluctuations strongly predict behavioural variability in both health (e.g. slower reaction times and reduced perceptual accuracy) and disease (e.g. age-related cognitive decline and schizophrenia) by regulating neuronal excitability. In our study, we extended this concept to olfactory function, focusing on "broad ongoing fluctuations" (encompassing multiple frequency bands) and linking their dynamics to OB output. Specifically, these fluctuations influence the ability of the OB to support rapid recovery from olfactory adaptation and to mediate complex odour discrimination, thereby uncovering a novel role for endogenous neural fluctuation status in olfactory circuit function.

Reviewer #3 (Recommendations for the authors):

(1) For Figure 1C, I strongly recommend that the authors include full, uncropped Western blot images for both p-ErbB4 and total ErbB4 as supplementary material to improve the transparency and credibility of the findings.

As you suggested, we have supplemented the full, uncropped western blot images for both phosphorylated ErbB4 (p-ErbB4) and total ErbB4, as the revised Figures 1 and 2 and Supplementary Figure 1.

(2) The authors should either validate the specificity of the sc-283 antibody in their experimental conditions or consider using a more specific antibody for ErbB4 to ensure the reliability of their data.

Thank you for your comments. As shown in the revised Figure 2I and 2J, the absence of ErbB4 expression in PV-ErbB4 knock-out mice proved the specificity of the antibody for immunofluorescence staining. To address these potential concerns, we conducted a rigorous validation tailored to our model. In the revised manuscript, we complemented the protein-level detection with RT-PCR, which directly validated the specific knockout of Erbb4 mRNA in PV neurones (the revised Figure 2F-H). Additionally, we performed follow-up immunofluorescence staining experiments using a well-validated anti-ErbB4 antibody (19943-1-AP, Proteintech; see PMID: 37060105) to assess the expression of ErbB4 in PV neurones (the revised Figure 9B and 9C).

(3) I recommend that the authors clarify in the methods section whether a post hoc test with multiple comparison adjustments (e.g., Tukey's HSD or Bonferroni) was conducted for the behavioral tests presented throughout the study. If no such test was performed, it would be beneficial to include one to ensure accurate identification of significant differences between groups across the different experimental days or conditions.

For the behavioural tests, we confirmed that two-way ANOVA was followed by Tukey’s post hoc adjustment for multiple pairwise comparisons, which has been clarified in the revised Methods section (Lines 832-833). Additionally, we have revised the asterisks in the relevant figures to clearly indicate significant post-hoc differences between the groups.

(4) The rationale for using a whole-brain PV-specific knockout model (PV-ErbB4-/-) to assess OB-specific phenomena in Figures 4 and 5 should be clarified. I also recommend the authors provide or conduct additional experiments with OB-specific knockouts to determine whether the effects observed in Figures 4 and 5 are directly attributable to the loss of ErbB4 in PV interneurons within the OB.

To clarify the rationale, the whole-brain PV-Erbb4^−/− mouse line was initially used to characterise the overall functional relevance of ErbB4 in PV interneurons for OB activity and odour-evoked responses. We clarified this rationale in the revised Results section (Lines 210-211 and Lines 250-251).

As elaborated above, AAV-mediated OB-specific knockouts suffer from injection-induced tissue injury, incomplete mosaic recombination, and uncertainties in local electrophysiological recordings, which would compromise data reliability. Moreover, OB PV interneurons act predominantly as local circuit neurones; therefore, the phenotypes observed in Figures 4 and 5 most plausibly arise from intrinsic ErbB4 deletion within bulbar PV cells rather than from long-range brain-wide network changes. For these reasons, we consider our current global PV-Erbb4^−/− model appropriate to interpret the local OB circuit effects here, and we outline OB-targeted knockout manipulations as a direction for future independent follow-up research.

(5) To strengthen their conclusions drawn from Figures 6-8, the authors should perform direct recordings from PV interneurons in the OB to assess how the ErbB4 knockout affects their intrinsic properties, synaptic inputs, and firing behavior. This would provide more direct evidence of PV interneuron dysfunction in the knockout mice.

According to your recommendations, we have performed direct recordings of OB PV interneurons using ex vivo patch-clamp electrophysiology (the revised Figure 7—figures supplementary 1). These recordings reveal that ErbB4 knockout affects the intrinsic properties of PV interneurons, synaptic inputs (reduced mEPSC frequency and amplitude, the revised Figure 7—figures supplementary 1A and 1B), and firing behaviour (increased action potential frequency, the revised Figure 7—figures supplementary 1C), directly confirming that ErbB4 deletion affects PV interneurons themselves. These results provide direct evidence of PV interneuron dysfunction in the knockout mice.

(6) I recommend that the authors assess molecular markers of PV interneuron inhibition, such as GAD67, in PVIs. This would help link the ErbB4 knockout to impaired PV interneuron function and its downstream effects on MC inhibition.

Thank you for the comment. We attempted to detect GAD67 in OB PVIs using multiple commercial antibodies but encountered consistent technical limitations (e.g. weak signal specificity and high background) that precluded reliable quantitative analysis.

To address this gap, we used optogenetics to activate PV interneurons and recorded optogenetically evoked inhibitory postsynaptic currents (oIPSCs) in MCs. As shown in the revised Figure 9 and Supplementary Figure 1F and 1G, deleting ErbB4 of PV interneurons in the OB weakened the inhibitory modulation of MCs by PV interneurons. This result provides direct electrophysiological evidence linking ErbB4 knockout to impaired PV interneuronal function and its downstream effects on MC inhibition.

(7) For Figure 9, I recommend that the authors perform IHC to show that ErbB4 is selectively knocked out in PV neurons but not in non-PV neurons within the OB, as the Western blot analysis alone does not provide this specificity.

According to your recommendations, we performed an additional key experiment: immunofluorescence co-staining for PV and ErbB4, with quantitative analysis confirming a significant reduction in ErbB4 in OB PV interneurons, but not in non-PV cells, in AAV-PV-Cre–injected mice (the revised Figure 9B and 9C).

eLife Assessment

This work presents valuable new data on the role of D-Serine and how it competes with its stereoisomer L-Serine to influence metabolism. The work presents a variety of convincing experimental data combined with simulated results to investigate the mechanisms focused on one-carbon metabolism, which is relevant for several research fields. However, some claims are only partially supported by data, and critical areas comparing L- vs D-Serine and further mechanistic studies are required. Furthermore, while the work has potential for various fields, the work has only been studied in a neural context.

Reviewer #1 (Public review):

[Editor's note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The inclusion of a summary schematic in this revision significantly improves the manuscript. We have also modified the last sentence of the original editorial assessment.]

Summary:

The authors demonstrate the stereoselective role of D-serine in 1C metabolism showing that D-serine competes with L-serine and inhibits mitochondrial L-serine transport. They observe expression of 1C metabolites in their metabolomics approach in primary cortical neurons treated with L-serine, D-serine and mixture of both. Their conclusions are based on the reduction in levels of glycine, polyamines and their intermediates and formate. Single cell RNA sequencing of N2a cells showed that cells treated with D-serine enhanced expression of genes associated with mitochondrial functions such as respiratory chain complex assembly and mitochondrial functions with downregulation of genes related to amino acid transport, cellular growth and neuron projection extension. Their work demonstrates that D-serine inhibits tumor cell proliferation and induces apoptosis in neural progenitor cells highlighting the importance of D-serine in neurodevelopment.

Strengths:

D-amino acids do not merely function as ligands at receptors but have underlying roles in signaling and metabolism. These roles are just beginning to be uncovered. The authors elucidate the metabolic role of D-serine in the context of neuronal maturation by its suppression of mitochondrial L-serine availability for SHMT2 and 1C flux. This is the strength of the manuscript. The implications for the metabolic role of D-serine in neurons is a highlight and underlines its roles in neuronal metabolism.

Comments on previous version.

This reviewer is satisfied by the effort made by the authors based on the prior comments raised.

Reviewer #2 (Public review):

Summary:

This study by Suzuki et al. reports an interesting stereo-selective role of D-serine in regulating one-carbon metabolism during neurodevelopment to adapt the functional transition, probably through the competition with mitochondrial transport of L-serine. The authors provide a multi-layered set of evidence, including metabolomics, enzyme assays, mitochondrial transport competition and functional assays in immature/neural progenitor cells, to build up a conceptual integration of D-serine as both a neurotransmitter and a metabolic regulator in central neural system, which raises a broad potential interest to the neuroscience and metabolism communities.

Strengths:

This work provides a conceptual advance that D-serine is not only serves as a traditional neurotransmitter in central neural system but also critically contributes to metabolic regulation of neural cells. The authors performed solid metabolomic assays to validate the suppressive effect of D-serine on one-carbon metabolic pathway, providing some evidence that D-serine competitively inhibits mitochondrial serine transport, but not directly impairs SHMT2 enzymatic activity. All these data indicate a critical role of D-serine synthesis during neural maturation and suggest a potential translational strategy for targeting serine metabolism in neural tumors.

Comments on previous version.

My previous concerns have been appropriately addressed or discussed in this revised version of manuscript. I have to say that, at this stage, I have no further questions.

Reviewer #3 (Public review):

Summary:

This manuscript presents a comprehensive and well-executed investigation into the metabolic role of D-serine in the central nervous system. The authors provide solid evidence that D-serine competitively inhibits mitochondrial L-serine transport, thereby impairing one-carbon metabolism. This stereoselective mechanism reduces glycine and formate production, suppresses cellular proliferation, and induces apoptosis in immature neural cells and glioblastoma stem cells. Developmental analyses further reveal a physiological enantiomeric shift in serine metabolism during neurogenesis, aligning with the transition from proliferation to maturation. Overall, the study bridges developmental neurobiology, cancer metabolism, and amino acid transport, uncovering a previously unrecognized metabolic function of D-serine beyond its role in neurotransmission.

Strengths:

(1) The discovery that D-serine inhibits one-carbon metabolism by competing for mitochondrial L-serine transport-rather than through enzymatic inhibition or receptor-mediated signaling-represents a significant and previously underappreciated mechanism. This finding has broad implications for understanding metabolic regulation during neurodevelopment and offers potential relevance for targeting metabolic vulnerabilities in cancer.

(2) The authors integrate metabolomics, mitochondrial transport assays, molecular dynamics simulations, genetic and pharmacologic perturbations, transcriptomics, and both in vitro and ex vivo models. The breadth of experimental approaches, combined with the coherence of the findings across systems, provides strong support for the central conclusions and enhances the overall impact of the study.

(3) The temporal shift in D-/L-serine levels during neurodevelopment is elegantly linked to the transition from proliferative to mature neuronal states. The selective vulnerability of neural progenitors and tumor cells-contrasted with the resistance of mature neurons-highlights a biologically meaningful and potentially targetable metabolic distinction.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors demonstrate the stereoselective role of D-serine in 1C metabolism showing that D-serine competes with L-serine and inhibits mitochondrial L-serine transport. They observe expression of 1C metabolites in their metabolomics approach in primary cortical neurons treated with L-serine, D-serine and mixture of both. Their conclusions are based on the reduction in levels of glycine, polyamines and their intermediates and formate. Single cell RNA sequencing of N2a cells showed that cells treated with D-serine enhanced expression of genes associated with mitochondrial functions such as respiratory chain complex assembly and mitochondrial functions with downregulation of genes related to amino acid transport, cellular growth and neuron projection extension. Their work demonstrates that D-serine inhibits tumor cell proliferation and induces apoptosis in neural progenitor cells highlighting the importance of D-serine in neurodevelopment.

Strengths:

D-amino acids do not merely function as ligands at receptors but have underlying roles in signaling and metabolism. These roles are just beginning to be uncovered. The authors elucidate the metabolic role of D-serine in the context of neuronal maturation by its suppression of mitochondrial L-serine availability for SHMT2 and 1C flux. This is the strength of the manuscript. The implications for the metabolic role of D-serine in neurons is a highlight and underlines its roles in neuronal metabolism.

Thank you for the positive assessment.

Weaknesses:

These are some minor issues that come up on critical assessment of the manuscript and is only intended to strengthen the manuscript. The comments below are based on the revisions made by the authors including the justification of their approach and rebuttal.

(1) Kinetic assessment of D-serine versus L-serine: The authors have made reference to prior work by Miyamoto et al. and justify their rationale. This is acceptable.

(2) Molecular Dynamics simulations while a good first step in modeling interactions at the active site, relies on force fields. The authors state that any elaborate study into longer simulations is beyond the scope and their simulations data are supported by other experimental work. This is justified.

(3) The use of N2a cell line is also justified to reflect the proliferative nature of immature neurons.

(4) With regards to caspase 3 comment, the whole blot is convincing and shows cleaved caspase-3 band at approx. 15 kDa.

(5) Scale Bars are clearly visible and Fig S6 which was earlier S5 is legible. If possible, the authors can include an magnified inset in the merged image to show the clear activation of caspase-3.

(6) Issue of phosphatidyl serine standard in LC-MS is justified by the use of L-serine standard due to lack of availability.

(7) The authors mention about enantiomeric shift of serine metabolism during neural development which appears to be a discussion of prior published data from Hubbard et al 2013, Burk et al 2020, and Bella et al 2021 in Supplementary Figure panels 8 A-E.

The authors justify by citing references to the work which may be acceptable and also the current norms of publication. This reviewer felt contrary to the fact, however it is left to the editors to make a decision on this.

(8) The discussion section has been substantially revised and now reads well.

(9) The relevant references have been cited. In doing so, the work integrates and elucidates a mechanistic and functional role of D-serine in neurons.

(10) Figure S7A in the revised manuscript shows the specificity of D-serine in the cleaved caspase-3 assay which is informative.

Comments on revised version.

This reviewer is satisfied by the effort made by the authors based on the prior comments raised.

Thank you for understanding our explanation.

Reviewer #2 (Public review):

Summary:

This study by Suzuki et al. reports an interesting stereo-selective role of D-serine in regulating one-carbon metabolism during neurodevelopment to adapt the functional transition, probably through the competition with mitochondrial transport of L-serine. The authors provide a multi-layered set of evidence, including metabolomics, enzyme assays, mitochondrial transport competition and functional assays in immature/neural progenitor cells, to build up a conceptual integration of D-serine as both a neurotransmitter and a metabolic regulator in central neural system, which raises a broad potential interest to the neuroscience and metabolism communities.

Strengths:

This work provides a conceptual advance that D-serine is not only serves as a traditional neurotransmitter in central neural system but also critically contributes to metabolic regulation of neural cells. The authors performed solid metabolomic assays to validate the suppressive effect of D-serine on one-carbon metabolic pathway, providing some evidence that D-serine competitively inhibits mitochondrial serine transport, but not directly impairs SHMT2 enzymatic activity. All these data indicate a critical role of D-serine synthesis during neural maturation and suggest a potential translational strategy for targeting serine metabolism in neural tumors.

Comments on revised version.

My previous concerns have been appropriately addressed or discussed in this revised version of manuscript. I have to say that, at this stage, I have no further questions.

Thank you for understanding the value of our study.

Reviewer #3 (Public review):

Summary:

This manuscript presents a comprehensive and well-executed investigation into the metabolic role of D-serine in the central nervous system. The authors provide solid evidence that D-serine competitively inhibits mitochondrial L-serine transport, thereby impairing one-carbon metabolism. This stereoselective mechanism reduces glycine and formate production, suppresses cellular proliferation, and induces apoptosis in immature neural cells and glioblastoma stem cells. Developmental analyses further reveal a physiological enantiomeric shift in serine metabolism during neurogenesis, aligning with the transition from proliferation to maturation. Overall, the study bridges developmental neurobiology, cancer metabolism, and amino acid transport, uncovering a previously unrecognized metabolic function of D-serine beyond its role in neurotransmission.

Strengths:

(1) The discovery that D-serine inhibits one-carbon metabolism by competing for mitochondrial L-serine transport-rather than through enzymatic inhibition or receptor-mediated signaling-represents a significant and previously underappreciated mechanism. This finding has broad implications for understanding metabolic regulation during neurodevelopment and offers potential relevance for targeting metabolic vulnerabilities in cancer.

(2) The authors integrate metabolomics, mitochondrial transport assays, molecular dynamics simulations, genetic and pharmacologic perturbations, transcriptomics, and both in vitro and ex vivo models. The breadth of experimental approaches, combined with the coherence of the findings across systems, provides strong support for the central conclusions and enhances the overall impact of the study.

(3) The temporal shift in D-/L-serine levels during neurodevelopment is elegantly linked to the transition from proliferative to mature neuronal states. The selective vulnerability of neural progenitors and tumor cells-contrasted with the resistance of mature neurons-highlights a biologically meaningful and potentially targetable metabolic distinction.

Thank you for recognizing the importance of our study.

Weaknesses:

(1) While the authors attribute D-serine's metabolic effects to competition with mitochondrial L-serine transport, the specific identity of the transporter(s) mediating this process remains undefined. This represents a meaningful mechanistic gap, as the central conclusion depends on D-serine limiting mitochondrial L-serine availability to inhibit one-carbon metabolism.

Thank you for the comments. We acknowledge that defining the specific identity of L-serine transporters that are interfered by D-serine requires further studies.

(2) The effective concentrations of D-serine used in vitro (IC₅₀ ≈ 1-2 mM) exceed typical brain levels (~0.3 mM). While the authors acknowledge this, a more focused discussion on whether higher local D-serine concentrations could arise in specific microenvironments-such as synaptic compartments, tumor niches, or pathological states-would help contextualize the in vitro findings and strengthen their physiological relevance. For example, disruptions in D-serine clearance or altered expression of serine racemase and transporters in disease contexts could lead to localized accumulation. Moreover, differences between extracellular and intracellular D-serine pools-and the mechanisms governing their regulation-may further influence its metabolic impact in vivo.

Thank you for this suggestion. However, since we have discussed these points in lines 479-492, we chose not add any additional sentences in order to avoid losing the focus of this paper with speculation.

(3) While the manuscript focuses on neural stem/progenitor cells and neural tumors, it remains unclear whether the anti-proliferative effects of D-serine are specific to neural lineages or extend to other highly proliferative non-neural cell types. A brief discussion addressing this point would help clarify the scope of D-serine's metabolic impact and whether its mechanism of action reflects a unique vulnerability in neural cells or a more general feature of proliferative metabolism. This distinction is particularly relevant for assessing the broader therapeutic potential of targeting mitochondrial L-serine transport.

The reviewer raised an interesting point. While we speculate that other cancers requiring L-serine for proliferation may also be affected by D-serine, in this study we focused on central nervous system tumors because D-serine is produced in the central nervous system. Since cancer cells may have different metabolic profiles regardless of their tissue of origin, it would be meaningful to compare the expression profiles of metabolism-related genes with their sensitivity to D-serine.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

A minor suggestion to the authors is to include a summary schematic to clarify the main finding of the paper.

Since there is a lot of data in the manuscript ranging from MD simulations and metabolomics and other data sets, an overall summary schematic will help the readers with a take home message.

Thank you for the great suggestion. We added a summary schematic in Figure 4 -figure supplement 10.

Reviewer #2 (Recommendations for the authors):

The authors have addressed my previous concerns and further included appropriate discussions on their conclusions. I have no further questions.

eLife Assessment

This important study identifies two pairs of dopaminergic neurons (DA-WED) in Drosophila that coordinate cardiac deceleration and locomotor responses to a mechanical threat. The conclusions are supported by convincing evidence from a combination of optogenetic, physiological, and behavioral experiments. The work provides invaluable insights into the neural coordination of behavioral and cardiac responses to threat and will be broad interest to neuroscientists, physiologists, and researchers studying brain-body interactions.

Reviewer #1 (Public review):

Summary:

This study by Tsuji et al. explores a mechanical threat model in Drosophila using air puffs as a stimulus. The authors first establish the paradigm and show that air puffs induce cardiac deceleration along with increased locomotion. They then identify dopamine as a key regulator of this response and go on to map the underlying circuit. In doing so, they pinpoint two pairs of DA-WED neurons as critical players. They carefully used intersectional strategies to achieve relatively clean labeling of these neurons, which helps ensure that the observed effects can be attributed specifically to DA-WED neurons. They further show that DA-WED neurons are both required and sufficient to drive cardiac deceleration, and that their activity increases in response to air puff stimulation. These neurons also contribute to the locomotor response. Directly inducing cardiac deceleration via optogenetic manipulation of cardiomyocytes also increases locomotion, suggesting a link between cardiac state and behavioral output.

Strengths:

Overall, the experiments are thoughtfully designed, well controlled, and clearly presented. The figures are easy to follow, and the conclusions are generally well supported by the data. The manuscript is also clearly written, with a discussion that acknowledges potential caveats and outlines future directions. The genetic tools, behavioral paradigm, heart rate measurement approaches, and stimulation methods introduced here will be valuable resources for the community. few minor comments:

Comments on revised version:

I thank the authors for responding to my questions, they have thoroughly addressed my comments. I am therefore supportive of publication of the manuscript.

Reviewer #2 (Public review):

Summary:

The authors study cardiac deceleration during threat responses in Drosophila. Particularly, it focuses on identifying the neuronal control of this deceleration. Using behavioral and cardiac tracking and analysis, genetics and calcium imaging, they identify two pairs of dopaminergic neurons involved in cardiac deceleration during air puff responses.

Strengths:

The study is overall well done and the paper clearly written. Particularly the work on identifying the two pairs of dopaminergic neurons involved in cardiac deceleration using a series of drivers and generating new ones is rigorous and extensive. Finally, the authors manipulate heartbeat to investigate how it influences threat responses. Although the evidence that heartbeat deceleration-induced increase in locomotion remains partial, the study reveals important insight in the relationship between locomotion, cardiac activity and neuromodulatory system during threat responses.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

This study by Tsuji et al. explores a mechanical threat model in Drosophila using air puffs as a stimulus. The authors first establish the paradigm and show that air puffs induce cardiac deceleration along with increased locomotion. They then identify dopamine as a key regulator of this response and go on to map the underlying circuit. In doing so, they pinpoint two pairs of DA-WED neurons as critical players. They carefully used intersectional strategies to achieve relatively clean labeling of these neurons, which helps ensure that the observed effects can be attributed specifically to DA-WED neurons. They further show that DA-WED neurons are both required and sufficient to drive cardiac deceleration, and that their activity increases in response to air puff stimulation. These neurons also contribute to the locomotor response. Directly inducing cardiac deceleration via optogenetic manipulation of cardiomyocytes also increases locomotion, suggesting a link between cardiac state and behavioral output.

Strengths:

Overall, the experiments are thoughtfully designed, well-controlled, and clearly presented. The figures are easy to follow, and the conclusions are generally well supported by the data. The manuscript is also clearly written, with a discussion that acknowledges potential caveats and outlines future directions. The genetic tools, behavioral paradigm, heart rate measurement approaches, and stimulation methods introduced here will be valuable resources for the community.

Weaknesses:

A few minor points to add to the clarity of the manuscript:

(1) The DA-WED driver (R48A08-AD ∩ VT008692-DBD ∩ TH-FLP) appears quite clean in the brain. However, since the study focuses on cardiac function and locomotion, it would be helpful to check expression in cardiomyocytes and the ventral nerve cord. This would help rule out any off-target expression that might contribute to the phenotypes and further support the idea of a descending pathway from brain dopaminergic neurons.

We thank reviewer#1 for this thoughtful comment. We have indeed checked the expression pattern of the DA-WED driver (R48A08-AD ∩ VT008692-DBD ∩ TH-FLP) throughout the body, and failed to detect noticeable expression outside the brain (Supplementary Fig. 6d in the original manuscript). As for the heart, the whole tissue was cleared and confocal-imaged from the dorsal side. As cardiomyocytes are positioned directly beneath the dorsal cuticle of the abdomen, the expression of a driver can be best assessed this way. In contrast to the lack of apparent expression in cardiomyocytes, we could see clear expression in DA-WED neurons in the brain in the same sample (Supplementary Fig. 6d). In this experiment, however, potential expression in VNC could have been missed, as VNC is positioned on the ventral side and thus further away from the objective lens. We thus conducted an additional experiment in which we dissected out the brain + VNC and checked the driver’s expression. Our confocal imaging results revealed no expression in VNC. We have included this data in our updated manuscript as Supplementary Fig. 6e.

(2) Since DA-WED>Kir2.1 abolishes the puff-induced locomotor response (Figure 5b), suggesting that DA-WED neurons are directly involved in mediating locomotion. In the model (Figure 5L), it might make more sense for the pathway from mechanical threat to locomotion to pass through DA-WED neurons. The authors could consider adjusting the schematic if they agree.

As the reviewer pointed out, Kir2.1-mediated suppression of DA-WED neurons almost completely abolished puff-induced locomotion. In our original schematic, we included a pathway from mechanical threat to locomotion that bypassed DA-WED, reflecting the possibility that locomotion could be driven by multiple pathways. We agree, however, that this possibility was not directly supported by our data and that the near-complete suppression of puff-induced locomotion upon DA-WED silencing more strongly supports a model in which the predominant flow of information from puff detection to locomotion passes through DA-WED neurons. We have therefore removed the direct arrow from mechanical threat to locomotion and added a dashed arrow from DA-WED neurons to locomotion, representing a pathway that may promote locomotion independently of DA-WED-mediated cardiac deceleration (Fig. 5m; Supplementary Fig. 9i).

We have updated the legend of Fig. 5m as follows:

“Model. DA-WED neurons promote locomotion, through and/or bypassing interoception of the heartbeat (dashed arrows).”

We also updated the legend of Supplementary Fig. 9i as follows:

“Model. Air puff promotes locomotion while suppressing grooming. Mechanistically, air puff activates DA-WED neurons. DA-WED neurons drive not only cardiac deceleration but locomotion, potentially through a feedback from cardiac deceleration (dashed arrows). Such a DA-WED-based mechanism fails to influence grooming.”

(3) In line 408, Figure 5K should be 5L as it's a discussion of the model.

We thank reviewer#1 for pointing this out. Accordingly, we have made the following corrections:

L436-439: “cardiac deceleration is accompanied by increased locomotion (Fig. 5e-l). Our data points to a scenario in which DA-WED neurons induce cardiac deceleration upon mechanical threat, which may in turn influence threat-induced locomotion (Fig. 5m).”

(4) In Figure 5j, the x-axis is missing time labels. Even if it matches Figure 5h, adding labels would make it easier to interpret at a glance.

We agree with reviewer#1 and added the x labels to Fig. 5j, and also to the second panels of h, j, and l.

(5) In line 312, it would be helpful to briefly explain why a 28 ms light pulse was used, compared to other pulse durations elsewhere in the paper.

Our system synchronized behavior, heartbeat, and optogenetics at 70 Hz (14-ms intervals). Because mechanical threat transiently elongated beat-to-beat intervals typically by 20 - 23ms, we chose 28-ms (2-frame) stimulation to best mimic this effect, over 14 (1-frame) or 42 (3-frame) ms (Fig. 1; Supplementary Fig. 2b, e, f). We have added this explanation to our manuscript, as follows:

L326 - 328: “To approximate the 20–23 ms beat-to-beat prolongation evoked by mechanical threat (Fig. 1; Supplementary Fig. 2e, f), we used 28-ms (2-frame) green light pulses.”

(6) The cardiac deceleration seems to recover quickly after the air puff ends, whereas the locomotor response persists longer (around 10-15 seconds; see Figure 1 and Figure 5). This difference might suggest that DA-WED neurons influence locomotion through an additional or partially independent pathway, beyond their role in cardiac regulation. It could be worth briefly discussing this possibility.

We agree that the time course of puff-induced locomotion (Fig. 1f, Fig. 5a) is more persistent than DA-WED activation (Fig. 4f) or cardiac deceleration (Fig. 1d, Fig. 4b). Also, silencing DA-WED neurons partially reduced puff-induced cardiac deceleration (Fig. 4c) but appears to have more strongly abolished puff-induced locomotion (Fig. 5a, b). Given that our interoception model is only partially tested, we agree that it is better to mention the possibility that DA-WED neurons might influence locomotion through an additional or partially independent pathway beyond their role in cardiac regulation.

Our “Limitation of this study” section in our original manuscript touched on this issue, as follows (L430-432):

“Although cardiac deceleration can promote locomotion, we have not directly tested whether it is required for DA-WED–evoked locomotion, leaving open the possibility of parallel downstream pathways.”

But to be more explicit, in our updated manuscript, we also included the following discussion:

L439-441: “We note, however, that our data does not exclude the possibility that DA-WED neurons might influence locomotion at least partially independent of its role in cardiac deceleration”

Reviewer #2 (Public review):

Summary:

The authors study cardiac deceleration during threat responses in Drosophila. Particularly, it focuses on identifying the neuronal control of this deceleration. Using behavioral and cardiac tracking and analysis, genetics, and calcium imaging, they identify two pairs of dopaminergic neurons involved in cardiac deceleration during air puff responses

Strengths:

The study is overall well done, and the paper is clearly written. Particularly, the work on identifying the two pairs of dopaminergic neurons involved in cardiac deceleration using a series of drivers and generating new ones is rigorous and extensive. Finally, the authors manipulate the heartbeat to investigate how it influences threat responses

Weaknesses:

There are, however, several points that need to be clarified, as some claims are not entirely supported by evidence.

The authors, for example, claim that dopaminergic neurons are responsible for cardiac deceleration (during the air puff, lines 182-3, page 9). However, based on the work in this study, it seems that other neurons could be involved in this control as well. In addition to dopaminergic neurons, the authors test serotonergic and octopaminergic neurons, which, based on silencing experiments, also show an implication in heart-beat deceleration. Furthermore, because they find that dopaminergic neurons are the only ones that, upon thermogenetic activation, lead to lower heart beat frequency, they conclude that the dopaminergic neurons are responsible for air -puff induced cardiac deceleration.

However, these activation experiments are done in a different context than the air puff experiments (at a higher temperature, which could have an effect on the heartbeat changes upon activation of different neuron groups), and because silencing of other monoaminergic neuron types during the air puff also resulted in less cardiac deceleration, one cannot exclude the implication of octopaminergic or serotonergic neurons in air-puff-induced deceleration.

Activation experiments without high temperatures (using, for example, optogenetics) and/or in the presence of the air puff would be important to determine that the dopaminergic neurons are the main type of monoaminergic neurons involved in air-puff-induced cardiac deceleration. Otherwise, the related claims should be rephrased in a way that clearly doesn't exclude a possible implication of other monoaminergic neurons.

We thank reviewer#2 for raising an important point. Our intention was not to claim that dopaminergic system is solely responsible for puff-induced cardiac deceleration. What we meant to convey in our original manuscript was that, after conducting experiments shown in Fig. 2, dopaminergic system was the one that made most sense to focus on, as the other system failed to show consistent results across loss- and gain-of-function experiments. We acknowledge, as reviewer #2 rightly pointed out, that consistency across loss- and gain-of-function experiments is not a logical necessity for given neurons to be causally contributing to a brain function, especially in this case where loss- and gain- of-function experiments were conducted at different temperatures. We have revised our manuscript to more precisely express our view outlined above, as follows:

L194-196: “Together, our data demonstrates that, while not excluding the contributions of serotonergic or octopaminergic systems, dopamine system is responsible for puff-induced cardiac deceleration.”

Regarding the interactions between the cardiac deceleration and locomotion, the authors propose, based on the results, that the optogenetic cardiac deceleration is sufficient to induce an increase in locomotion, and that it is the decrease in heartbeat that would be responsible via interoceptive pathways to trigger an increase in locomotion. In the model they propose, the DA-WED neurons would induce a decrease in heartbeat that, in turn, would trigger an increase in locomotion. There is not enough proof that cardiac deceleration is the one that triggers an increase in locomotion during air puff responses. As the authors themselves state, the experiments that would demonstrate this would involve preventing cardiac deceleration while optogenetically activating DA-WED. It can therefore not be excluded that the DA-WED neurons trigger an increase in locomotion that is possibly modulated by the cardiac activity. Both alternatives should be considered (models in Figures 4 and 5).

We thank the reviewer for this insightful comment, which is also consistent with reviewer#1’s view. We agree that it is better to mention the possibility that DA-WED neurons might influence locomotion through an additional or partially independent pathway beyond their role in cardiac regulation, which our data does not exclude. In our updated manuscript, we included the discussion as below:

L433-441: “Our results provide partial support for such a model. We show that (1) mechanical threat induces cardiac deceleration and locomotion (Fig. 1), (2) DA-WED neurons in the brain are responsible for threat-induced cardiac deceleration and locomotion (Fig. 4; Fig. 5a-d), (3) optogenetic cardiac deceleration is accompanied by increased locomotion (Fig. 5e-l). Our data points to a scenario in which DA-WED neurons induce cardiac deceleration upon mechanical threat, which may in turn influence threat-induced locomotion (Fig. 5m). We note, however, that our data does not exclude the possibility that DA-WED neurons might influence locomotion at least partially independent of its role in cardiac deceleration.”

Reviewer #3 (Public review):

Summary:

In this elegant study, Tsuji et al. identify a relationship in Drosophila between cardiodynamics and threatening stimuli where mild air puffs elicit a brief bradycardia that coincides with locomotion increases. They then take advantage of the arsenal of genetic tools available in the fruit fly to reveal the indispensability of dopamine, through the action of Dop1R2, in this phenomenon. Further, they pinpoint the source of this dopamine to two specific pairs of neurons - DA-WED that are threat-activated. They then test and find a potential role for cardiac interoception from the heart in linking behavior and cardiodynamics.

Strengths:

This is an interesting and timely story that brings together the tools of fruit fly systems neuroscience and links it with physiology. The experiments are well done and tell a very nice story. In particular, the primary message of the paper - that the authors have identified specific dopaminergic neurons that regulate cardiac activity - is sound.

Weaknesses:

There are no important problems with the scientific approach. Rather, there are some interpretive changes I would consider.

(1) The changes in heart rate are small (10% or so), and, as far as I can tell, are evident for a beat or two. So the data may be better interpreted not as a change in rate but as a lengthening of diastole for a beat or two. That may seem a petty difference, but it might point to particular stretch-activated systems or changes in blood flow as the determinant.

Heart rate must be averaged over time, and so might be blurring the effects. It may be useful to produce figures centered on beat count and duration rather than time. Because the effect may even be just on a single beat, we suggest the authors try plotting the average beat duration for each beat that follows the air puff. If it's really just the first beat, using a quantification of the change of this duration relative to the average that precedes the puff may produce more striking figures.

We thank reviewer#3 for this insightful suggestion, which also relates to the reviewer’s next comment. We tried plotting the average beat duration for each beat that follows the air puff, relative to beats preceding the puff. The reviewer’s prediction turned out to be partially true – our plot of beat-to-beat interval vs beat count around puff revealed that the interval indeed increased for the three beats following puff across different puff strengths (Supplementary Fig. 2e). Consistently, the average interval over three beats before and after puff per fly revealed larger differences for 50mL/min and especially 200mL/min but no noticeable changes for 20mL/min, so that now we detect similar cardiac deceleration across puff strength (Supplementary Fig. 2f). However, this novel analysis unfortunately came with slightly larger variance overall (Supplementary Fig. 2f). We thus prefer to stick to HR in the present manuscript, but still agree that this is valuable for a more comprehensive understanding of cardiac puff responses. We included our new plot as Supplementary Fig. 2e, f, with the following description in the Result section:

L121-122: “Indeed, the beat-to-beat interval increased over the three beats following puff onset, and this increase was comparable across puff strengths (Supplementary Fig. 2e, f).”

(2) The author's model that cardiac deceleration leads to walking data is only partially supported by their data. In the first figure, the relationship between cardiac deceleration and walking probability seems to be inverted relative to their model (weak stimulus -> strong cardiac effect and weak locomotor effect; strong stimulus-> weak cardiac effect and strong locomotor effect). It is possible that this discrepancy may disappear when the authors look at beat duration rather than heart rate (for instance, if following the strong stimulus, there is a very long beat that is followed by tachycardia, thus weakening their observed HR change). It would also be easier to relate this data in Figure 1 to their interoceptive model if some data were shown that illustrated the relative timing of the cardiac change and the locomotor start.

Following the reviewer's suggestion, we re-examined the cardiac response using the beat-to-beat interval rather than heart rate. As the reviewer anticipated, this partially resolves the apparent discrepancy. Averaging the interval over the three beats before and after the puff (per fly) reveals a significant increase in beat-to-beat interval at every puff strength, with comparable magnitude across strengths (Supplementary Fig. 2e, f). However, since a comparable cardiac response, in terms of beat-to-beat interval, is accompanied by a graded locomotor response, locomotion cannot be driven by cardiac deceleration alone. Indeed, puff-induced cardiac deceleration and locomotion fail to correlate (added as Supplementary Fig. 10a in our updated manuscript). In contrast, optogenetic cardiac deceleration does increase locomotion (Fig. 5h-k), and with a strong correlation between optogenetically activated cardiac deceleration and locomotion (added as Fig. 5l in our updated manuscript). Taken together, locomotion is likely driven by both cardiac deceleration and a parallel, cardiac-independent pathway, with the latter accounting for the puff-strength dependence (Fig. 5m). We added the following description in the Result section:

L121-122: “Indeed, the beat-to-beat interval increased over the three beats following puff onset, and this increase was comparable across puff strengths (Supplementary Fig. 2e, f).”

Regarding the relative timing, we added a dedicated figure as Fig. 1h, which shows that the onsets of the cardiac and locomotor changes are approximately concurrent across puff strengths. They differ in their peak timing — the cardiac deceleration reaches its extremum in roughly 0.2 – 0.5 s, while the walking fraction peaks later (roughly 1 – 1.5 s). We have added the following to the Result section:

L127-139: “A close inspection of the peri-stimulus dynamics revealed that the onsets of the cardiac and locomotor changes were approximately concurrent at puff onset across strengths (Fig. 1h). Furthermore, the peak of cardiac deceleration (roughly 0.2 – 0.5 s post puff onset) preceded that of locomotor increase (roughly 1 – 1.5 s post puff onset). Especially at lower puff strengths, locomotor response showed an additional transient dip, embedded in dominant and persistent locomotion, almost concurrently, but in the case of 50 mL/min following, cardiac deceleration (Fig. 1h). Overall, cardiac and locomotor responses begin together, and the cardiac response reaches its peak before the locomotor response does. These data are consistent with the idea that cardiac and locomotor responses are coordinated, and that cardiac deceleration may causally contribute to the locomotor response. We therefore set out to identify the neurons responsible for cardiac deceleration and to test their causal role directly.”

We have also updated our corresponding Discussion section (chapter 1) accordingly:

L369-389: “Cardiac deceleration is coordinated with defensive locomotion

Our data show that air puff stimulation induces cardiac deceleration together with increased locomotion (Fig. 1). Across puff strengths, locomotion scaled with stimulus strength, whereas the cardiac deceleration did not: measured as time-binned HR, it was less pronounced at stronger puffs, while measured as the beat-to-beat interval it was comparable across strengths (Fig. 1d–h; Supplementary Fig. 2f). This indicates that locomotion is not simply a graded readout of cardiac deceleration.

Although cardiac deceleration is often associated with freezing in mammals [1], cardiac deceleration has also been observed during early threat detection and orienting responses [78], suggesting that it reflects engagement of a defensive processing rather than motor inhibition per se. Consistent with this view, dissociation between threat-evoked cardiac modulation and freezing has been reported across taxa. In mammals, for example, disruption of autonomic output can abolish cardiac deceleration without eliminating freezing [79]. Similarly, in decapod crustaceans and Drosophila, visual stimuli elicit reproducible and time-locked changes in heart rate that accompany both freezing and fleeing [8,9,80]. While these studies establish that cardiac modulation can be dissociated from freezing, our findings additionally demonstrate that cardiac deceleration can co-occur with a predominantly locomotor response to mechanical threat and can itself causally contribute to locomotion (Fig. 5e-m). Cardiac deceleration is therefore not restricted to passive defensive states, but can act upstream of active locomotor responses.”

We also examined HR-versus-locomotion relationships trial-by-trial and fly-by-fly, but trial-to-trial fluctuations in both signals were too large to pinpoint the timing of changes at that level, which is why we present the population peri-stimulus dynamics in Fig. 1h.

(3) Also, since the locomotor and cardiac changes are probabilistic, it would be very useful to see how their respective probabilities change when conditioned on the other. According to their interoceptive model, locomotion should preferentially increase on trials where cardiac deceleration occurs. The authors should discuss this incongruity and also potential alternative interpretations of their cardiac manipulation experiments. Perhaps the bradycardia makes them more sensitive to threats - as suggested in the introduction? Control flies show a mild increase in locomotion following green light (Figure 5j), so perhaps by slowing the heart, they are more sensitive and thus respond more strongly to this stimulus?

We believe that cardiac and locomotor changes are predicted to be congruent across flies only if locomotion is solely dependent on cardiac changes. Our optogenetic cardiac deceleration experiment likely corresponds to such a situation, and we indeed observed a strong correlation between cardiac and locomotor changes for retinal-fed test group but not for retinal-unfed control group (which we now present as Fig. 5l).

As for stimulus-induced locomotion, it may reflect contributions from cardiac-state-independent locomotor pathways (Fig. 5m). In such a situation, cardiac and locomotor changes are not necessarily predicted to be congruent. Indeed, puff-induced locomotion and cardiac changes do not appear to be congruent across flies (which we now present as Supplementary Fig. 10a).

We have added these plots as Fig. 5l and Supplementary Fig. 10a, and added the following explanation in the main text:

L335 - 341: “Furthermore, optogenetically induced changes in HR and locomotion were strongly anti-correlated fly-wise in the retinal-fed test group (Pearson’s correlation coefficient: -0.81, p-value: 0.00538) but not in the control group without retinal (Pearson’s correlation coefficient: 0.43, p-value: 0.436) (Fig. 5l). This contrasts with puff-induced responses, in which changes in HR and locomotion were not significantly correlated (Supplementary Fig. 10a). This lack of correlation may reflect contributions from cardiac-state-independent locomotor pathways.”

As for the possibility of bradycardia promoting locomotor response to threats (light), We thank the reviewer for raising this alternative interpretation. We agree that our experiment does not distinguish whether cardiac deceleration directly promotes locomotion or instead enhances the locomotor response to the green-light stimulus itself. Indeed, retinal-unfed control flies showed a modest increase in locomotion following green-light illumination, indicating that the light can itself influence behavior. Thus, the larger locomotor response in retinal-fed flies could in principle reflect an interaction between the optogenetically induced cardiac state and this sensory response.

We note, however, that this interpretation would still imply a causal influence of cardiac state on behavior: cardiac deceleration would increase the gain of an externally evoked locomotor response rather than directly initiating locomotion. Because our experiment does not discriminate between these possibilities, we have revised our manuscript to state that experimentally induced cardiac deceleration can promote locomotion or potentiate a light-evoked locomotion.

L346–L350: "We also note that retinal-unfed controls showed a modest locomotor increase following green-light illumination (Fig. 5j, k), leaving open whether cardiac deceleration directly promotes locomotion or potentiates a light-evoked locomotion. In either case, our data support a modulatory influence of cardiac state on locomotor output.”

We thank the reviewer for significantly strengthening our manuscript.

(4) Looking at the example shapes of the beats in Figure 5g versus Figure 1c, the optogenetically induced diastole has a very different shape from the naturally occurring long beat. Thus, the exact cardiac stimulus may be unnatural. If this is true across trials and animals, it may be worth considering that the funny beat (like an anxiogenic atrial fibrillation in mammals) is the source of the fear and, in turn, locomotor behavior (also interesting!) rather than being a true replication of the cardiac events seen following the puff stimulus.

We thank the reviewer for this insightful comment. We agree that the optogenetically induced cardiac deceleration does not fully reproduce every feature of the puff-induced heartbeat, and we have been careful not to claim that it does. Our aim was to test whether cardiac deceleration can influence locomotor state, not to claim that the optogenetically induced beat is an exact replica of the puff-induced response.

At the level of HR, the two manipulations were closely matched: both produced an ~10% reduction with a similar time course. They differed in the finer beat-to-beat structure — the optogenetic deceleration was concentrated in a transient prolongation at stimulus onset, whereas the puff-induced deceleration was distributed across several beats (Supplementary Fig. 10). In both cases, however, the M-mode recordings show that the heart maintained an organized rhythm, with the manipulation producing a transient diastolic prolongation rather than sustained beat-to-beat irregularity (Fig. 1c, Fig. 5g). This argues against the possibility that a fibrillation-like funny beat underlies the behavioral effect, but leaves open the possibility of light-induced one long interval, vs puff-induced long intervals across three consecutive beats, may have contributed to behavior. We also note that our current imaging system does not yet permit a comprehensive comparison of finer heartbeat features, such as full contraction dynamics, waveform shape, stroke volume/ejection fraction, or beat-pattern irregularity quantified across trials and animals. We therefore do not intend to exclude the possibility that specific features of the optogenetically induced beat pattern contribute to the behavioral effect — an interesting question that we agree is worth pursuing.

We have revised the manuscript accordingly. We now state, with an added figure (Supplementary Fig. 10), that the optogenetic experiment supports that experimentally induced cardiac deceleration can modulate locomotor behavior, but does not establish that the artificial cardiac pattern is identical to the mechanically evoked response. We also note the open possibility that distinct heartbeat patterns may themselves influence defensive behavior.

L343–L346: " One limitation is that our optogenetically induced cardiac pattern (Fig. 5g; Supplementary Fig. 10b, c) is not necessarily identical to its mechanically induced counterpart (Fig. 1c; Supplementary Fig. 10d, e), which could have influenced locomotion.”

L321-324: “To directly test whether cardiac deceleration can influence locomotion, we optogenetically induced cardiac deceleration of a magnitude comparable to the puff-induced response by expressing the light-gated anion channel GtACR1 in cardiomyocytes using Hand-GAL4 (Fig. 5e).”

Assessing how heartbeat pattern influences behavior will require improved whole-heart imaging and more detailed quantitative analysis of heartbeat dynamics, which we are currently developing but consider beyond the scope of the present study.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors):

Recommendation for the authors:

(1) In the introduction section (around line 43), it would be important to also include relevant work on the relationship between cardiac activity and threat responses in Drosophila (Barrios et al, 2021) as this study describes the internal physiological states of fruit flies during defensive responses (including the changes in heartbeat).

We thank the reviewer for pointing this out. The suggested paper was referenced in our original Introduction but slightly later (L78) than the location suggested by the reviewer (L44). We have now included this reference in L44 as well.

(2) Activation experiments could be performed in the presence of an air puff. The authors show that the stronger the air puff rate, the stronger the response. Low or medium air puff could be used during optogenetic activation of serotonergic and octopaminergic neurons, and see if the activation of different neuron types enhances cardiac deceleration and increase of locomotion (to levels similar to those observed, for example, at high puff rates). Otherwise, the claim that the dopaminergic system is solely responsible should be rephrased.

We agree that the suggested experiment can more reliably assess the potential contribution of serotonergic and octopaminergic systems. However, our intention was not to claim that dopaminergic system is solely responsible for puff-induced cardiac deceleration. Rather, consistency across loss- and gain-of-function experiments led us to focus on dopaminergic system, which allowed us to identify DA-WED neurons. We have revised our manuscript to more precisely express our view outlined above, as follows:

L194-196: “Together, our data demonstrates that, while not excluding the contributions of serotonergic or octopaminergic systems, dopamine system is responsible for puff-induced cardiac deceleration.”

(3) Since there is not enough proof that cardiac deceleration is the one that triggers an increase in locomotion during air puff responses and cannot be excluded that the DA-WED neurons trigger an increase in locomotion that is further modulated by the cardiac deceleration. Both options should be considered (models in Figures 4 and 5, results and discussion).

We thank the reviewer for this insightful comment, which is also consistent with reviewer#1’s view. We agree that it is better to mention the possibility that DA-WED neurons might influence locomotion through an additional or partially independent pathway beyond their role in cardiac regulation, which our data does not exclude. In our updated manuscript, we added to our schematic model an arrow emanating from DA-WED to locomotion directly (Fig. 5m), explaining our interpretation in the corresponding legend. We also included the discussion as below:

L439 - 441: “We note, however, that our data does not exclude the possibility that DA-WED neurons might influence locomotion at least partially independent of its role in cardiac deceleration.”

(4) Pages 14-15 (and in discussion). It appears that in some cases, puff and threat are used interchangeably, which may be confusing, as for example the authors ask if DA-WED could be general threat detectors in this context, meaning, if I understand correctly, that they are involved in all the aspects of the behavioral response to an air puff. However, the term general threat detector may be understood to mean not specific to air puff but also other types of threats, which is not addressed in this study. I suggest this be rephrased to avoid confusion. Also, they should be clear in the text when they refer specifically to air puff threats and when to threat responses.

We thank the reviewer for this important clarification. We swapped the phrase “general threat detectors” with “puff detectors” for precision. We have revised our manuscript accordingly, as follows:

L306 - 320: “An alternative interpretation is that DA-WED neurons function as mere puff detectors that broadly regulate puff-evoked behaviors. If this were the case, manipulating these neurons should affect all behavioral outputs induced by puff. In addition to promoting locomotion, air puff suppressed grooming predominantly within 3 s of puff onset, consistently across puff strengths; grooming during the later 10–20 s window was variable, either remaining suppressed, returning to baseline, or reversing to an increase depending on puff strength (Supplementary Fig. 9a–c). However, silencing DA-WED neurons failed to disrupt puff-induced grooming suppression (Supplementary Fig. 9d–f), despite markedly reducing cardiac deceleration (Fig. 4b, c) and locomotion (Fig. 5a, b). Likewise, photoactivation of DA-WED neurons failed to suppress grooming (Supplementary Fig. 9g–i), while robustly inducing cardiac deceleration (Fig. 4i, j) and locomotion (Fig. 5c, d). Thus, manipulation of DA-WED neurons does not reproduce the full behavioral repertoire triggered by puff. These findings argue against the idea that DA-WED neurons function as mere puff detectors and instead support a more specific role in coordinating cardiac and locomotor responses.”

To better distinguish “external threat” vs “threat responses”, we have also made the following changes in our main text:

L303-305: “These findings suggest that DA-WED neurons coordinate cardiac deceleration with locomotor responses upon mechanical threat.”

L341-343: “These observations are consistent with the possibility that cardiac deceleration may influence locomotor responses upon mechanical threat”

L437-439: “Our data points to a scenario in which DA-WED neurons induce cardiac deceleration upon mechanical threat, which may in turn influence threat-induced locomotion (Fig. 5m).”

L477-478: “While our data are consistent with a model in which DA-WED neurons couple cardiac deceleration to locomotor responses upon mechanical threat”

(5) In Supplementary Figure 3, the R74C07 labels anterior and posterior sensilla. Since the external sensilla on the posterior abdomen were excluded based on the Aes1N driver's result, the phenotype is likely due to the anterior sensilla. However, those are not included in the summary in the i) panel. Why is that?

We thank the reviewer for pointing this out. Our policy was that “When multiple drivers labeled the given neurons, we considered such neurons as indispensable for cardiac deceleration only when those drivers produced consistent results (hence, e.g. external sensilla on the posterior abdomen were excluded based on Aes1N driver's result, despite R74C07 driver's result suggests otherwise)”. Accordingly, sensilla on the eye compound (“anterior” sensilla in reviewer’s comment) labeled by R74C07 driver were excluded by “R38B08AD-VT043775DBD” split combination’s result. Our original legend of Supplementary Fig. 3i explained our policy, but to avoid confusion, our updated legend now includes above explanation as well, as follows:

“(hence, e.g. external sensilla on the posterior abdomen were excluded based on Aes1N driver's result, despite R74C07 driver's result suggests otherwise; similarly, sensilla on the eye compound labeled by R74C07 driver were excluded by “R38B08-AD VT043775-DBD” split combination’s result).”

(6) There seems to be a strong off response to air puff in medial and not in lateral DA-WED neurons, suggesting that the single pair of neurons could have a different function in the context of air puff responses. That should be discussed.

We agree with the reviewer that there may be differences in puff responses in medial vs lateral DA-WED neurons (Supplementary Fig. 8c). We have included this in our Result section as follows:

L281 – 283: “We note, though, that the medial neurons tended to show offset response-like activities more pronounced than their lateral counterparts (Supplementary Fig. 8c), potentially hinting at functional heterogeneity within DA-WED neurons.”

(7) Grooming: In Supplementary Figure 9c, although not significant, the silencing of DA-WED does seem to have an effect on the grooming profile- averaging over the 10-second time window does not reveal a significant effect, but a finer analysis may reveal some differences. There seems to be an increase in grooming after 10s? (in e-h).

We added the 10–20 s window data as Supplementary Fig. 9c, f, i, l. We note, however, that this window falls outside the period of robust puff-evoked grooming suppression. In wild-type flies, air puff suppressed grooming predominantly within 3 s of puff onset, and this suppression was consistent in direction across all puff strengths (Supplementary Fig. 9a, b). Grooming during the 10–20 s window, by contrast, was variable, either remaining suppressed, returning to baseline, or reversing to an increase depending on puff strength (Supplementary Fig. 9a). We therefore regard the 0–3 s window as reflecting the most reliable puff-evoked grooming response. Within this window, we did not detect statistically significant effects of neuronal or cardiac manipulation on the grooming suppression. Independently of this, we also examined the average grooming fraction during the 10–20 s window: the apparent differences there are driven by a few unusually high-magnitude values and are not statistically significant even when those values are retained (Supplementary Fig. 9f, i, l).

We also made the following changes to the Results section (and added Supplementary Fig. 9c, f, i, l):

L308 – 317: "In addition to promoting locomotion, air puff suppressed grooming predominantly within 3 s of puff onset, consistently across puff strengths; grooming during the later 10–20 s window was variable, either remaining suppressed, returning to baseline, or reversing to an increase depending on puff strength (Supplementary Fig. 9a–c). Silencing DA-WED neurons failed to disrupt puff-induced grooming suppression (Supplementary Fig. 9d–f), despite markedly reducing cardiac deceleration (Fig. 4b, c) and locomotion (Fig. 5a, b). Likewise, photoactivation of DA-WED neurons failed to suppress grooming (Supplementary Fig. 9g–i), while robustly inducing cardiac deceleration (Fig. 4b, c) and locomotion (Fig. 5c, d)."

L333 – 334: “In contrast, this manipulation failed to alter grooming behavior (Supplementary Fig. 9j-l).”

We also spotted an inaccurate description in our original legend, which read “Change in grooming fraction, calculated for each fly as the % difference between the mean fraction during the 10 s following stimulation onset and the mean fraction during the 10 s baseline period preceding stimulation.” This was an error, and we have instead used “during the 3 s following stimulation onset” window as we accurately described in our original Method (L579 – 582: “Puff-induced changes in walking and grooming were quantified by comparing the mean fraction during the 10 s and 3 s, respectively, following puff onset with the mean fraction during the 10 s baseline period preceding puff onset.”). We have corrected this error in our updated legend. We apologize for any confusion this error might have caused.

Reviewer #3 (Recommendations for the authors):

(1) Showing the model in Figure 4k as puff in parallel leads to locomotor effects independently of DA-WED, which is incongruent with the next figure where they discuss their interoceptive model. It might be useful to present the interoceptive model as a competing model in 4K that will be tested in the following figure.

Our intention there was to graphically summarize the tentative model that was plausible consistent with data presented in Fig.1-4. While we much appreciate the reviewer’s suggestion, we prefer not to present interoception idea in Fig. 4k: since the interoception idea comes from observations in Fig. 5a-d, we fear that presenting interoception model in Fig. 4k could be confusing without looking at Fig. 5a-d. Still, the reviewer’s suggestion gave us an opportunity to revisit our manuscript’s structure and for that we are thankful.

(2) The effect the authors observe is similar to one that has been [1.1] recently published in zebrafish larvae (Herrera et al. 2026 and Hernandez-Nunez et al. 2026), where lengthened heartbeats accompany active avoidance behavior following threats. Interestingly, this effect seems to be particularly conserved amongst smaller animals like fish larvae and insects, where heartbeats may represent a stronger sensory stimulus. It may be worthwhile for the authors to discuss these parallels. Further, some of the difficulties about interpreting direct optogenetic manipulation of cardiac function are discussed and explored in the Herrera et al. paper.

We thank the reviewer for highlighting these studies and this intriguing parallel. We agree the correspondence across distant taxa is striking. In larval zebrafish, cardiac dynamics and behavioral state are coupled following threat (Herrera et al., 2026), and dedicated vagal sensory neurons encode HR, including subsets responsive to cardiac deceleration, through a largely stimulus (including optogenetic cardiac pacing)-invariant population code (Hernandez-Nunez et al., 2025). Together with the demonstration in mice that optogenetically induced tachycardia can drive an anxiety-like state (Hsueh et al., 2023), and our present findings in Drosophila, these results suggest that cardiac-interoceptive influence on behavior is a broadly shared feature of nervous systems across taxa.

We find intriguing the reviewer's suggestion that such coupling may be especially relevant in small animals. Indeed, in small animals, the heart may perturb the internal state more globally - in small animals, the per-beat change in physiological state relative to body size may be proportionally larger, rendering each beat a stronger signal rather than being buffered by a large, compartmentalized vasculature. Alternatively, the compactness of small nervous systems could allow an interoceptive cardiac signal to reach motor circuits through few intervening synapses. In Drosophila, almost any neuron can reach any other within roughly four synapses (Scheffer et al., 2020; Dorkenwald et al., 2024; Schlegel et al., 2024), although whether synaptic path lengths are shorter in Drosophila than in larger brains remains unknown. We note, however, that the demonstration of cardiogenic affective control in mammals (Hsueh et al., 2023) indicates the phenomenon is not restricted to small animals, and so we consider the size question open.

Finally, following the reviewer's pointer to the interpretive considerations raised by Herrera et al., we have added text acknowledging that our optogenetic manipulation may engage the cardiac state through multiple routes. The term interoception is used broadly to include neural as well as humoral, blood pressure, or gas-exchange-related sensing of the cardiac state, and we note that distinguishing among these routes would require targeted manipulations we have not performed.

We have added the following section to our Discussion:

L453 – 474: “ Cardiac dynamics can modulate behavior

Our findings echo a growing body of work linking cardiac dynamics to behavior across distant taxa. In larval zebrafish, cardiac dynamics and behavior are coupled during threat [85], and dedicated vagal sensory neurons encode HR, including its decreases, through a stimulus-invariant population code86. In mice, optogenetically induced tachycardia can drive anxiety-like behavior [87]. And here we find that cardiac deceleration accompanies, and can promote, locomotion in Drosophila. Cardiac influence on behavior thus spans insects, fish, and mammals.

How cardiac state might be sensed in Drosophila is unknown, and several routes are conceivable. Drosophila heart receives rich neuronal innervation [51,52], although whether these neurons relay sensory information from the heart is unknown. If so, neurons could directly detect the heartbeat, analogous to the cardiac afferents that mediate cardiac interoception in vertebrates [86,87]. Alternatively, the heartbeat could be sensed mechanically or through gas-exchange-related cues. In mammals, the cardiac pressure pulse can be detected directly within the brain by mechanosensitive central neurons88. In zebrafish, cardiac activity can causally influence behavior through its oxygen supply [85]. In insects, although the circulatory architecture differs, the heartbeat contributes to pressure changes in the hemocoel and tracheal system and resultant gas exchange [89,90], which could in principle be detected. While present study does not distinguish among these possibilities, establishing how cardiac state is sensed would provide a foundation for dissecting how central commands and peripheral physiological states interact to coordinate survival behaviors.

(3) Serotonergic inhibition or knockout seems to diminish cardiac deceleration (lines 159-162), yet activation leads to increased heart rate (lines 170-171). The authors should discuss some possible explanations for this paradoxical result.

While we can only speculate, one possibility is that serotonergic neurons modulate cardiac control in a context-dependent manner. During mechanical threat, serotonergic activity is indispensable for cardiac deceleration, perhaps by gating the threat-responsive cardiac deceleration pathway involving DA-WED. In contrast, thermogenetic activation of serotonergic neurons outside this context may recruit broader downstream targets, including circuits that induce cardiac acceleration. We have added the following sentences in our Result:

L183 – 187: “Serotonergic system, on the other hand, provided paradoxical results (Fig. 2g-k, Supplementary Fig. 5c-g): this could be because serotonergic system is required for threat-induced cardiac deceleration, while broad serotonergic activation outside this context possibly engages downstream circuits that promote cardiac acceleration.”

(4) The split-GAL4 technology is a well-established approach in Drosophila research. The authors don't need to go into so much detail about its mechanism in the results (lines 199-205).

We have compressed our detailed explanation about split-GAL4 technology as follows:

L212 – 214: “In this approach, GAL4 is divided into DBD and AD hemi-drivers, and functional GAL4 is reconstituted only in cells expressing both components (Fig. 3c).”

eLife Assessment

This study utilizes a two-choice assay to examine how groups of Drosophila melanogaster respond to temperature changes. The findings are useful, offering focused importance for the collective decision-making community. However, the strength of the evidence is incomplete, as the current data only partially support the primary claims and the work lacks the necessary integration into the existing literature to contextualize the findings.

Reviewer #1 (Public Review):

Summary:

This study by Z. Moradi, S. Jahedazad, and colleagues presents a refined two-choice assay paradigm for investigating thermotactic behavior in groups of Drosophila melanogaster. The authors developed an experimental setup that maintains a stable temperature profile and utilized supervised classifiers based on extracted locomotion parameters to distinguish between temperature conditions within the arena as well as antennal ablation status. The integration of a Braitenberg vehicle model to simulate fruit fly exploration patterns and U-turn behaviors at thermal boundaries represents a valuable contribution to the study of collective thermotaxis.

Strengths:

(1) The detailed description of the behavioral assay and the rigorous calibration methods make the experimental setup reproducible.

(2) The incorporation of a state-of-the-art animal tracking system enables the precise extraction of multiple locomotion parameters, providing valuable insights into the fruit fly exploration as well as for predicting temperature-dependent behavioral responses. The methodology could be potentially adopted in other studies.

(3) The authors' decision to study thermotactic behavior in groups is a noteworthy addition. The finding that overall staccato movement patterns remain largely unaffected by antennal ablation, while the frequency and duration of pairwise encounters increase with temperature, offers valuable insights into individual sensory processing vs the influence of collective behavior.

Weaknesses:

(1) The study investigates groups of fruit flies, which is an interesting contrast to previous single-fly studies. While group dynamics are mentioned as potentially important, no direct comparison or analysis was conducted between individual vs group, which would have been a straightforward and valuable addition. The strength of intergroup interactions remains unclear, as the pairwise encounter parameters appeared to have low significance according to the importance factor analysis.

(2) The manuscript does not clearly articulate the novelty of the work; the innovative aspects are present but not effectively highlighted.

(3) The descriptions of the locomotor parameters and Braitenberg modeling, which constitute the most critical components of the study, are not optimally presented in the main text. The authors should provide a more detailed rationale for why certain locomotion parameters emerged as more influential (i.e., with higher importance scores) than others. Additionally, clearer explanations from a physical or mechanistic perspective would help readers better appreciate and interpret the significance of these findings.

Reviewer #2 (Public Review):

The authors develop an experimental framework to precisely control temperature in an arena and make quantitative measurements of thermal avoidance behavior in Drosophila. They find that wild-type animals show robust heat avoidance behavior and suggest that antennal ablation reduces this heat avoidance. They also make a novel claim that antennal inputs do more than simple heat detection but play a role in navigation strategies like centrophobism and are needed for decision-making at thermal boundaries. The fundamental two-choice behavior paradigm or the use of peltier modules for temperature control have been employed previously in the field. The strength of this work lies in detailed behavioral measurements. However, the study falls short due to a lack of adequate experimental controls. The authors then model this behavior as a classical Braitenberg vehicle of sensory-motor transformation. Overall, the paper has some interesting insights into in-depth analysis of Drosophila navigation in thermal gradients, but their evidence only partially supports their claims.

Reviewer #3 (Public Review):

Per my training and expertise, my evaluation and feedback below are focused on the behavioral components of this study and not the modeling. The authors use heat avoidance thermotaxis behavior in adult flies as a behavioral read out of decision-making with the goal of providing a novel quantitative analysis of the navigational strategies that contribute to heat avoidance behavior. Breaking down a specific behavior (here, heat avoidance) into component behavioral motifs can provide a more detailed analysis that can contribute to deeper understanding of navigational strategies within Drosophila and potentially aid understanding of navigational strategies at a broader level. The authors describe temperature preference behaviors, as has been previously reported and examine the role(s) of thermoreceptors localized to the antennae compared to the rest of the fly. The thermotaxis behavior is then broken down into several parameters with the goal of gaining deeper insight into how and when the fly initiates and generates heat avoidance behavior. The work would be further strengthened and clarified by more completely setting the findings within the context of the field both by including relevant citations and via text that describes this within the manuscript. Additionally, while the authors discuss choosing the temperatures in the assay to allow for more "straightforward interpretation", it would be important to note that this might not be what the fly actually experiences and although it might drive a "decision" it might not be revealing the ethologically accurate behavior. Additionally, Figure 2 in particular can be revised to aid the reader in understanding and evaluating the results. For example, writing out the terms on the y-axis (or making a table a reader could easily reference) and including statistically significant findings on the figure itself would allow the reader to more easily draw their own conclusions. Finally, a more thorough discussion would help make the findings more accessible and demonstrate how these findings have the potential to be of broad interest to many disciplines. For example, placing findings in the context of navigation and quantitative behavioral analysis across species and referring to figures within the discussion section itself.

Author response:

Reviewer #1 (Public Review):

Summary:

This study by Z. Moradi, S. Jahedazad, and colleagues presents a refined two-choice assay paradigm for investigating thermotactic behavior in groups of Drosophila melanogaster. The authors developed an experimental setup that maintains a stable temperature profile and utilized supervised classifiers based on extracted locomotion parameters to distinguish between temperature conditions within the arena as well as antennal ablation status. The integration of a Braitenberg vehicle model to simulate fruit fly exploration patterns and U-turn behaviors at thermal boundaries represents a valuable contribution to the study of collective thermotaxis.

Strengths:

(1) The detailed description of the behavioral assay and the rigorous calibration methods make the experimental setup reproducible.

(2) The incorporation of a state-of-the-art animal tracking system enables the precise extraction of multiple locomotion parameters, providing valuable insights into the fruit fly exploration as well as for predicting temperature-dependent behavioral responses. The methodology could be potentially adopted in other studies.

(3) The authors' decision to study thermotactic behavior in groups is a noteworthy addition. The finding that overall staccato movement patterns remain largely unaffected by antennal ablation, while the frequency and duration of pairwise encounters increase with temperature, offers valuable insights into individual sensory processing vs the influence of collective behavior.

Weaknesses:

(1.1) The study investigates groups of fruit flies, which is an interesting contrast to previous single-fly studies. While group dynamics are mentioned as potentially important, no direct comparison or analysis was conducted between individual vs group, which would have been a straightforward and valuable addition. 

We appreciate the reviewers' suggestion regarding a direct single-fly versus group-fly comparison. In response, we have added a preliminary analysis comparing single-fly and group-fly trials, with the results presented in the revised manuscript under “Group-Enhanced Heat Avoidance in Drosophila Correlated with Pairwise Encounter Duration” This analysis provides an initial assessment of whether the group-level heat avoidance response differs from that observed in individual flies.

We must, however, be transparent about the limitations of this additional data. Due to exceptional conditions in Iran during this period, and the completion of the degree program under which this research was conducted, a full comparative dataset could not be collected. The analysis presented is therefore based on preliminary data and should be interpreted as indicative rather than conclusive, and we have clearly noted this limitation in the manuscript.

We would also like to clarify the scope and intent of this work. As engineers, our primary contribution is methodological: we aimed to develop tools and a framework for temperature-response analysis in Drosophila, with particular focus on enabling such analysis at the collective level, a setting that has received comparatively less attention. A direct behavioral comparison between single and multi-fly conditions was not the original objective of this study. We believe the core contribution of the paper, the arena implementation methodology, analysis pipeline, and collective behavior characterization, stands independently, and we hope the added preliminary comparison addresses the reviewers' concern in the spirit in which it was intended.

(1.2) The strength of intergroup interactions remains unclear, as the pairwise encounter parameters appeared to have low significance according to the importance factor analysis.

Thank you for raising this point. First of all, we would like to clarify that, in response to this and concern 21, we have revised the figure, which was previously based on one-way ANOVA scores, to a permutationbased feature importance score (Figure 3). This updated method better reflects the functional contribution of each feature in relation to our trained classifiers. As demonstrated, for the classification of antennal deficits, encounter duration emerges as the fourth most significant feature affecting classification performance. 

Moreover, we inserted a new section “Group-Enhanced Heat Avoidance in Drosophila Correlated with Pairwise Encounter Duration”, deeply explain the relationship between enhanced AI values in groups (than individual trials) and pairwise interactions.

We recognize that the impact of social interactions on group thermal responses is a compelling yet underexplored area that deserves further investigation. Therefore, we included some suggestions in the “Limitations and Future Directions” section of our updated manuscript.

(2) The manuscript does not clearly articulate the novelty of the work; the innovative aspects are present but not effectively highlighted.

We appreciate your comment and have revised the manuscript accordingly. In particular, we have revised the Introduction section, which now includes a detailed explanation of the novelty of our work and highlighting its distinctions from previous related studies.

(3.1) The descriptions of the locomotor parameters and Braitenberg modeling, which constitute the most critical components of the study, are not optimally presented in the main text. 

We revised our manuscript. For your reference, please check out “Table 1” and the “A Simple Braitenberg Vehicle Recapitulated Fly Heat-Avoidance Navigational Strategy” section.

(3.2) The authors should provide a more detailed rationale for why certain locomotion parameters emerged as more influential (i.e., with higher importance scores) than others. Additionally, clearer explanations from a physical or mechanistic perspective would help readers better appreciate and interpret the significance of these findings.

We tried to consider this in the entire revision process of our manuscript.

Reviewer #2 (Public Review)

The authors develop an experimental framework to precisely control temperature in an arena and make quantitative measurements of thermal avoidance behavior in Drosophila. They find that wild-type animals show robust heat avoidance behavior and suggest that antennal ablation reduces this heat avoidance. They also make a novel claim that antennal inputs do more than simple heat detection but play a role in navigation strategies like centrophobism and are needed for decision-making at thermal boundaries. The fundamental two-choice behavior paradigm or the use of peltier modules for temperature control have been employed previously in the field. The strength of this work lies in detailed behavioral measurements. However, the study falls short due to a lack of adequate experimental controls. The authors then model this behavior as a classical Braitenberg vehicle of sensory-motor transformation. Overall, the paper has some interesting insights into in-depth analysis of Drosophila navigation in thermal gradients, but their evidence only partially supports their claims.

Reviewer #3 (Public Review)

Per my training and expertise, my evaluation and feedback below are focused on the behavioral components of this study and not the modeling. The authors use heat avoidance thermotaxis behavior in adult flies as a behavioral read out of decision-making with the goal of providing a novel quantitative analysis of the navigational strategies that contribute to heat avoidance behavior. Breaking down a specific behavior (here, heat avoidance) into component behavioral motifs can provide a more detailed analysis that can contribute to deeper understanding of navigational strategies within Drosophila and potentially aid understanding of navigational strategies at a broader level. The authors describe temperature preference behaviors, as has been previously reported and examine the role(s) of thermoreceptors localized to the antennae compared to the rest of the fly. The thermotaxis behavior is then broken down into several parameters with the goal of gaining deeper insight into how and when the fly initiates and generates heat avoidance behavior. The work would be further strengthened and clarified by more completely setting the findings within the context of the field both by including relevant citations and via text that describes this within the manuscript. Additionally, while the authors discuss choosing the temperatures in the assay to allow for more "straightforward interpretation", it would be important to note that this might not be what the fly actually experiences and although it might drive a "decision" it might not be revealing the ethologically accurate behavior.  

We interpret the reviewer’s point as a request to distinguish the temperatures used in our assay from the temperatures a fly would typically encounter in a natural setting. Our selected temperatures were chosen to provide experimentally clear thermal contrasts and facilitate quantitative interpretation; however, the temperature experienced by a fly in nature can vary spatially and temporally and may not correspond to the controlled conditions used here. Thus, a behavioral choice observed in our assay should not automatically be interpreted as an ethologically exact representation of natural thermal decision-making. We have clarified this point in the revised manuscript and now frame the assay as a controlled test of thermotactic behavior rather than a direct reproduction of a natural thermal environment.

Additionally, Figure 2 in particular can be revised to aid the reader in understanding and evaluating the results. For example, writing out the terms on the y-axis (or making a table a reader could easily reference) and including statistically significant findings on the figure itself would allow the reader to more easily draw their own conclusions. 

Table 1, included in the revised version, defines locomotor features along with their corresponding symbols. In Figure 2, blue median lines depict significant deviation from the baseline at 25°C for the wild-type group. Moreover, in the same figure, red median lines denote significant differences between wild-type and ablated groups at the warm temperature of 35°C. Features’ values (Mean ± SD) extracted from trials as well as the p-values for suitable overall statistical tests (ANOVA, welch ANOVA, Kruskal-Wallis) are included in the supplementary File 1. Adjusted p-values for all pairwise comparisons in the wild-type group are provided in Table 2.

Finally, a more thorough discussion would help make the findings more accessible and demonstrate how these findings have the potential to be of broad interest to many disciplines. For example, placing findings in the context of navigation and quantitative behavioral analysis across species and referring to figures within the discussion section itself.

We agree with the reviewer and have substantially revised the Discussion to better connect our findings to navigation and quantitative behavioral analysis. The revised Discussion places the locomotor features and thermal decision-making results in the broader context of behavioral navigation and explicitly refers to the relevant figures when interpreting the findings. We have also clarified which conclusions are directly supported by the present data and which represent potential broader applications or directions for future work.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Comments:

(1) The current study focuses on transcriptional regulation of the sex chromosomes. It would be interesting to know whether perturbation of pre-rRNA synthesis also affects transcription of autosomal genes.

Our RNA-seq analysis shows that BHM-21 treatment markedly affected the transcription of X- and Y-linked genes, with no obvious influence on the transcription of autosomal genes in pachytene spermatocytes (Figure 4B).

(2) Is ribosome biogenesis still active during prophase I of male meiosis? Additional discussion of the timing and extent of rRNA synthesis at this stage would help place the findings in a broader biological context.

To determine whether ribosome biogenesis remains active during meiotic prophase I, we examined 47S pre-rRNA levels by qPCR in spermatogonia (Spg) and prophase I spermatocytes. As shown in Figure S2, pre-rRNA transcript levels were extremely low in prophase I spermatocytes compared to Spg, indicating an arrest of rRNA synthesis upon meiotic entry. Notably, we observed prominent redistribution of nucleolar components to the XY body during prophase I. These findings suggest that reduced engagement of nucleolar components in rRNA biogenesis may facilitate their redistribution and their participation in other biological processes. We have added this discussion to the revised manuscript (Lines 283-289).

(3) A recent preprint reports active RNA polymerase II-mediated transcription of Y chromosome genes within nucleolus-like bodies (NLBs) during prophase I of meiosis in Drosophila male germ cells (https://doi.org/10.64898/2026.05.20.726666). These findings suggest that the meiotic nucleolus may have species-specific roles in regulating sex chromosome gene expression. It would be valuable for the authors to discuss how their findings compare with these observations and the potential evolutionary implications.

In Drosophila spermatocytes, the nucleolus is reorganized into a nucleolus-like body (NLB). This structure serves as a platform for RNA polymerase II-mediated transcription of Y-linked fertility genes. In contrast, in mouse spermatocytes, nucleolar components such as pre-rRNA and NPM1 are redistributed to the XY body. Their accumulation promotes Pol II exclusion and meiotic sex chromosome inactivation (MSCI). The transcriptional outcomes differ between the two species: Y-linked genes remain transcriptionally active in Drosophila, whereas sex chromosomes are transcriptionally silenced in the mouse. These findings suggest that nucleolar organization has been evolutionarily adapted in a species-specific manner to meet distinct requirements for sex chromosome regulation during spermatogenesis. We have discussed this point in the revised manuscript (Lines 317-326).

Reviewer #2 (Public review):

Summary:

The authors showed the localization pattern of nucleolus components, including Pre-rRNA, a precursor of rRNAs, changes during meiotic prophase I, particularly with the localization of these nucleolar components to the X-Y body, which shows inactivation of RNA polymerase II transcription, during pachynema. The localization of Pre-rRNA depends on ATR kinase and gammaH2AX. The chemical inhibition of rRNA transcription disrupts the binding of pre-rRNA to the X-Y body and suppresses the inhibition of the RNA polymerase II-mediated transcription on the sex chromosomes.

Strengths:

The cytological analysis, combined with the chemical inhibition, provided solid evidence to support the idea that, together with the remodeling of the nucleolus structure, pre-rRNA is an essential component of sex chromosome inactivation in male mouse meiosis. The role of pre-rRNA in sex chromosome inactivation in male meiosis helps our understanding of how the X-Y body, which would be a biological condensate, would be formed; e.g. for example, this Pre-rRNA may promote phase separation.

Weaknesses:

However, there is limited information on how Pre-rRNA is recruited to only sex chromosomes and how the RNA promotes the inactivation of sex chromosomes. Of course, these will be a target of future study. One major weakness of this paper is a poor description of the results, with fair presentation and interpretation of the data.

Following the reviewers suggestions, we have modified the manuscript. As listed below, we have addressed all the concerns raised by the reviewers point by point.

Recommendations for the authors:

Reviewer #2 (Public review):

Major points:

(1) The authors studied the localization of pre-rRNA, which is the center of this study. However, the general structure of this RNA was poorly described in the text. For example, how this pre-rRNA is different from mature rRNAs. Importantly, how the authors detected pre-rRNA, not other rRNA species, should be described in the Figure and the text.

The primary pre-rRNA transcript contains the sequences that give rise to the mature rRNAs, together with the external transcribed spacers (5′-ETS and 3′-ETS) and internal transcribed spacers (ITS1 and ITS2). During rRNA maturation, these spacer regions are progressively removed through a series of processing steps and are therefore absent from mature rRNAs (PMID: 22342225). To specifically detect pre-rRNA, we performed RNA-FISH using probes targeting the ITS1 and ITS2 regions as described in the Methods section (Line 380-398). This probe design strategy has been reported in our previous study and cited in the Methods section (PMID: 34980897). Following your suggestion, we have now added a brief description of the general structural features of pre-rRNA and its distinction from mature rRNAs in the Introduction (Lines 36-39).

(2) For the dynamics of the localization of pre-RNA and/or NMP1 during the meiotic prophase I, it is very important to check the localization of these components in zygonema (or leptonema) by ChIP-seq (Figure 2C) to address whether these signals bind to the sex chromosome prior to pachynema.

As suggested, we performed additional ChIRP-seq and CUT&Tag-seq to detect the localization of pre-rRNA and NPM1 in leptotene/zygotene spermatocytes (GSE346625, GSE346622). The results indicate that neither pre-rRNA nor NPM1 associates with sex chromosomes at the leptonema-zygonema stage prior to pachynema (Revised Figure S3).

(3) For Pre-rRNA ChIP, how did the author IP the RNA? Do they have specific antibodies against the RNA? Please explain this in the text.

The pre-rRNA-associated chromatin was not isolated using an RNA-specific antibody. Instead, we employed Chromatin Isolation by RNA Purification (ChIRP), a well-established method for mapping RNA–chromatin interactions (PMID: 24457971). In this assay, biotinylated antisense DNA oligonucleotides complementary to pre-rRNA sequences were used to specifically capture pre-rRNA together with its associated chromatin, followed by streptavidin-based purification. The experimental procedure has been described in the Methods section (Lines 467-475).

(4) For Pre-rRNA ChIP-seq and CUT&Tag-seq (Fig. 2C, 3C), the signals are enriched in specific genomic regions. Please show where they are and discuss them.

Pre-rRNA/NPM1 signals appear to be enriched within γH2AX-positive regions of the sex chromosomes (Figure S8). This spatial overlap is consistent with our finding that ATR-mediated DNA damage response signaling drives the sequestration of nucleolar components to the XY body. Supporting this interpretation, inhibition of ATR or loss of H2AX markedly reduced the enrichment of both pre-rRNA and NPM1 on the sex chromosomes. We have discussed this point in the revised manuscript (Lines 155-158).

(5) In Figure S2, the staining patterns of PLRE1 and FBL are different; FBL seems to bind to chromosome axes, which should be mentioned in the text and, if possible, discussed in terms of the differential localization of nucleolar components on the X-Y body.

FBL was specifically enriched on the unsynapsed axis of the X and Y chromosomes, where SPO11-mediated DSBs occur. However, POLR1E exhibited a diffuse distribution in the XY body. Thus, it is possible that these nucleolar proteins may be involved in meiosis-related events in the XY body, such as meiotic DSB repair. The differential localization pattern suggests that the nucleolar components may play distinct roles in the meiotic prophase I. We have included the description and discussion of this part in the revised manuscript (Lines 302-306).

(6) Graphs in Figure 3 showed the order of the binding after ATRi washing: gammaH2AX>NPM1>Pre-rRNA. This kinetic difference is confirmed and should be mentioned in the text (and discussed).

We agree that the distinct recovery kinetics of γH2AX, NPM1, and pre-rRNA upon ATRi washout support a hierarchical recruitment model. After ATRi washout, γH2AX signals reappeared most rapidly, followed by the gradual re-accumulation of NPM1 and pre-rRNA. This distinct kinetics suggests that ATR-dependent γH2AX signaling precedes the recruitment of NPM1 and pre-rRNA. We have included related information in the revised manuscript (Lines 155-158).

(7) Figure 3C: It is clear that the absence of pre-rRNA and NMP1 in the gammaH2AX KO testis. It is important to show the localization patterns of these components in pachytene KO cells; e.g., no change in the localization.

As suggested, we performed additional IF-FISH analysis in wild-type and H2AX-knockout pachytene spermatocytes to examine the subcellular localization of pre-rRNA and NPM1. Both pre-rRNA and NPM1 were prominently enriched at the XY body in wild-type pachytene spermatocytes. In contrast, this XY body-associated enrichment was lost in H2AX-knockout pachytene spermatocytes. Notably, pre-rRNA remained colocalized with NPM1 in H2AX-knockout cells. These observations suggest that H2AX is required for the recruitment of pre-rRNA/NPM1 to the XY body, but is not required for their spatial association. These new data have been included in Figure S9 and described in the revised manuscript (Lines 164-166).

(8) EU incorporation experiment in Figure 4: Signals of EU in this figure are different from those in Figure1 C. This suggests the distinct loci for rRNA production between leptonema and pachynema. Please explain more. More importantly, it is very important to show the amounts and production rates of ribosomes or rRNAs during distinct stages of prophase I.

5-EU (5-ethynyl uridine) is a nucleoside analog that is incorporated into all newly synthesized RNA, including both rRNA and mRNA, during active transcription. Therefore, the 5-EU signals we observed reflect the global transcriptional activity of all RNA polymerases (Pol I, Pol II, and Pol III), rather than rRNA production specifically. In pachytene spermatocytes, the XY body undergoes meiotic sex chromosome inactivation (MSCI), a process in which transcription on the sex chromosomes is largely silenced. However, Pol II-mediated transcription is robustly active on the autosomes during this stage. Therefore, the 5-EU signals we detected in pachytene cells predominantly originate from active Pol II transcription on autosomes, while the XY body remains largely devoid of 5-EU incorporation due to MSCI.

Following the reviewer’s suggestion, we further quantified pre-rRNA transcript levels at different stages of meiotic prophase I using quantitative RT-qPCR. Pre-rRNA transcript levels were extremely low in prophase I spermatocytes compared to Spg, indicating an arrest of rRNA synthesis upon meiotic entry (Revised Figure S2). These findings suggest that reduced engagement of nucleolar components in rRNA biogenesis may facilitate their redistribution and their participation in other biological processes. We have discussed this point in the revised manuscript (Lines 283-289).

(9) It is very important to show the localization of X-Y body markers such as Brca1, ATR, etc., in spermatocytes treated with BMH-21 to know which process is prevented with this inhibitor.

As suggested, we examined the localization of ATR and BRCA1 in pachytene spermatocytes following BMH-21 treatment. ATR localization to the XY body was unaffected by BMH-21 treatment. Together with our finding that ATR inhibition disrupts pre-rRNA accumulation at the XY body, these observations support the notion that ATR signaling is required upstream of pre-rRNA recruitment. In contrast, BRCA1 enrichment at the XY body was reduced by BMH-21, suggesting that BRCA1 accumulation at the XY body is potentially regulated by pre-rRNA. We have included these data in Figure S11C and S11D and added the corresponding description in the revised manuscript (Lines 180-181).

(10) Figure 4A clearly shows the localization of pre-rRNA in the X-Y body depends on active PolI transcription. Where is pre-rRNAs, which resided in the X-Y body after the treatment? Spreading to the nucleoplasm or degradation. Is it very important to analyze the amounts of rRNA species, including pre-rRNAs, under these conditions (and also during normal prophase I)?

As shown in Figure 4A, the prominent enrichment of pre-rRNA at the XY body was lost upon BMH-21 treatment. Instead, the residual pre-rRNAs were dispersed in the nucleoplasm. We also observed that BMH-21 treatment markedly reduced total pre-rRNA levels in pachytene spermatocytes (Author response image 1). Therefore, these results indicate that BMH-21 disrupts the enrichment of pre-rRNA at the XY body and reduces overall pre-rRNA abundance.

Author response image 1.

BMH-21 reduces pre-rRNA levels in pachytene spermatocytes. RT-qPCR analysis of pre-rRNA in pachytene spermatocytes treated with BMH-21 (2.5 μM, 12 hours). Data are presented as mean ± SEM (n=3 biological replicates).

(11) The authors need to study the localization of nucleolar components including Pre-rRNA, in female meiosis.

We agree that examining the localization and dynamics of nucleolar components during female meiosis would be of considerable interest. However, the present study focuses specifically on the redistribution of nucleolar components to the XY body during male meiotic prophase I. The XY body is a male-specific chromatin domain formed by the largely unsynapsed X and Y chromosomes and is closely associated with meiotic sex chromosome inactivation and the ATR-dependent DNA damage response. Because female meiotic cells lack an equivalent XY body structure, investigating nucleolar organization during female meiosis would address a distinct biological question that is beyond the scope of the present study.

Minor points:

(12) Line 61-62, subtelomeric region: "pseudo-autosomal region" is better.

We have replaced "subtelomeric region" with "pseudoautosomal region (PAR)" in the revised manuscript.

(13) Figure 1A: Please add the localization of pre-RNA (and other nucleolar components) in the mouse ES cells, etc., as a control. How the nucleolus in germ cells is similar or different from that in typical somatic cells is important.

We examined the localization of pre-rRNA in mouse embryonic fibroblasts (MEF) as a representative somatic cell type (Figure S1A). Pre-rRNA showed a similar nucleolar localization pattern in MEFs and spermatogonia (SPG), indicating that the general organization of pre-rRNA within the nucleolus is comparable between these two cell types.

(14) Line 83, in situ: Should be italicized.

We have corrected all instances of "in situ" to italicized format (in situ) throughout the manuscript.

(15) Figure 1A and text: Please explain what Scp3 is and why the authors used this.

Synaptonemal complex protein 3 (SYCP3) is a major structural component of the lateral elements of the synaptonemal complex, a proteinaceous structure formed between homologous chromosomes during meiotic prophase I. SYCP3 localizes along meiotic chromosome axes and is widely used as a marker to visualize chromosome organization and distinguish different substages of meiotic prophase I based on characteristic patterns of chromosome synapsis (PMID: 42059593). Therefore, we used SYCP3 staining to visualize meiotic chromosome axes and to identify the developmental stage of spermatocytes. We have now added a clear description of SYCP3 and the rationale for its use in the revised manuscript (Lines 54-56).

(16) Figure 1A and others, dispersion ratio: Please explain what the ratio is in the text and/or the legend.

The dispersion ratio is defined as the area of pre-rRNA signal divided by the total nuclear area (delineated by DAPI staining). This ratio reflects the degree of spatial dispersion of pre-rRNA within the nucleus: a higher ratio indicates a more diffuse distribution, whereas a lower ratio indicates a more concentrated localization. The dispersion ratio of NPM1 was calculated using the same method as that for pre-rRNA. We have now added a detailed explanation of this definition to the legend of Figure 1A.

(17) Figure S1A, top panels: A small portion of the signal of POLR1E (yellow) is out of the nucleus (white circle) and overlaps with c-Kit. What does this mean? Need to clarify.

The appearance of this small portion of POLR1E signal was most likely due to non-specific staining or an imaging artifact, as it was not reproducibly observed in other cells or independent experimental replicates. To avoid potential misinterpretation, we have replaced the original image with a new representative panel from the same experimental condition in the revised Figure S1A.

(18) Line 93. GC: Should be "NPM1", since this is the result.

We have corrected “GC” to “NPM1” in the corresponding sentence in the revised manuscript.

(19) Line 98-, Figure 1C: Please explain how the EU is incorporated into cells and what the "transient" means here in the legend.

5-EU (5-ethynyl uridine) is a cell-permeable uridine analog that is incorporated into newly synthesized RNA during active transcription (PMID: 40701178; PMID: 37683610; PMID: 29431736). In our experiment, we added 5-EU to the culture medium at a concentration of 1 mM for 2 hours to transiently label newly synthesized RNA. The term "transient" refers to the short pulse-labeling period (2 hours), which allows us to capture a snapshot of active transcription at the time of labeling, rather than steady-state RNA accumulation.

(20) Line 110, coalesce: Without live imaging, it is very hard to show that NPM1 (or pre-rRNA) bodies coalesce. The authors need to soften the wording here.

We have replaced "coalesce" with " aggregate" (Line 116).

(21) Line 114, pre-rRNA and NPM1 are recruited and sequestered in the X-Y body; however, in Figure 2A, B. No evidence of the colocalization of the signals with a marker for X-Y bodies or X, Y chromosomes. How is it possible to quantify the signals without the reference marker, or how is the area of the X-Y body defined (right graphs, line 114-116)? It is essential to show the colocalization analysis in cytology.

In our study, the XY body territory was identified based on the following criteria: a. SYCP3 staining was used as the primary marker to delineate the axes of the X and Y chromosomes, which remain unsynapsed during pachynema and are embedded within the XY body. b. DAPI staining served as a secondary marker, as the XY body typically appears as a DAPI-dense region (heterochromatin-enriched) within the nucleus. Accordingly, the XY body area in our quantification (Figure 2A and 2B) was defined as the SYCP3-defined XY chromosome axis region plus the surrounding DAPI-dense chromatin territory, following the established criteria in references 20 and 38. We have now added this information to the corresponding Materials and methods section (Lines 496-501) to clarify our morphological analysis criteria.

(22) Line 115 gradual accumulation: This is true for Pre-rRNA, but not NPM1, which showed a sudden increase mid- to late-pachynema.

We have revised the wording from "gradual" to "dynamic" in the revised manuscript (Line 121).

(23) Figure S3, statistical analysis: What kind of statistics did the authors use here? It seems to be a simple quantification of the ratios.

In Figure S5 (originally presented in Figure S3), we quantified the dispersion ratio of pre-rRNA (or NPM1) across different meiotic stages. As defined in the figure legend, the dispersion ratio was calculated as the pre-rRNA (or NPM1) signal area divided by the total nuclear area (delineated by DAPI staining), which reflects the degree of spatial dispersion of the signal within the nucleus. The numbers indicated above the data points represent the mean value for each group, which facilitates direct comparison of the overall trend across different stages.

(24) Figure S4, line 126: Lamin is not a marker for the nuclear envelope. If the authors insist on an intact nuclear envelope (NE), please use the other NE markers.

We agree that Lamin is not a direct marker for the nuclear envelope (NE) per se. While the nuclear lamina is closely associated with the NE and provides structural support to the nucleus, it is not the envelope itself. Notably, Lamin B1 is widely used as a marker to assess nuclear envelope integrity (PMID: 15353547), because disruption of Lamin B1 localization is frequently accompanied by nuclear envelope abnormalities (PMID: 32954426; PMID: 23166514). Therefore, we consider it reasonable to use Lamin B1 staining as an indicator of NE integrity in our experimental context.

(25) Line 143-144, Figure S5A: How did the authors define the morphology of the X-Y body (what marker was used for the morphological analysis)?

Upon re-evaluation, we realized that SYCP3 staining, which we had originally used as a reference, lacks sufficient specificity to define the morphology of the XY body. Therefore, in the revised manuscript, we have deleted the morphological description in Page 7.

(26) Line 145, released: No experimental evidence for the release. Please rephrase here.

We have replaced "released" with "absent" in the revised manuscript (Line 151).

(27) Line 165, the treatment of spermatocytes with BMH-21: This is an in vitro experiment using isolated spermatocytes? As shown above, please mention the system precisely.

As the reviewer noted, the experiments were performed in vitro on isolated spermatocytes. Specifically, spermatocytes were isolated from mouse testes and treated with BMH-21 at 2.5 μM for 12 hours in culture medium, followed by fixation and immunofluorescence and RNA FISH analysis. We have now clarified the experimental system (Lines 172-173) and in the Methods section.

(28) Figure S7B. Please measure the weight of the testis.

We have now included the testis weight in the revised Figure S12B (corresponding to the original Figure S7B).

(29) Line 194, PNA: What is PNA? Please explain it to readers.

PNA (peanut agglutinin) is used as a marker for the acrosome of post-meiotic germ cells (round and elongated spermatids), and PNA-positive cells therefore represent post-meiotic spermatids. We have added this information in the revised manuscript (Lines 202-203).

(30) Figure 5A, PolII staining: The images shown in the figure are not convincing (cannot see any signals), which is not consistent with the quantification. Moreover, the graph shows the "relative" amount of PolII. How did the authors define 100% values here? Mean values are not 100% in the two graphs.

We have now corrected the y-axis to "Pol II-positive area within the XY body", which accurately reflects what was quantified. In the quantification graphs, the "Pol II-positive area within the XY body" refers to the percentage of the XY body territory that is occupied by Pol II signal. Specifically, the XY body region was defined as the region of interest based on SYCP3 staining and DAPI-dense chromatin. We have now added this definition to the Methods section (lines 496-501) to ensure clarity for readers.

(31) For the ATAC-seq in Figure 6, what happens to the accessibility of chromatin of autosomes (and sex chromosomes) in the BMH-21-treated cells relative to the mock control? In the text, the authors compared the accessibility of sex chromosomes relative to the autosomes only in the treated cells. It is critical to show the difference in the accessibility of autosomes in the treated and untreated cells.

The overall chromatin accessibility of autosomes remained unchanged upon BMH-21 treatment, whereas the accessibility of sex chromosomes was significantly increased under the same conditions (Figure 6B). Thus, the effect of BMH-21 on chromatin accessibility is preferentially associated with the sex chromosomes. The results have been included in Figure 6B and described in the Results section (Lines 239-245).

eLife Assessment

This fundamental paper describes the role of the Pre-rRNA in meiotic sex chromosome inactivation in mouse spermatocytes. The cytological and genome-wide analyses on components of the nucleolus and chemical inhibition of RNA polymerase I for rDNA transcription provided solid evidence to support the authors' conclusion. The results shown in the paper are of interest to researchers in meiotic chromosome structure and the nucleolus.

Reviewer #1 (Public review):

The authors show that during prophase I of male meiosis, nucleoli disassemble and nucleolar components relocalize to the sex chromosome (XY) body. They further demonstrate that this process is regulated by the ATR-dependent signaling pathway that mediates meiotic sex chromosome inactivation (MSCI). Pharmacological disruption of pre-rRNA synthesis using the RNA polymerase I inhibitor BMH-21 leads to the recruitment of RNA polymerase II to the sex chromosomes and ectopic expression of sex chromosome-linked genes. These findings uncover a previously unrecognized role for pre-rRNAs in maintaining transcriptional silencing during meiosis. The study employs a combination of cell biology, genetics, and genomics approaches, and the conclusions are supported by compelling, well-organized data.

Comments on revised version.

I have read the revision and I have no additional questions.

eLife Assessment

This study presents an important examination of the role of cis-acting versus trans-acting genetic variation in DNA methylation divergence between humans and chimpanzees and its relationship to gene expression. By differentiating fused interspecies tetraploid cell lines into multiple cell types, the study provides compelling evidence that cis-acting changes predominate in DNA methylation divergence. The evidence linking directional methylation divergence to lineage-specific selection and, particularly, to adaptive human phenotypes is less clear. This work will be of interest to biologists and evolutionary anthropologists studying the evolution and genetics of gene regulation, particularly in primates.

Reviewer #1 (Public review):

[Editor's note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

Ma et al. use human-chimpanzee tetraploid cells, across different cell types, to identify the genetic causes and then transcriptomic consequences of divergence in DNA methylation. They conclude that the evolution of DNA methylation is driven primarily by cis-regulatory changes, and that the evolution of CpG sites contributes to cis-regulation while transcription factor expression underlies some trans changes. They then argue that divergence in DNA methylation is associated with changes in gene expression and may contribute to human phenotypes.

The tetraploid model is able to provide compelling evidence that most regulatory evolution occurs due to cis-regulatory changes, and that sites that have diverged in DNA methylation level cluster together into DMRs which share similar divergence patterns and genetic architecture. This stands in intriguing contrast to trans-mechanisms which explain most variation within modern human populations. The authors proceed to show that trans-effects can be explained, in part, by nearby TF binding motifs and many cis-regulatory changes are explained by CpG-disrupting variants. While these mechanisms clearly contribution to divergence in DNA methylation, the degree to which they explain cis- and trans-regulatory divergence would be valuable.

Next, the authors seek to show that differences in DNA methylation are functionally relevant. Consistent with previous results, they show that differences in DNA methylation are (weakly) associated with changes in gene expression. They hypothesize that genes with concordant regulatory should exhibit great methylation-expression coupling than other genes and show that cis-expression/cis-methylation pairs are more strongly correlated than trans/trans pairs. I think that looking at cis/trans or trans/cis changes would also be useful. Another limitation is that this analysis is limited to promoter regions. It is not clear how many divergent DMRs that includes and how many of those genes have differences in expression. The key question is whether differences in DNA methylation are functionally important, and the answer provided by these analyses is "sometimes".

Finally, the authors make a case for lineage-specific selection on DNA methylation that is connected to human traits. Using a sign-test, they identify sets of genes which have likely experienced selection in humans and may underlie human phenotypes.

In conclusion, I think this study provides a valuable resource for differences in DNA methylation between humans and chimpanzees across tissues and provides important insight into the relative abundance of cis and trans regulatory divergence. Additional research is necessary to investigate the underlying regulatory mechanisms and more care needs to be taken in exploring the functional consequences.

Reviewer #3 (Public review):

Summary:

Ma et al. use human-chimpanzee tetraploid cells to examine species differences in DNA methylation. They identify differentially methylated regions under cis or trans regulation. Cis-DMRs are enriched near SNVs that disrupt or create CpGs, providing a plausible mechanism for cis changes in methylation. They also seek to identify transcription factors that might affect methylation in trans, as well as gene sets with evidence for consistent changes in methylation and expression between humans and chimpanzees, suggesting that they may underlie lineage-specific traits.

Strengths:

The authors have generated a new dataset across multiple different cell types examining differences in DNA methylation between humans and chimpanzees using human diploid cells, chimpanzee diploid cells, and human-chimpanzee tetraploid cells. Using this dataset, they identify that cis-DMRs are enriched near SNVs that disrupt or create CpGs compared to trans-DMRs and identify transcription factors as candidate trans-acting factors. Both identified SNVs and transcription factors are good candidates for future experimentation. The authors also find that cis-DMRs are more highly correlated with cis-expressed genes than trans-DMRs with trans-expressed genes, providing evidence that methylation and expression are linked genome-wide. Further, they apply a clever sign test approach to identify gene pathways with evidence for lineage-specific selection.

Comments on revised version.

They authors have addressed my concerns from their initial manuscript. In particular, the authors have noted the limitations of their study as appropriate, including the use of permissive FDR cut-offs.

Author response:

Public Reviews:

Reviewer #1 (Public review):

Ma et al. use human-chimpanzee tetraploid cells, across different cell types, to identify the genetic causes and then transcriptomic consequences of divergence in DNA methylation. They conclude that the evolution of DNA methylation is driven primarily by cis-regulatory changes, and that the evolution of CpG sites contributes to cis-regulation while transcription factor expression underlies some trans changes. They then argue that divergence in DNA methylation is associated with changes in gene expression and may contribute to human phenotypes.

The tetraploid model is able to provide compelling evidence that most regulatory evolution occurs due to cis-regulatory changes, and that sites that have diverged in DNA methylation level cluster together into DMRs which share similar divergence patterns and genetic architecture. This stands in intriguing contrast to trans-mechanisms which explain most variation within modern human populations. The authors proceed to show that trans-effects can be explained, in part, by nearby TF binding motifs and many cis-regulatory changes are explained by CpG-disrupting variants. While these mechanisms clearly contribution to divergence in DNA methylation, the degree to which they explain cis- and trans-regulatory divergence would be valuable.

Next, the authors seek to show that differences in DNA methylation are functionally relevant. Consistent with previous results, they show that differences in DNA methylation are (weakly) associated with changes in gene expression. They hypothesize that genes with concordant regulatory should exhibit great methylation-expression coupling than other genes and show that cis-expression/cis-methylation pairs are more strongly correlated than trans/trans pairs. I think that looking at cis/trans or trans/cis changes would also be useful. Another limitation is that this analysis is limited to promoter regions. It is not clear how many divergent DMRs that includes and how many of those genes have differences in expression. The key question is whether differences in DNA methylation are functionally important, and the answer provided by these analyses is "sometimes".

Finally, the authors make a case for lineage-specific selection on DNA methylation that is connected to human traits. Using a sign-test, they identify sets of genes which have likely experienced selection in humans and may underlie human phenotypes.

In conclusion, I think this study provides a valuable resource for differences in DNA methylation between humans and chimpanzees across tissues and provides important insight into the relative abundance of cis and trans regulatory divergence. Additional research is necessary to investigate the underlying regulatory mechanisms and more care needs to be taken in exploring the functional consequences.

Reviewer #3 (Public review):

Summary:

Ma et al. use human-chimpanzee tetraploid cells to examine species differences in DNA methylation. They identify differentially methylated regions under cis or trans regulation. Cis-DMRs are enriched near SNVs that disrupt or create CpGs, providing a plausible mechanism for cis changes in methylation. They also seek to identify transcription factors that might affect methylation in trans, as well as gene sets with evidence for consistent changes in methylation and expression between humans and chimpanzees, suggesting that they may underlie lineage-specific traits.

Strengths:

The authors have generated a new dataset across multiple different cell types examining differences in DNA methylation between humans and chimpanzees using human diploid cells, chimpanzee diploid cells, and human-chimpanzee tetraploid cells. Using this dataset, they identify that cis-DMRs are enriched near SNVs that disrupt or create CpGs compared to trans-DMRs and identify transcription factors as candidate trans-acting factors. Both identified SNVs and transcription factors are good candidates for future experimentation. The authors also find that cis-DMRs are more highly correlated with cis-expressed genes than trans-DMRs with trans-expressed genes, providing evidence that methylation and expression are linked genome-wide. Further, they apply a clever sign test approach to identify gene pathways with evidence for lineage-specific selection.

Comments on revised version.

They authors have addressed my concerns from their initial manuscript. In particular, the authors have noted the limitations of their study as appropriate, including the use of permissive FDR cut-offs.

We thank the Reviewing Editor, Senior Editor, and reviewers for their careful reading of our manuscript and their constructive feedback.' We have made two changes in response to the 2nd round of reviews.

We have replaced the paragraph “To identify the molecular mechanisms driving methylation divergence, we analyzed transcription factors (TFs) associated with trans-DMRs and sequence variants associated with cis-DMRs” with “To explore potential molecular mechanisms driving methylation divergence, we focused on two of many possible contributors: transcription factors (TFs) associated with trans-DMRs and sequence variants associated with cis-DMRs.”

We now report the absolute numbers of sites underlying each proportion: at FDR < 0.05, 663,368 cis-regulated and 140,017 trans-regulated sites out of 932,573 divergent sites, among 24,935,901 CpG sites tested across cell types. Only CpG sites covered in all parental and hybrid samples of a cell type can be tested, so this is fewer than the 49,049,052 unique CpG positions covered by any library, and we have clarified this distinction in the Results. Per-cell-type counts are provided in Supplemental File 1.

Second, we performed a variance partitioning analysis analogous to heritability partitioning. We decomposed total interspecies divergence (parental human minus chimpanzee) into a cis component (the allelic difference within hybrids) and a trans component (the remainder), and estimated the proportion of variance in divergence explained by each, together with their covariance. To limit sampling noise, we pooled read counts within 1-kb genomic bins. We used noise-corrected estimators based on covariances between independent replicates. For cell types with a single parental line per species, we report bounds, whose validity we confirmed in CNCC, where two parental replicates allow a direct replicate-based estimate.

In CNCC, cis effects explained 93% (95% CI 92–95%) of the variance in divergence and trans effects 71% (70–73%), with a strongly negative covariance term (−0.64). This indicates that cis and trans changes often act in opposite directions. In the other cell types, cis effects explained 58–98% and trans effects 3–48% of the variance, depending on cell type and bound, with small covariance terms. Cis exceeded trans in every cell type under both bounds, and the results were robust to a stricter coverage threshold and consistent at single-CpG resolution.

These analyses confirm that cis-regulation is the dominant contributor to human–chimpanzee methylation divergence. They also show that trans effects contribute more to divergence variance than the proportions of classified sites alone would suggest. We have added these results to the Results, described the method in a new Methods subsection, "Variance partitioning of methylation divergence”, and provided full results in Supplemental File 1.

eLife Assessment

The authors established a novel three-dimensional culture system for stratified epithelia that allows epithelial cells to undergo epithelial-to-mesenchymal transition (EMT) and subsequent mesenchymal-to-epithelial transition (MET) while migrating through a membrane with 3.0-µm micropores, and, thus, provides a highly valuable tool to study EMT and possibly wound regeneration or metastasis. Furthermore, a set of experiments provides compelling data suggesting that TGF beta signaling and actin polymerization promote movement of epithelial cells into the pores, while Piezo1 and Keratin 6 prevent keratinocyte migration and EMT.

Reviewer #1 (Public review):

[Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

Summary:

The study describes the migration of epidermal keratinocytes through porous membranes and observed a unique size selection whereby only on 3-micron membrane are keratinocytes able to migrate and reform an intact epidermis. The authors propose that the model replicates three cell states of intact epidermis, EMT, and MET. They also show that this response depends on the actin cytoskeleton, keratin 6, and Piezo1, and the migration could be stimulated with TGFbeta ligands.

Strengths:

Strengths of the study include the establishment of a simple yet robust in vitro model that capture all three cell states, which could be useful for future investigation of wound healing or metastasis across a range of epithelia. There is also good quantitative characterisation of the pore size effects, providing some interesting observations about the physical regulation of keratinocyte migration, and the study provides some initial insights into the regulatory mechanisms. The images and presentation are clear.

Reviewer #2 (Public review):

Summary:

The revised manuscript by Nohara et al. presents a novel 3D assay that allows for stratification of epithelia, active EMT through small pores, and active MET. They show that 3um pores allow for keratinocytes to sample the pore through filopodia and up regulate EMT genes to transverse the pores to the other side of the membrane where EMT genes are down regulated as the cells re-establish stratified epithelia. The TGFbeta pathway and actin polymerization promotes movement of cells into the pores and Piezo1 and KRT6 actively block this movement. This work provides a novel 3D assay that is likely to become a benchmark to analyze these processes using a more complex system than other current culture-specific EMT and MET assays.

Strengths:

The strengths of the manuscript include the foundational analysis of the pathways involved in establishing the 3-compartment epithelium. The authors have incorporated live imaging, drug studies, KO analysis, and RNA-sequencing to show the relevant pathways involved. The authors were responsive to the original round of reviews and present a much-improved study with greater details on molecular markers, quantification, and clarity.

Reviewer #3 (Public review):

Summary:

The authors established an experimental system that reproduced three-dimensional triphasic epithelia, i.e., the original epithelium, its EMT, and MET. Keratinocytes (KCs), skin epithelial cells, placed on a microporous membrane migrated through 3.0-um or larger micropores. The 3.0-um-pored membrane induced an epithelial structure with three states: stratified KCs above the membrane, KCs showing EMT within the micropores, and a new stratified epithelium under the membrane. The membrane with larger micropores failed to maintain this triphasic epithelia. Live imaging revealed that KCs moved in a reciprocating manner, with actin-rich filopodia-like KC structures extending into and out of the 3.0-um micropores, while the cells migrated unidirectionally into larger micropores. KO of Piezo1 and keratin 6 increased KC entry to and exit from the 3.0-um micropores. Their results demonstrate that benign keratinocytes migrate through confined spaces in a reciprocating manner, which might help form triphasic epithelia, recapitulating wound healing processes.

Strengths:

Careful observation of the behaviour of keratinocytes on the different sized pores. CrispR-Cas9 gene editing to KO Piezo 1 and keratin 6 isoforms in HaCaT keratinocytes.

Author response:

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

Changes made independently of the reviewers' comments

In addition to the revisions described below, we have made the following three changes, none of which affects the conclusions of the paper.

(1) Shared control data (Figure 4c, d, l, m). The parental cells and the PIEZO1 KO and pan-KRT6 KO cell lines were assayed in parallel in the same experiments. We have therefore stated in the figure legend that the parental-cell data are shared between panels c, d and panels l, m.

(2) Number of DEGs (Figure 4h). We have corrected the number of genes upregulated in cells cultured on the 3.0-µm-pored membrane from 69 to 68.

(3) Yoda1 data (Figure 4e, f). Some values had been incorrectly transferred during data compilation. We have corrected these values and the corresponding graphs. The statistical results are unchanged.

We have also provided two new files: Source data 1 (numerical data underlying all graphs) and Source data 2 (uncropped western blot images).

Public Reviews:

Reviewer #1 (Public review):

Summary:

The study describes the migration of epidermal keratinocytes through porous membranes and observed a unique size selection whereby only on 3-micron membrane are keratinocytes able to migrate and reform an intact epidermis. The authors propose that the model replicates three cell states of intact epidermis, EMT, and MET. They also show that this response depends on the actin cytoskeleton, keratin 6, and Piezo1, and the migration could be stimulated with TGFbeta ligands.

Strengths:

Strengths of the study include the establishment of a simple yet robust in vitro model that capture all three cell states, which could be useful for future investigation of wound healing or metastasis across a range of epithelia. There is also good quantitative characterisation of the pore size effects, providing some interesting observations about the physical regulation of keratinocyte migration, and the study provides some initial insights into the regulatory mechanisms. The images and presentation are clear.

Weaknesses:

While there are some insights into the underlying mechanisms, there could be more detailed analysis of the key signalling pathways and functional relationships between actin, keratins, Piezo, and EMT genes. Although the extent of new biological insight is somewhat limited, the study still demonstrates the utility of the model, and these points may be important questions to address in future studies.

We agree with the reviewer that these are important questions. We will address the functional relationships among actin, keratins, PIEZO1, and EMT genes, and the underlying signaling pathways, in future studies.

Reviewer #2 (Public review):

Summary:

The revised manuscript by Nohara et al. presents a novel 3D assay that allows for stratification of epithelia, active EMT through small pores, and active MET. They show that 3um pores allow for keratinocytes to sample the pore through filopodia and up regulate EMT genes to transverse the pores to the other side of the membrane where EMT genes are down regulated as the cells re-establish stratified epithelia. The TGFbeta pathway and actin polymerization promotes movement of cells into the pores and Piezo1 and KRT6 actively block this movement. This work provides a novel 3D assay that is likely to become a benchmark to analyze these processes using a more complex system than other current culture-specific EMT and MET assays.

Strengths:

The strengths of the manuscript include the foundational analysis of the pathways involved in establishing the 3-compartment epithelium. The authors have incorporated live imaging, drug studies, KO analysis, and RNA-sequencing to show the relevant pathways involved. The authors were responsive to the original round of reviews and present a much-improved study with greater details on molecular markers, quantification, and clarity.

We thank the reviewer for the positive evaluation of our revised manuscript.

Reviewer #3 (Public review):

Summary:

The authors established an experimental system that reproduced three-dimensional triphasic epithelia, i.e., the original epithelium, its EMT, and MET. Keratinocytes (KCs), skin epithelial cells, placed on a microporous membrane migrated through 3.0-um or larger micropores. The 3.0-um-pored membrane induced an epithelial structure with three states: stratified KCs above the membrane, KCs showing EMT within the micropores, and a new stratified epithelium under the membrane. The membrane with larger micropores failed to maintain this triphasic epithelia. Live imaging revealed that KCs moved in a reciprocating manner, with actin-rich filopodia-like KC structures extending into and out of the 3.0-um micropores, while the cells migrated unidirectionally into larger micropores. KO of Piezo1 and keratin 6 increased KC entry to and exit from the 3.0-um micropores. Their results demonstrate that benign keratinocytes migrate through confined spaces in a reciprocating manner, which might help form triphasic epithelia, recapitulating wound healing processes.

Strengths:

Careful observation of the behaviour of keratinocytes on the different sized pores. CrispR-Cas9 gene editing to KO Piezo 1 and keratin 6 isoforms in HaCaT keratinocytes.

Weaknesses:

The authors have improved the observed weaknesses including better statistical analysis and more detail on replicates. They have also performed staining for laminin beta3 chain.

We thank the reviewer for these comments.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

The authors have significantly improved the manuscript in the recent revision. One minor point is to please make the RNA-seq analysis, including gene lists available in the supplementary information.

As suggested, we have added Figure 4—source data 1, which lists the differentially expressed genes (DEGs) identified by bulk RNA-seq of HaCaT keratinocytes cultured on 0.4-µm- or 3.0-µm-pored membranes.

Reviewer #3 (Recommendations for the authors):

I am not suggesting any further changes to the manuscript.

We thank the reviewer for evaluating our revised manuscript.

eLife Assessment

This important study describes computationally designed proteins that bind to the chemokine CCL25. The authors present evidence that some binders simply prevent chemokine binding to the CCR9 receptor, while one binder changes the downstream signaling triggered by chemokine binding. The evidence is convincing, showing a degree of selectivity between chemokine ligands that could be further improved. The study will be of interest to pharmacologists, cell biologists studying chemokine function, and protein engineers.

Reviewer #1 (Public review):

Summary:

In this manuscript, the authors describe the use of BindCraft computational protein design to create a series of binders to the chemokine CCL25. This chemokine normally mediates CCR9-dependent trafficking of immune cells to the gut, making it a potential target for the treatment of inflammatory bowel disease and related conditions. Importantly, CCL25 also binds a scavenging receptor, ACKR4. The computational protein design approach used does not involve defining particular epitopes to be targeted, allowing a free search for any potential interaction surface.

Among four designs tested, three were predicted to interact at a similar site on the chemokine, while a fourth clone, VUP25111, was predicted to bind to a different site. All four designs showed binding to CCL25, with similar high-nM KD values in all cases. The first three clones showed evidence of direct competition with the receptor for CCL25 binding, while VUP25111 showed incomplete inhibition of binding. In functional assays, all clones acted as antagonists except for VUP25111, which inhibited arrestin recruitment by CCR9, but did not affect G protein activation by CCR9 or arrestin recruitment by ACKR4 (which signals exclusively through arrestin and not G protein).

Strengths:

The work is completed to a high technical standard, and the functional diversity of the clones is intriguing. It is exciting to see pathway-selective modulation of signaling, and this basic paradigm is likely to generalize to other chemokine/receptor systems. The exceptional complexity of chemokine signaling makes this an excellent area to explore the development of modulators that can restrict signaling to a specific subset of receptors.

Weaknesses:

No major weaknesses were noted by this reviewer.

Comments on revised version:

The revision looks fine to me.

Reviewer #2 (Public review):

This study from de Boer, Lamme, Verdwaald and Schafer describes the de novo AI-guided design of miniproteins that target the chemokine CCL25, with the aim to modulate the activation and signalling of the chemokine receptors CCR9 and ACKR4. The study focuses on characterising four miniproteins that all bind CCL25 with good affinity. Three designs appear to prevent CCL25 binding to both CCR9 and ACKR4, with increasing concentrations of miniproteins resulting in decreased arrestin (both receptors) and mini G protein recruitment (CCR9), less inhibition of forskolin-stimulated cAMP (CCR9), and decreased GRK3 recruitment and receptor internalisation (CCR9). One miniprotein, VUP25111, changes the properties of CCL25 rather than preventing ligand/receptor interactions, resulting in greater selectivity for CCR9 over ACKR4 and a G protein-biased signalling profile (maintenance of mini G protein recruitment, GRK3 recruitment, inhibition of cAMP and receptor internalisation, but loss of arrestin recruitment). VUP25111 also maintained chemotactic migration in MOLT-4 T lymphoblast cells, whereas this response was lost in the presence of the other three miniproteins.

Overall, this is a very interesting application of AI-designed de novo miniproteins to modulate GPCR responses by directly binding the ligand rather than the receptor. This is a conceptually very intriguing approach that could, in principle, be extended to other GPCR systems beyond the chemokine family. The authors deploy an impressive array of assays spanning multiple signalling endpoints, providing a thorough picture of how each miniprotein influences receptor activation and downstream signalling. The presentation of concentration-response relationships for CCL25 alone and in the presence of each miniprotein is particularly informative, and the figures are very well constructed throughout. The inclusion of clear cartoons illustrating the basis of each assay is a nice touch that will help readers from outside the immediate field follow the logic of each experiment.

Reviewer #3 (Public review):

The authors employed the BindCraft platform to develop binders targeting the chemokine CCL25, a selective activator of the chemokine receptor CCR9. They successfully generated two classes of binders: one that inhibits all CCL25-mediated CCR9 activation, and another that permits CCR9 G protein signaling while simultaneously preventing arrestin recruitment. These tools will enable the dissection of arrestin involvement in regulating cell migration.

Author response:

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

Reviewer #1 (Public review):

Summary:

In this manuscript, the authors describe the use of BindCraft computational protein design to create a series of binders to the chemokine CCL25. This chemokine normally mediates CCR9-dependent trafficking of immune cells to the gut, making it a potential target for the treatment of inflammatory bowel disease and related conditions. Importantly, CCL25 also binds a scavenging receptor, ACKR4. The computational protein design approach used does not involve defining particular epitopes to be targeted, allowing a free search for any potential interaction surface.

Among four designs tested, three were predicted to interact at a similar site on the chemokine, while a fourth clone, VUP25111, was predicted to bind to a different site. All four designs showed binding to CCL25, with similar high-nM KD values in all cases. The first three clones showed evidence of direct competition with the receptor for CCL25 binding, while VUP25111 showed incomplete inhibition of binding. In functional assays, all clones acted as antagonists except for VUP25111, which inhibited arrestin recruitment by CCR9, but did not affect G protein activation by CCR9 or arrestin recruitment by ACKR4 (which signals exclusively through arrestin and not G protein).

Strengths:

The work is completed to a high technical standard, and the functional diversity of the clones is intriguing. It is exciting to see pathway-selective modulation of signaling, and this basic paradigm is likely to generalize to other chemokine/receptor systems. The exceptional complexity of chemokine signaling makes this an excellent area to explore the development of modulators that can restrict signaling to a specific subset of receptors.

Weaknesses:

No major weaknesses were noted by this reviewer.

We thank the reviewer for their positive response and thorough evaluation of our work. Below, we address their recommendations in detail in the private review.

Reviewer #1 (Recommendations for the authors):

Minor points for potential improvement:

(1) Consider shortening the discussion section, which could benefit from a more focused discussion of the key conclusions and their implications. As written, it is rather long, and parts of it repeat material already covered elsewhere (e.g., the two-step binding mechanism is described in lines 55-64 and again in 331-335).

We thank the reviewer for the suggestion to make the discussion more focused. We agree that parts repeat material from the introduction. As such, we have revised the discussion to reduce repetition with the introduction, including removal of the repeated description of the two-step binding mechanism, and condensing several passages.

(2) The figures are clear, but it may be helpful to combine panels into a smaller number of figures to make direct comparisons among signaling curves more straightforward.

We thank the reviewer for the suggestion to reduce the number of figures for easier curve comparisons. We prefer our current presentation of the figures so as not to conflate disparate themes and concepts; for example, GRK recruitment is a distinct event from receptor internalization. The pharmacological results have been compiled in Figure S8 in the supplementary to allow visual, side-by-side comparisons of the DRCs presented in the publication.

Reviewer #2 (Public review):

This study from de Boer, Lamme, Verdwaald and Schafer describes the de novo AI-guided design of miniproteins that target the chemokine CCL25, with the aim to modulate the activation and signalling of the chemokine receptors CCR9 and ACKR4. The study focuses on characterising four miniproteins that all bind CCL25 with good affinity. Three designs appear to prevent CCL25 binding to both CCR9 and ACKR4, with increasing concentrations of miniproteins resulting in decreased arrestin (both receptors) and mini G protein recruitment (CCR9), less inhibition of forskolin-stimulated cAMP (CCR9), and decreased GRK3 recruitment and receptor internalisation (CCR9). One miniprotein, VUP25111, changes the properties of CCL25 rather than preventing ligand/receptor interactions, resulting in greater selectivity for CCR9 over ACKR4 and a G protein-biased signalling profile (maintenance of mini G protein recruitment, GRK3 recruitment, inhibition of cAMP and receptor internalisation, but loss of arrestin recruitment). VUP25111 also maintained chemotactic migration in MOLT-4 T lymphoblast cells, whereas this response was lost in the presence of the other three miniproteins.

Overall, this is a very interesting application of AI-designed de novo miniproteins to modulate GPCR responses by directly binding the ligand rather than the receptor. This is a conceptually very intriguing approach that could, in principle, be extended to other GPCR systems beyond the chemokine family. The authors deploy an impressive array of assays spanning multiple signalling endpoints, providing a thorough picture of how each miniprotein influences receptor activation and downstream signalling. The presentation of concentration-response relationships for CCL25 alone and in the presence of each miniprotein is particularly informative, and the figures are very well constructed throughout. The inclusion of clear cartoons illustrating the basis of each assay is a nice touch that will help readers from outside the immediate field follow the logic of each experiment.

There are two main conclusions that are not currently as well-supported by the evidence as they might be, and that would benefit from some qualification. The first concerns the selectivity of the miniproteins for CCL25. Testing the impact of the miniproteins on CXCL12 activation of CXCR4 is an important and welcome experiment, but it addresses selectivity against only one other chemokine system, and the current claim of specificity is therefore stronger than the data allow. Additionally, at the highest concentration tested (10 µM), the more potent miniproteins (VUP25101, VUP25107) appear to show some inhibition of arrestin recruitment to CXCR4 - perhaps unsurprising given the degree of structural conservation among chemokines. The statements regarding selectivity and the lack of effect on the CXCL12/CXCR4 system would benefit from revision to more accurately reflect these observations.

We thank the reviewer for the detailed review and suggestions for our publication. We agree with the reviewer that testing against only one additional chemokine receptor: chemokine pair is insufficient to support strong claims of selectivity. To further assess the selectivity of the miniproteins, we evaluated the effects of the VUPs against CCL20, a chemokine closely related to CCL25. CCL20 binds both CCR6 and ACKR4, providing an optimal platform to compare the effects on this non-specified target. Receptor activation was determined by β-arrestin recruitment (Fig. S6). For ACKR4, VUP25107 and VUP25112 have similar IC50 values for CCL20 and CCL25. In contrast, VUP25101 showed reduced potency towards CCL20 relative to CCL25 (CCL20 pIC50 = 6.06 ± 0.13, CCL25 pIC50 = 7.73 ± 0.08), suggesting greater selectivity. VUP25111 had no impact on activation of ACKR4 by CCL20, which may reflect a lack of engagement with the other chemokine or that it binds in a similar way as to CCL25 and does not impact chemokine-promoted ACKR4 activation. CCL20-mediated activation of CCR6 further supports the specificity of VUP25101, which showed only a modest effect at the highest concentration tested. In contrast, VUP25107 and VUP25112 showed a dose-dependent inhibition of arrestin recruitment (pIC50_VUP25107 = 6.99 ± 0.15, pIC50_VUP25112 = 7.49 ± 0.07).

We agree that a reduction in β-arrestin recruitment to CXCR4 is apparent at the highest concentration tested (10 µM) for VUP25101 and VUP25107. To further address the selectivity of these miniproteins, we have included β-arrestin recruitment data for ACKR3 (Fig. S5), the other chemokine receptor that binds CXCL12. A similar suppression was observed for the atypical receptor as for CXCR4 at the highest concentrations, suggesting cross-reactivity at high concentrations.

Based on the CCL20 and CXCL12 data, we have revised the text to more accurately reflect these observations. We now frame these results as an initial proof of concept demonstrating the feasibility of this design approach, while noting that further optimization and characterization of selectivity will be required to advance the miniproteins for additional applications.

To reflect this, we have added the following text on Pg.9:

“CCL25 specificity was experimentally evaluated by testing the miniproteins against CXCL12 by tracking arrestin recruitment to CXCR4 and ACKR3, the native agonist for this canonical and atypical receptor pair (Fig. S5). VUP25111 and VUP25112 showed no effect on CXCL12-mediated activation at any concentration, while VUP25101 and VUP25107 showed partial inhibition at the highest tested concentration (10 µM).

In addition, the effects of the VUPs were evaluated against CCL20, a closely related chemokine to CCL25, by assessing arrestin recruitment to CCR6 and ACKR4 (Fig. S6). VUP25107 and VUP25112 displayed similar inhibitory potencies of CCL20 as to CCL25 for ACKR4 (VUP25107: pIC50 = 7.72 ± 0.21; VUP25112: pIC50 = 6.93 ± 1.03). In contrast, VUP25101 showed reduced inhibition of CCL20 relative to CCL25 (VUP25101: pIC50_CCL20 = 6.06 ± 0.13 v pIC50_CCL25 = 7.73 ± 0.08), indicating greater selectivity for the designed target. CCL20 activation of CCR6 further supports the specificity of VUP25101, which only showed a modest effect at the highest concentration tested. VUP25111 showed no inhibitory effects on CCL20 activation of either CCR6 nor ACKR4, which may indicate a lack of binding, but is also indiscernible from forming a receptor-competent complex similar to the proposed CCL25:VUP25111:ACKR4 interaction.”

And added the following in the discussion on Pg.15:

“The miniproteins displayed varying degrees of chemokine selectivity among CCL25, CCL20, and CXCL12, including cross-reactivity for some designs. Given that only two additional chemokines were examined, these experiments provide an initial assessment of selectivity. As chemokine selectivity was not explicitly optimized during design, broader experimental profiling and negative-design strategies may be required to establish and improve miniprotein selectivity.”

The second concern relates to the interpretation of the preserved GRK3 recruitment, but the complete loss of arrestin recruitment observed with VUP25111. In the GRK3 recruitment experiments, 20 nM CCL25 was used, representing an EC40 concentration in this assay. VUP25111 causes a concentration-dependent reduction in CCL25-induced GRK3 recruitment, down to approximately 15% of the maximal response. It is worth considering whether this degree of reduction in GRK3 recruitment could itself be sufficient to disrupt patterns of receptor phosphorylation and thereby prevent observable arrestin recruitment. Both interpretations are complicated by the fact that the GRK3 recruitment and arrestin recruitment assays likely differ in their sensitivity and dynamic windows, making direct quantitative comparisons between them difficult. In the absence of direct measurements of CCR9 phosphorylation in the presence of VUP25111, the alternative interpretation remains open and would benefit from acknowledgement. Given recent work from the same group demonstrating that receptor internalisation is only partially dependent on arrestins (Lamme et al., 2025, J Biol Chem), further discussion of the relationship between GRK and arrestin recruitment and CCR9 internalisation would be of value to the broader GPCR audience this work is likely to attract.

We appreciate the reviewer’s careful analysis of our results and highlighting earlier work from our group. The reviewer is correct that the discrepancy between VUP25111’s impact on GRK3 and arrestin recruitment can also be explained by the decrease in GRK3 recruitment reflecting a sufficient change in receptor phosphorylation to account for the loss of arrestin engagement and not solely by the presented interpretation of the adoption of an arrestin-incompatible conformation with retained phosphorylation. The main text has been modified on pg. 11 to reflect this potential:

“This suggests that either activation of CCR9 by the VUP25111:CCL25 complex promotes a conformation incompatible with arrestin engagement, but still accessible for GRK phosphorylation, or the reduced GRK interaction represents the formation of a phosphorylation barcode incapable of supporting arrestin recruitment. Without direct phosphorylation measurements these possibilities cannot be separated.”

And in the discussion on pg.14:

“The extent of phosphorylation is not clear, and the inhibited GRK recruitment may reflect the formation of sub-phosphorylated states which cannot support arrestin recruitment nor promote efficient receptor internalization. Alternatively, the miniprotein: chemokine complex may induce a conformational state incompatible with arrestin engagement.”

A partial phosphorylation explanation has also been added on internalization on pg.12:

“These results suggest that the VUP25111:CCL25 complex suppresses both the arrestin-dependent and independent internalization mechanisms used by CCR9, despite likely allowing for some, albeit perhaps insufficient, GRK phosphorylation.”

Finally, some additional justification for the use of 20 nM CCL25 across all assays would strengthen the study, as this concentration represents different points on the concentration-response curve depending on the assay and receptor in question. It ranges from an EC40 for CCR9 GRK3 recruitment and internalisation, to an EC50 for CCR9 arrestin and mini-Gi recruitment, an EC80 for CCR9 cAMP inhibition, and an Emax for ACKR4 arrestin recruitment. This has potential consequences for the interpretation and cross-assay comparison of miniprotein potency, and the authors are encouraged to acknowledge and discuss this in the context of their conclusions.

We thank the reviewer for highlighting the need to discuss CCL25 concentration across assays. Our primary aim was to use a single CCL25 concentration to enable direct comparison of all assay outcomes against the same CCL25:miniproteins ratio, since the miniproteins are designed to antagonize the chemokine, rather than the receptor. We set the concentration at 20 nM based on the first arrestin recruitment assays to CCR9, where we initially characterized miniprotein: CCL25 activity. Importantly, CCL25 has lower potency for CCR9 than ACKR4, making it the limiting case. Selecting 20 nM ensured responses within the EC40-EC80 range for all assays. To improve clarity, we have now explicitly stated this rationale in the manuscript and noted the implications of using a single CCL25 concentration across diverse readouts on Pg.10:

“Despite the higher potency of CCL25 for cAMP inhibition compared to arrestin recruitment, a fixed concentration of 20 nM CCL25 was used throughout this and subsequent assays to maintain a consistent CCL25:miniprotein stoichiometry. As this concentration represents different points on the respective concentration-response curves, cross-assay comparisons of apparent miniprotein potency should be interpreted qualitatively.”

Reviewer #2 (Recommendations for the authors):

Major points:

(1) The conclusions regarding miniprotein selectivity for CCL25 would benefit from revision to reflect that only the CXCL12/CXCR4 system was tested as a counter-screen. The apparent inhibition of arrestin recruitment to CXCR4 by VUP25101 and VUP25107 at 10 µM should be acknowledged/discussed.

We thank the reviewer for this comment and agree that selectivity should be more explicitly acknowledged. This point has been thoroughly addressed in our response in the public review. Briefly, we have added a similar selectivity assay but now against CCL20, which is closely related to CCL25, as well as revisiting the CXCL12 assay with ACKR3 that was originally presented in the manuscript. Statements in the text have been added to note the apparent inhibition at 10 µM for VUP25101 and VUP25107.

(2) The conclusions regarding G protein-biased signalling driven by VUP25111 would benefit from qualification. In particular, it is worth acknowledging that a substantial reduction in GRK3 recruitment may itself be sufficient to account for the observed loss of arrestin recruitment, and that direct measurement of CCR9 phosphorylation in the presence of VUP25111 would help to resolve this. Discussion of the relationship between GRK recruitment, receptor phosphorylation, arrestin recruitment and internalisation in this context would also be of value to the broader GPCR audience this work is likely to attract.

We agree with the reviewer that an incomplete or insufficient level of phosphorylation could also explain our arrestin and internalization results presented by the VUP25111:CCL25 complex. We have added the following to the discussion on pg. 14 and modified the results to better reflect the multiple interpretations:

“The extent of phosphorylation is not clear and the inhibited GRK recruitment may reflect the formation of sub-phosphorylated states which cannot support arrestin recruitment, nor promote efficient receptor internalization. Alternatively, the complete inhibition of arrestin engagement cannot be easily explained by phosphorylation, and instead suggests the miniprotein: chemokine complex may induce a conformational state incompatible with arrestin engagement. Due to the clash with ECL2, the loop would need to distort to accommodate the larger ligand complex and this displacement may translate into the G protein and GRK-permissive structural state.”

(3) The choice of 20 nM CCL25 across assays should be justified or its implications discussed, particularly given that this concentration represents different effective concentrations for each endpoint and receptor.

The rationale for using a fixed 20nM CCL25 concentration and its implications is now discussed within the main text as well as our response in the public review.

Modelling:

Modelling of VUP25111 and VUP25101 binding to CCL25/receptor complexes is presented. Was the same analysis performed for VUP25112 and VUP25107, and do they similarly appear to interfere with direct receptor binding?

We initially selected VUP25101 as a representative for VUP25112 and VUP25107 to simplify visualization. We have now extended the same modeling analysis to VUP25107 and VUP25112. These additional analyses have now been included in the revised manuscript in Fig. 3D.

Methods:

(1) Percentage concentrations throughout the methods are missing w/v or v/v designations.

We thank the reviewer for their careful reading and apologize for the oversight. We have added all w/v or v/v designations in the methods.

(2) More specific information for each BRET experiment would be useful. For example, incubation times with CCL25 and miniproteins (included for some but not all assays), and clarification of whether experiments were endpoint reads or timecourse experiments.

We agree with the reviewer that the methods section could be clarified, as well as removing a lot of repetition. We have revised the methods.

Specific text revisions:

(1) Abstract, line 20: "Most drug discovery efforts target the receptors, whereas natural regulatory mechanisms focus on the chemokines". This is unclear for a non-chemokine audience without additional context clarifying that this refers to binding to or modification of the chemokine itself to modulate receptor activation.

To clarify this distinction for a broader readership, we revised the abstract to state explicitly that natural regulatory mechanisms modulate immune responses by directly binding to or modifying chemokines, thereby altering their ability to bind receptors. We have changed the text in the abstract on Pg.2 to:

“Most drug discovery efforts target the receptors, whereas natural regulatory mechanisms modulate immune responses by directly binding to or modifying the chemokine ligands.”

(2) Results, line 177: "In the case of VUP25111, only the highest tested concentrations decreased CCL25 binding to either receptor, but did not completely eliminate the binding signal". This should be revised as there is a clear concentration-dependent effect of VUP25111 on CCL25 binding.

We have revised the text to reflect that VUP25111 inhibits CCL25 binding in a concentration-dependent manner, while also making clear that this inhibition occurs with significantly lower potency than the other miniproteins. We have changed the following text on Pg.7:

“VUP25111 shows a drastically lower inhibitory potency compared to the other binders and did not completely eliminate the binding signal.”

(3) Line 179: "...this suggests that CCL25:VUP25111 may bind the receptors cooperatively". Please revise; the meaning is unclear.

Our results indicate that VUP25111 does not prevent CCL25 from engaging either receptor. To avoid ambiguity, we have revised the text to explicitly state that our results are consistent with the binding of a pre-formed VUP25111:CCL25 complex to the receptor, resulting in the formation of a VUP25111:CCL25:receptor complex. We have changed the following text on Pg.7:

“Since the binder readily complexes with the chemokine at these concentrations (Fig. 2), this suggests that the CCL25:VUP25111 complex retains the ability to bind the receptor.”

(4) Line 216: "VUP25112 also completely impaired CCL25-mediated activation of the atypical receptor at the highest concentrations, but with significantly less potency". Please provide statistical tests to support this statement.

We thank the reviewer for requesting clarification regarding the statistical support for this statement. Statistical analyses supporting the conclusions in line 216, as well as other figures, were included in the initial manuscript but were provided in the supplementary information. To improve clarity in the manuscript, we have added explicit references to the corresponding supplementary statistical table in the text, added appropriate significance indicators to the relevant figures, and expanded the figure legends to state the statistical test used.

(5) Line 219: "The unique binder VUP25111, on the other hand, did not inhibit the activation of ACKR4 by CCL25 at any tested concentration". Please rephrase. While there was no concentration-dependent inhibition of arrestin recruitment to ACKR4, the addition of VUP25111 at any concentration appeared to reduce total arrestin recruitment to approximately 80% of the CCL25-alone condition.

We would like to thank the reviewer for this careful analysis of the data. Although this may appear to be the case in the figure, there is no significant difference between the lowest concentration (2.5 nM) and vehicle condition (∞) for any miniprotein and no real difference can be claimed.

We have specifically noted this on Pg. 9:

“The unique binder VUP25111, on the other hand, did not significantly impact the activation of ACKR4 by CCL25 at any tested miniprotein concentration.”

(6) Line 240: "suggest that the specific ligand interactions driving arrestin engagement with CCR9 may be more strict than for the G proteins". An alternative interpretation is that the miniprotein specifically impairs the receptor conformation important for arrestin engagement, without affecting the conformation required for G protein engagement, and this should be acknowledged.

We agree with the reviewer that these statements would benefit from further clarification. Indeed, the observed effects may not be only due to specific ligand interactions but could also depend on specific receptor conformations required for the specific protein-protein interactions. We have adjusted the following text on Pg.10:

“The deviations between the arrestin results (Fig. 4) and G protein (Fig. 5B) suggest that the specific ligand interactions or receptor conformations driving arrestin engagement with CCR9 may be more strict than for the G proteins.”

Reviewer #3 (Public review):

Summary:

The authors employed the BindCraft platform to develop binders targeting the chemokine CCL25, a selective activator of the chemokine receptor CCR9. They successfully generated two classes of binders: one that inhibits all CCL25-mediated CCR9 activation, and another that permits CCR9 G protein signaling while simultaneously preventing arrestin recruitment. These tools will enable the dissection of arrestin involvement in regulating cell migration.

My comments, in the order of reading:

(1) Title: I strongly recommend removing the term "biasing" from the title. In this context, it does not convey a specific mechanistic concept. The term "biased signaling" has been used for a very broad range of mechanistically distinct pharmacological phenomena, and without a precise definition, it adds more confusion than clarity. I therefore suggest refraining from using it in the title.

We appreciate the reviewer’s concern regarding the use of the term “biasing” given its broad and sometimes ambiguous use in the literature. Most commonly it refers to G protein versus arrestin-mediated signalling but is also used in regard to GRK specificity as well as to discrete differences in receptor trafficking (Motso et al. 2025, Cell). In our study, we use this in a specific mechanistic sense, referring to the bias VUP25111 induces upon binding CCL25. The miniprotein: CCL25 complex inhibits CCR9 arrestin recruitment, receptor internalization, and GRK3 engagement while not affecting G protein recruitment (mGi), G protein-mediated signaling (CAMYEL), and cell migration. We compare these directly to the effects of the chemokine alone, which shows robust responses across all of these receptor outputs. The use of the term “biasing” in the title is reflective of this observed change in receptor response to the modified agonist. Additionally, as noted by the reviewer below, we compare similar outputs to compare the impact on effector coupling, eg. Arrestin and mini-G protein recruitment. We strongly feel that the term “biasing” most accurately describes the phenomenon we are describing with regard to the altered receptor response to the VUP25111:CCL25 complex and thus prefer to keep the title without alteration.

(2) Abstract, line 34: The term "bias" should be replaced. As currently used, it appears to suggest a dichotomy between G protein signaling and arrestin recruitment. However, arrestin recruitment is a consequence of G protein signaling, and it is not conceptually sound to compare a signaling event mediated by one protein family with a protein-protein interaction involving another protein family. A meaningful comparison requires experimental paradigms that differ by a single variable; in this case, there are two - distinct protein families and fundamentally different types of readouts (signaling versus protein-protein interaction).

We agree that the term “bias” must be used with precision. One point of clarification, arrestin recruitment is not explicitly a consequence of G protein signaling. The atypical chemokine receptors, of which ACKR4 is a member, are well-established examples of GPCRs that do not bind nor activate G proteins yet still recruit arrestins (Lamme et al., 2026). While biased responses within the GPCR system are often the comparison of different events as noted by the reviewer, we compare the effects of miniproteins on the recruitment of mGi and arrestin via a BRET-based assay that assesses the direct protein: protein interactions between the receptor and effector. The lack of effect on mGi recruitment was further supported by measuring the effect of the miniproteins on downstream cAMP levels. To avoid any misunderstanding, we have revised the abstract to explicitly state that the observed effect induced by VUP25111 is evaluated through direct recruitment of effectors, rather than implying a comparison between disparate signaling outputs.

(3) Abstract, line 34: The term "balanced agonist" should be removed. Any chosen reference ligand is, by definition, the "balanced" agonist for that analysis, regardless of its actual signaling profile. Consequently, the expression "balanced agonist" adds no mechanistic information beyond "the agonist used as reference in a particular bias calculation" and is potentially misleading, as it implies that this ligand possesses a uniquely unbiased, system‑independent signaling profile, which is not the case.

We agree with the reviewer that the term “bias” is always relative to a specific reference ligand. Indeed, it is known that some natural chemokine:receptor pairs show some bias on its own. Therefore, we adjusted the text in the abstract on Pg.2 and removed the word “balanced”.

(4) Abstract, line 36: I also recommend removing the term "bias" at this point. The concept of bias typically arises from experiments that quantitatively compare more than one variable. As currently written, the phrasing suggests a dichotomy between G protein- and arrestin-mediated signaling, yet the study does not assess arrestin signaling, only arrestin recruitment. Under these conditions, the use of "bias" is not appropriate. The data are clear and compelling on their own without the need for this potentially misleading terminology.

We thank the reviewer for this comment. We would like to emphasize that our conclusions are based on the fact that we measure both G protein (mGi recruitment) and arrestin (arrestin recruitment) as protein:protein interactions. We have changed the text in the abstract on pg.2 to further distinguish between signaling and recruitment:

“Additionally, G protein recruitment to CCR9 was unimpeded by VUP25111, suggesting that the miniprotein biased the native agonist towards G proteins.”

(5) Introduction: This is interesting to read and generally well written, though certain statements would benefit from improved semantic precision. For example, in lines 110-111, the phrase "G protein-biased complex" should be reconsidered, as it relies on the notion of G protein- versus arrestin-mediated signaling. Arrestins themselves do not signal; what is measured here is their recruitment. Comparing G protein signaling with arrestin recruitment is therefore conceptually unsound, since arrestin engagement is a downstream consequence of G protein activation. Comparisons become meaningful only when designed to differentiate between G protein-dependent and G protein-independent arrestin recruitment, which is not the case in this study.

Since the VUP25111:CCL25 complex suppresses arrestin recruitment to CCR9, but not G protein recruitment, we are confident representing the modified agonist as G protein-biased. Additionally, arrestin recruitment is only downstream of G protein activation in the occasion where the GPCR is phosphorylated specifically by GRK2/3 and not GRK5/6. CCR9 does not show any preference for GRKs and recruits arrestin as effectively when phosphorylated by either GRK2/3 or GRK5/6 (Lamme JBC 2025). Since G protein coupling is preserved when challenged with the VUP25111:CCL25 complex, both families of GRKs should be available for phosphorylation. We have modified the statement to better specify recruitment compared to signaling on Pg. 5:

“Most prevent activation by impeding chemokine binding to the receptor, thus impairing all signaling and migratory responses, but one forms a G protein-biased complex with CCL25, preserving G protein recruitment while suppressing arrestin recruitment to CCR9, and selectively inhibits CCR9, but not ACKR4.

(6) Results, 122,123: The authors should consider being more precise; possibly, the truncated CCL25 is somewhat less potent on CCR9. The authors should make a statistical test and then decide whether to rephrase or not for enhanced precision.

The reviewer is correct that we detect a small difference in potency between full-length CCL25 and CCL25-ΔCT at CCR9. This was quantified and is reported in the supplementary Table S2, where CCL25 shows a logEC50 of -7.46 ± 0.16 compared with -7.89 ± 0.29 for CCL25-ΔCT, indicating that the truncated chemokine is slightly more potent. While the difference was determined to be statistically significant per an Extra sum-of-squares F test, we don’t consider the of the truncation to be substantially impactful, which is also in line with previous publication (Pinheiro Nat. Comm. 2025). We have therefore revised the text to note the modest increase in potency of CCL25-ΔCT compared to the native, full-length CCL25 and added statistical information in the figure legend. We have changed the text on Pg. 5 to:

“The truncated CCL25 (CCL25_ΔCT, 1-77) activated both CCR9 and ACKR4 with similar efficacy while showing slightly enhanced potency only towards CCR9 compared to the full-length ligand (Fig. S1).”

(7) Figure S5: This figure is currently confusing and needs clarification. The authors state in the main text that CXCR4 is stimulated with CXCL12, yet the figure legend refers to CCL25; this discrepancy should be corrected to ensure consistency. In addition, inhibition of CXCR4 by the miniprotein binders should be analyzed and presented with normalization to CXCR4 responses, not to CCL25-stimulated CCR9. To avoid misinterpretation, inhibition by the miniproteins should be quantified separately for CCR9 and CXCR4, each normalized to its own receptor-specific and functionally equivalent stimulation condition, rather than to the "other" receptor.

We thank the reviewer for this comment and apologize for the oversight during editing. The mentions of CCL25 in Figure S5 were incorrect, and CXCL12 was indeed used in this cross-reactivity assay to stimulate CXCR4 and for subsequent normalization. We have adjusted the legend accordingly.

(8) Results, lines 211-213: The authors should be more semantically precise. They state that no binder has any effect on arrestin recruitment to CXCR4. If I see the data, this is not really true, as 25101 and 25107 inhibit arrestin recruitment by about 50 % or more at the highest applied concentrations; only 111 and 112 are completely inactive. As already commented, normalization should be done to arrestin recruitment of CXCR4 and not CCR9.

We thank the reviewer for this comment and agree that, at the highest tested concentrations of VUP25101 and VUP25107, substantial inhibition is observed. Additionally, we have extended our analysis of the miniprotein selectivity to include activation assays for the CCL20 receptors CCR6 and ACKR4, as well as the atypical CXCL12 receptor ACKR3. Based on these data, we have revised the text to more accurately reflect these observations. We now frame these results as an initial proof of concept demonstrating the feasibility of this design approach, while noting that further optimization and characterization of selectivity will be required to advance the miniproteins for additional applications.

We have adjusted the text on Pg.11:

“CCL25 specificity was experimentally evaluated by testing the miniproteins against CXCL12 by tracking arrestin recruitment to CXCR4 and ACKR3, the native agonist for this canonical and atypical receptor pair (Fig. S5). VUP25111 and VUP25112 showed no effect on CXCL12-mediated activation at any concentration, while VUP25101 and VUP25107 showed partial inhibition at the highest tested concentration (10 µM).

In addition, the effects of the VUPs were evaluated against CCL20, a closely related chemokine to CCL25, by assessing arrestin recruitment to CCR6 and ACKR4 (Fig. S6). VUP25107 and VUP25112 displayed similar potencies towards CCL20 comparable to those observed for CCL25 (VUP25107: pIC50 = 7.72 ± 0.21; VUP25112: pIC50 = 6.93 ± 1.03) to ACKR4. In contrast, VUP25101 showed reduced inhibition of CCL20 relative to CCL25 (VUP25101: pIC50_CCL20 = 6.06 ± 0.13 v pIC50_CCL25 = 7.73 ± 0.08), indicating greater selectivity. CCL20 activation of CCR6 further supports the specificity of VUP25101, which only showed a modest effect at the highest concentration tested. VUP25111 showed no inhibitory effects on CCL20 activation of either CCR6 nor ACKR4, which may indicate a lack of binding, but is indiscernible from forming a receptor-competent complex similar to how it interacts with CCL25 and ACKR4.”

Concerning the comment regarding the normalization to CCR9 rather than CXCR4, due to a mistake, the figure legend incorrectly stated CCL25, whereas the assay was performed using CXCL12. We thank the reviewer for bringing this up and have corrected the figure accordingly.

eLife Assessment

This study investigates the role of Interleukin-2-inducible T cell kinase (ITK) deficiency in autoimmune lung injury using a pristane-induced pulmonary hemorrhage (PH) model, suggesting that ITK-deficient regulatory T cells (Tregs) restrict severe tissue pathology. The work represents an important addition to the fields of autoimmunity, inflammation, and T-cell biology in the lung. The experimental evidence supporting the core ideas is solid and provides mechanistic insight into how ITK deficiency promotes a tissue-protective Treg phenotype, although additional studies will be needed to further define the molecular pathways involved.

Reviewer #1 (Public review):

[Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

In this study, Hossain et al. investigated the role of Interleukin-2-inducible T cell kinase (ITK) in autoimmune lung injury, demonstrating that ITK-deficient (Itk-/-) mice are protected against pristane-induced pulmonary hemorrhage (PH). Authors suggest that this protection correlates with a significant remodeling of the T cell compartment in Itk-/- mice, including increased frequency of memory-like CD4+ and CD8+ T cells (CD44⁺CD62L⁺) as well as higher frequency of Treg populations. Furthermore, adoptive transfer of ITK-deficient Treg isolated from injured ITK-deficient mice confers protection against pulmonary hemorrhage in WT recipients.

Strengths:

The adoptive transfer of congenically marked wildtype and Itk-/- Treg populations clearly demonstrates that ITK-deficient Treg can actively rescue pre-existing lung injury and reverse systemic secondary metrics like proteinuria in wild-type recipients, providing proof-of-concept validation for the therapeutic utility of the ITK-Treg axis.

Reviewer #2 (Public review):

Summary:

In this manuscript, Hossaim and colleagues investigate the role of the ITK kinase in modulating inflammation in a pristane-induced lung hemorrhage model. Using a germline ITK KO mouse, they report that loss of ITK skews the T cell compartment toward a memory-like state, expanding Tregs, and conferring protection against alveolar hemorrhage, inflammatory monocyte recruitment, proteinuria, and systemic cytokine elevation. They further show that transfer of ITK-deficient Tregs into wild-type hosts with established disease attenuates injury and shifts the cytokine balance toward resolution, and that ITK-deficient Tregs carry a transcriptional signature enriched for OXPHOS, mTORC1, MYC, and cell-cycle programs. While these observations are interesting for the development of potential immunotherapies, there are several issues with the methodological approach that support the authors' claims, tempering my enthusiasm for this manuscript.

Strengths:

(1) The clinical motivation and potential targeted therapies are relevant.

(2) The murine phenotype seems robust.

Comments on revised version.

The authors have addressed my comments appropriately.

Reviewer #3 (Public review):

Summary:

Hossain et al. investigate the role of ITK as a central regulator of autoimmune lung injury. They used ITK-deficient mice and the pristane-induced pulmonary hemorrhage (PH) model to show that ITK deficiency confers protection against PH. The adoptive cell transfer experiment suggests a possible role for altered Treg cells in ITK-deficient mice in regulating the inflammatory response in the lungs of pristane-injected mice. This study shows that targeting the ITK axis may be beneficial by reducing systemic inflammatory injury that contributes to poor outcomes in PH.

Strengths:

This study highlights the importance of ITK in regulating pulmonary hemorrhage. The enrichment of Treg cells is known to confer protection in autoimmunity-mediated alveolar damage. However, the involvement of ITK in regulating the Treg cells' function is interesting and could be explored as a novel therapeutic approach for chronic inflammation.

Comments on revised version:

The authors have answered my comments by incorporating new results and clarification to their observation.

Author response:

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

Reviewer #1 (Public review):

Strengths:

The adoptive transfer of congenically marked wildtype and Itk-/- Treg populations clearly demonstrates that ITK-deficient Treg can actively rescue pre-existing lung injury and reverse systemic secondary metrics like proteinuria in wild-type recipients, providing proof-of-concept validation for the therapeutic utility of the ITK-Treg axis.

We sincerely thank Reviewer #1 for recognizing the strength and translational significance of our adoptive-transfer experiments. We agree that these studies provide important proof-of-concept evidence that ITK-deficient Tregs are not simply associated with protection but can actively ameliorate established pulmonary hemorrhage and associated systemic manifestations when transferred into wild-type recipients. We appreciate the reviewer’s recognition of the potential therapeutic significance of targeting the ITK–Treg axis.

Weaknesses:

The molecular analysis of Tregs isolated from unchallenged wild-type ITKdeficient mice has limited value to understand how the ITK-deficient Tregs protect against autoimmune pulmonary hemorrhages in lungs. Detailed molecular analysis of Tregs isolated from the lung will be helpful in understanding the mechanistic regulation and identifying further therapeutic targets.

We thank the reviewer for this important observation and agree that detailed molecular characterization of lung-resident Tregs following pristane-induced pulmonary injury would provide additional insight into the mechanisms responsible for protection.

The transcriptional analyses presented in the current manuscript were designed to identify intrinsic molecular differences associated with ITK deficiency in Tregs independent of the inflammatory environment. These studies revealed alterations in pathways including oxidative phosphorylation, mTORC1 signaling, MYC-associated programs, and cell-cycle regulation, thereby providing a molecular framework for the enhanced function of ITK-deficient Tregs.

We recognize, however, that the phenotype and transcriptional program of Tregs within the injured lung microenvironment may differ from those observed under baseline conditions. We therefore consider direct molecular comparison of lung Tregs isolated from pristane-treated wild-type and ITK-deficient mice an important direction for subsequent mechanistic studies.

Importantly, the principal functional conclusion of the current study is supported by the adoptive transfer experiments shown in Figure 5. Transfer of ITK-deficient Tregs into wild-type mice with established disease significantly reduced pulmonary hemorrhage and systemic injury, demonstrating that ITK-deficient Tregs themselves are capable of mediating protection. We have clarified this limitation and the need for future studies examining lung-resident Tregs in the revised Discussion. See page 21 line 353- 361.

Reviewer #2 (Public review):

The authors have addressed my comments appropriately.

We sincerely thank Reviewer #2 for carefully evaluating our revised manuscript and for concluding that our previous responses appropriately addressed the reviewer’s concerns. We also appreciate the reviewer’s recognition of the potential relevance of our findings for the development of targeted therapeutic approaches.

Reviewer #2 (Recommendations for the authors):

I would strongly recommend removing Fig. 5J (IF), as there is no convincing signal for Foxp3. This is a well-known issue with Foxp3 Ab. If authors would like to include imaging, would recommend performing smFISH/RNAScope.

We thank the reviewer for this helpful recommendation. We agree that the Foxp3 immunofluorescence signal in the previous Figure 5J does not provide sufficiently convincing visualization to support our conclusions. We have therefore removed Figure 5J from the revised manuscript.

Importantly, our conclusions regarding the protective activity of ITK-deficient Tregs are independently supported by flow-cytometric analyses and, most importantly, by the adoptive-transfer experiments demonstrating that ITK-deficient Tregs can ameliorate established pulmonary hemorrhage in wild-type recipients.

We appreciate the reviewer’s suggestion regarding smFISH/RNAscope and agree that this approach

could provide valuable spatial information in future studies.

Reviewer #3 (Public review):

The novelty of this study lies in the association between ITK-deficient Tregs and pulmonary hemorrhage in autoimmunity. One limitation of this study is that the authors performed all experiments in ITK global knockout mice, in which not only T cells but all other cell types are deficient in ITK. Therefore, further studies are required for a mechanistic rationale for targeting ITK in severe inflammatory lung disease.

We sincerely thank Reviewer #3 for recognizing the novelty of the relationship between ITK-deficient Tregs and protection from autoimmune pulmonary hemorrhage, as well as the potential therapeutic relevance of this pathway in chronic inflammatory disease.

We agree that the use of a germline ITK-deficient mouse represents a limitation of the current study because it does not allow us to completely distinguish T-cell-intrinsic effects of ITK deficiency from potential contributions of other immune populations.

However, our adoptive-transfer experiments provide important functional evidence supporting a direct role for ITK-deficient Tregs. Specifically, purified ITK-deficient Tregs transferred into wild-type recipients with established pulmonary hemorrhage significantly reduced pulmonary injury and associated systemic manifestations. These findings demonstrate that ITK-deficient Tregs are themselves sufficient to confer substantial protection within an otherwise ITK-sufficient host.

Nevertheless, we agree that future studies using lineage-specific conditional ITK-deficient models will be important to define the Treg-intrinsic contribution of ITK signaling more precisely. Such studies, together with evaluation of ITK modulation in human Tregs and appropriate humanized models, will be important for determining the translational potential of targeting the ITK pathway in severe inflammatory lung disease.

We have revised the Discussion to more clearly acknowledge this limitation and highlight these future directions. We again thank the editors and reviewers for their thoughtful and constructive comments. Their suggestions have substantially strengthened the manuscript and helped us more clearly define both the significance of our findings and the remaining questions that will guide future investigation. See the limitations of this manuscript.

eLife Assessment

This manuscript seeks to understand the hierarchical temporal representation of social actions in the lateral visual pathway in the human brain. It introduces a potentially valuable methodological approach for characterizing the spatiotemporal dynamics of social-interaction representations, with the data supporting the hypothesis that the anterior portion of the superior temporal sulcus responds to communicative actions, and providing a temporal estimate of when it begins responding. However, the failure to recover several previously reported effects regarding mid-level representations in this and other visual brain regions, combined with unresolved concerns about the correspondence between EEG, fMRI, and stimulus-feature measures, limits confidence in the method's sensitivity and validity, rendering the current empirical evidence against a hierarchical feedforward architecture incomplete.

Reviewer #1 (Public review):

Summary:

The manuscript presents a study combining human functional MRI (fMRI), electroencephalography (EEG), and human judgement data within a two-step regression framework to determine when several regions of interest (ROIs) along the so-called third visual pathway for social perception are activated, and which aspects of visual representations of social interactions they encode. The goal is to determine whether social interactions are encoded through a hierarchy of representation along the third pathway, and whether this hierarchy is implemented in a feedforward manner.

The first analysis linking EEG to fMRI data identifies early activation (within 60 ms) in early visual cortex (EVC) and later (onset between 112 and 310 ms) activations in selected, more anterior ROIs along the pathway, i.e., motion-selective MT, extrastriate body area (EBA), object-selective lateral occipital cortex (LOC), and posterior and anterior social-interaction superior temporal sulcus regions (pSTS and aSTS). The second analysis linking EEG-fMRI data to stimulus features derived either computationally from the images or from human judgments shows that EVC activity arises early (within 68 ms) and is predicted by low-level visual features and motion energy. Activity in all other ROIs emerges after 140 ms (or later) and is predicted by motion energy (in MT, EBA and LOC), arousal (in MT), spatial expanse (in LOC, pSTS and aSTS), agent distance (in LOC, pSTS and aSTS) and communication (in aSTS). No effects of facingness, object-directedness, joint action, or valence are found in any ROI.

Two observations are highlighted: features important for recognizing social interactions (i.e., agent distance) are encoded with a similar latency in the mid-level LOC and in STS; the communication content is selectively encoded in anterior STS.

It is concluded that social interactions are processed in a hierarchical way, with more abstract information (i.e., communication) encoded in more anterior regions. Moreover, because this higher-level information is encoded at similar or even shorter latencies than mid-level features in mid- and high-level ROIs, these findings are taken to challenge a strictly feedforward hierarchy within the third visual pathway.

Strengths:

This is a timely manuscript concerning a newly discovered and now extensively investigated brain network, purportedly dedicated to social perception. In the manuscript, the selectivity of this network for the perception of social interactions is assumed rather than demonstrated. Instead, the study addresses two related questions that are critical for understanding how the network (or pathway) is organized and functions: (i) whether representations are hierarchically organized, progressing from more specific information (lower-level visual features) to increasingly abstract information (interaction category, e.g., communication), and (ii) whether processing follows a feedforward mechanism with prior and rapid encoding of lower-level features of social interactions, followed by encoding of mid-level and higher-level features.

Addressing these questions is important because it goes beyond the knowledge of where social information is represented, toward understanding what kinds of visual input and transformations support the recognition of social interactions, and how the processing stages are organized (e.g., in a feedforward manner or with recurrent interactions). Establishing both aspects would define the particular computational architecture supporting the human ability to understand visual social interactions.

The manuscript is also methodologically valuable, as it introduces a variation of more conventional fMRI-EEG fusion analyses. This method establishes fMRI and EEG correspondences, thus revealing spatiotemporal dynamics of brain activity (i.e., where activity occurs in the brain, at which time point), and further links these dynamics to stimulus features (i.e., which aspects of the stimuli predict activity in a given brain area at a given time). The description and discussion of the method in the manuscript are clear and transparent. This facilitates replication and allows appreciation of the rationale of the approach and its potential for future research.

Weaknesses:

The method introduced in the manuscript appears to be a valuable addition to the range of non-invasive approaches to study the spatiotemporal dynamics of representations in the human brain. The results presented in the manuscript, however, do not make a strong case for the sensitivity of the method.

The study fails to replicate several effects that have demonstrated reliability across previous studies. In particular, motion energy in the stimuli only poorly predicts - relatively late -activity in motion-sensitive MT. Moreover, unlike several previous studies, including work by McMahon, Isik, and colleagues, the study finds no effects of facingness (or any other mid-level features) in EBA, object-directedness in LOC, joint actions in pSTS, and valence in pSTS. The manuscript does not discuss discrepancies, raising concerns about the sensitivity and validity of the method. In particular, the absence of previously reported effects is difficult to interpret without considering whether the method is sufficiently sensitive to detect the relevant stimulus dimensions.

Another related and unexplored issue concerns the potentially limited correspondence between EEG and fMRI signals, and between stimuli used to derive the stimulus features for the analyses and stimuli used in the actual EEG experiment. EEG and fMRI data were collected in different samples, and different individuals may assign different social value to the same events, potentially affecting neural processing in regions such as the pSTS (see e.g., work by Varrier, Finn et al.). Furthermore, the stimuli used in the EEG experiment were derived from, but differed from those used in the fMRI experiment. Although this is not entirely clear from the manuscript, it appears that stimulus features were measured for the original stimuli used in the fMRI experiment and assumed to be valid also for stimuli in the EEG experiment. The manuscript does not acknowledge these issues, or other potential sources of noise arising from them, either on their own right or in relation to the above null effects. This makes it difficult to determine whether the absence of several expected effects reflects genuine characteristics of neural representations or limitations of the measurement and analysis pipeline.

Beyond the methodological concerns, the results themselves are less straightforward than the manuscript suggests. Evidence for hierarchical representation emerges from the effect of low-level visual features in EVC and the encoding of communication in aSTS. Activity in all ROIs 'in the middle', from MT, EBA and LOC to pSTS, is predicted by a mixture of low- and mid-level stimulus features, whose effects in some cases come online after the encoding of the interaction category in aSTS (at 195 ms). This raises important questions concerning how a higher-level representation of an interaction category can emerge before the encoding of the lower- or mid-level information that is presumably diagnostic of that category. The manuscript would benefit from a more thorough discussion of this issue, particularly because the temporal ordering of these effects is central to the authors' claims about hierarchical and feedforward processing.

Reviewer #2 (Public review):

Summary:

The authors seek to understand the hierarchical temporal representation of social communicative actions in the posterior superior temporal sulcus and other regions of the lateral pathway. They apply an analysis framework in which they predict fMRI responses from EEG responses to the same stimuli. (The EEG stimuli were shorter versions of the same clips shown as fMRI stimuli.) They examine which temporal segments of EEG data are most predictive of responses in which part of the brain. They find that early (< 100 ms post-stimulus) EEG activity is predictive of responses in early visual cortex, and that later (~ 150 ms post-stimulus) EEG activity is predictive of responses in the lateral visual stream. They then model the fMRI responses predicted by the EEG signals - which they term the "EEG-relevant" portion of the fMRI signal - as a function of various social and visual factors. They conclude that social interactions elicit responses in aSTS-SI later than low-level visual features elicit responses in early visual cortex, and they provide some evidence that mid-level social primitives elicit responses in mid-level areas somewhere in between.

Strengths:

I find the question and the framework interesting. The extraction of "EEG-relevant" fMRI signal as a target for visual and social feature-based encoding models has potential, though the approach in this instance suffers from low signal-to-noise. The evidence that the anterior portion of the superior temporal sulcus responds to communicative actions is clear and supports prior work in this regard.

Weaknesses:

I think there are several important points that the authors need to address to strengthen their case, and I think that many points could use more explicit description, explanation, and discussion.

Major concerns:

First, the basis of the EEG predictions of fMRI activity is not clear. The paradigm relies on predicting fMRI responses at each voxel from a weighted combination of EEG responses measured anywhere over the scalp. However, the authors never show any kind of weight map to visualize which electrodes on the scalp were weighted to predict responses in each fMRI voxel or ROI. I see that the use of PCA to orthogonalize the EEG responses creates problems with localizing the predictive signal in EEG to particular electrodes. Is the PCA step necessary? In any case, some form of analysis to estimate where in the brain the predictive EEG signal is coming from seems important. It's fairly clear that the best signal quality in the EEG recordings is in the early visual cortex. Thus, it could be the case that electrodes over EVC provide the main basis for predictions across the entire brain, with different temporal phases of EVC activity associated with activity in different regions in the fMRI data. It could also be the case (tho less likely in my opinion) that activity in regions away from EVC on the scalp provides the basis for predicting more anterior brain regions in the fMRI data. These two possibilities each suggest a substantially different interpretation of the present results: in the former case, there may be some sort of feedback loop between early visual areas and lateral areas; in the latter case, the hierarchical interpretation the authors have laid out seems more plausible. It seems important to draw a distinction between these possibilities.

In any case, the authors should say more in the discussion about the sources of EEG signals and fMRI signals and the relationship between the two.

On a similar note: the authors show that their EEG data was reliable based on a split-half analysis of the data. However, the authors only show average split-half correlation across electrodes and subjects. It stands to reason that some electrodes were much more reliable than others. The authors should show a map of electrode reliability across the scalp. One relevant question is: are there reliable electrode locations throughout the visual cortex? Or are the EEG signals only, or by far most reliable, in EVC? (How reliable is the EEG signal in regions that are at least somewhat near to the location of the lateral regions that is the principal object of investigation here?)

The use of predicted fMRI signals (predicted by EEG) as a target for image & social feature encoding models is only valid to the extent that the EEG signals accurately predict the fMRI signals. The prediction accuracy going from EEG to fMRI is not very high - a correlation of less than 0.2 in most regions besides EVC and the anterior extent of the activity along the ventral temporal lobe. In the best-predicted regions, the activity only reaches a correlation of 0.3. The majority of the signal (> 90% of the variance) is unmodeled in this analysis. Thus, I would hesitate to characterize the evidence for activation by particular stimulus features at particular times in this analysis as "strong". The authors themselves are somewhat ambivalent on this point, as well. While they write in the conclusion: "Finally, we find strong evidence that social interactions are computed via hierarchical computation of low- and mid-level social visual features in the lateral visual pathway." However, This statement conflicts somewhat with other statements the authors make in the discussion section: "Together, these results suggest that lateral visual regions may not be organized in a strict feedforward hierarchy as has previously been suggested" Later in the discussion there is more ambivalence about the hierarchical nature of the representation: "Our findings suggest that the processing of social action features from posterior to anterior brain regions does not adhere to a strict feedforward hierarchy." The authors should present a more consistent interpretation of their findings.

Finally, the low performance of most of the mid-level models in mid-level regions is a concern; it points to poor signal overall. The common threads in the models that do predict responses above chance seem to be space (spatial expanse and agent distance) and salience (to the extent that arousal may constitute salience). This raises a question: Were the social model features correlated with either of the low-level model features?

I think that the overriding issue here is poor signal quality in the EEG data. The estimate of repeatability from the split-half analysis is not high, even on average for the whole brain; it is likely to be very poor for areas outside EVC. I think this matters. An important null hypothesis to consider is that there may be a few fairly low-level, or at least simple and straightforward, elements of the response that are reflected in the noisy EEG signal. Chief among these is likely to be that the most salient or high-contrast retinotopic regions in the image, which both AlexNet and motion energy models would likely pick up on, are likely to elicit responses in EVC. This is likely why EVC is by far the best-predicted region, both based on the raw EEG data and in the two-step regression analysis. There could be elements of spatial priority at all levels of the visual system. This could be what the authors are observing: there is some relatively small (in terms of variance explained) component of the signal that is common to many regions, and which emerges somewhat later in later regions due to conduction delays.

I think an important exception to this signal problem is the finding of communicative actions eliciting responses in aSTS-SI. That seems very robust and clear.

Medium comments:

There should be a section in the methods on how the ROIs were defined. (There is a brief description in the results section, but this is not sufficient). The ROI outlines should also be displayed somewhere in the paper (not just circles showing the general region as in Figure 1). Figure 3 would be a nice place for them, as would Figure S3.

It's also worth noting in the methods and discussion that the use of template ROIs for hMT in particular is potentially sub-optimal. hMT lies in anatomically variable locations across subjects. One reason for the seemingly weak signal in hMT may be that the voxels being averaged do not fall within the true hMT for each subject.

In the results (section 3.2.1), the authors write, "We scored the regression by predicting the EEG-predicted fMRI activity (estimated fMRItest) to the held-out test videos and then correlating the actual fMRI response (fMRItest) in the test set to this prediction at each time point (estimated fMRItest)." Unless I am missing something, I think the authors can safely say that they predicted responses to the withheld stimuli without specifying the intended prediction target (whether it is the fMRI data or the estimated fMRI data), because the prediction is simply based on weights, labeled or estimated image features, and stimuli. So, the prediction might as well be a direct prediction of the fMRI test data. It seems to me that mentioning the estimated fMRI test in this sentence only muddies the waters.

In the two-stage regression analysis, were all the visual and social features fit to the EEG-predicted fMRI data in a single regression? Was this regression regularized? How many parameters total were fit? (This is clear for the AlexNet and motion energy models, but not clear for the other features - were they all single-feature models?)

Author response:

We thank the reviewers and editors for their constructive feedback and are in the process of preparing a detailed point-by-point response. In the meantime, we would like to underscore our main conclusions and highlight the consistency with this study and the prior fMRI literature.

First, while the reviewers are correct to highlight the difficulty in interpreting null effects, the claims of our paper are primarily concerned with the positive findings. To our knowledge, this is the first evidence that social interaction perception is extracted in a fast hierarchy, which progresses from early visual features to mid-level social primitives to high-level information about communicative actions in the STS. This is an exciting and novel finding that sheds new light on the neural computations that support this core human ability. Importantly, the main null finding is the lack of difference in the onset latency between mid-level and anterior ROIs, rather than an inability to decode information in these regions (see below) and does not undercut this central claim.

Second, we would like to highlight that, overwhelmingly, our results are consistent with the prior fMRI literature. For example:

• Using our two-step regression procedure, we recover representations of motion energy in MT validating the use of this method

• Further, we find representations of mid-level features in mid-level regions ~150 ms after stimulus onset, primarily the distance between people.

This consistency, and the decoding of high-level features like communication exclusively in STS regions (as highlighted by reviewer 1), provides support for our method and shows that the EEG-fMRI correspondence extends beyond early visual regions.

The reviewers also raise questions about potential inconsistencies with our prior fMRI study (McMahon et al., 2023). We would like to clarify those findings here, because the current manuscript relies on the same fMRI data, and thus we would not expect to see effects here that are not reported in that fMRI data. In our original fMRI study, we did not find an effect of facingness, joint action, or valence in any ROI. Further, in that study, the effect for communication was most robust in aSTS, rather than pSTS. The primary effect from our original fMRI study that we do not find with EEG-fMRI fusion is information about object-directed actions in LOC. We will address these in a more detailed point-by-point response and revision soon.

Beyond the practical limitations of what we could find in the current fMRI dataset, we would also like to discuss the theoretical reasons for differences between our findings and the broader literature. First, our lack of finding an effect of valence likely reflects our stimuli, not neural representations. The Moments in Time computer vision dataset that our videos are derived from does not sample strongly negative videos. This restricted valence range is also evident in the behavioral ratings that participants provided.

Further, there are many differences between how we have looked for an effect of facing direction compared to prior work in this area. First, we do not use a binary facing/not facing but a graded scale from facing to not facing, and, to our knowledge, ours is the only study that has investigated a facing effect in naturalistic videos. It is possible that due to the graded scale or the natural stimuli that the facing effect is too subtle for us to detect. Another possibility is that while facingness is a social cue that is encoded in the brain, it is not a dominant cue in naturalistic videos where other cues are present (e.g., the distance between people, physical contact, and contingent motion).

Finally, on joint action, a type of physical social interaction between people, there is now convergent evidence from multiple studies that this may not be the primary type of social interaction represented in the STS. McMahon et al. (2023) found that communicative, not joint interactions, predicted STS responses. This was replicated and extended in (Zhao et al., 2025), where we find that joint action representations are more represented in lateral occipital regions, while communicative interactions are represented in the STS. From ongoing work in controlled stimuli (McMahon & Kanwisher, 2026 VSS), we have found that regions in the STS respond more to communicative interactions than joint actions. This present study provides new insight into the computations that support this, suggesting that information about communicative interactions is extracted along a spatiotemporal hierarchy, consistent with the fast timescale of visual processing, along the newly proposed lateral stream.

References:

McMahon, E., Bonner, M. F. & Isik, L. Hierarchical organization of social action features along the lateral visual pathway. Current Biology 33, 5035-5047.e8 (2023).

Zhao, Y., McMahon, E. & Isik, L. Separate Neural Representations for Physical and Communicative Social Interactions: Evidence from Data-driven Voxel Decomposition. Preprint at https://doi.org/10.31234/osf.io/fmnq9_v1 (2025).

McMahon, E. & Nancy Kanwisher. The STS Responds Preferentially to Communicative Signals. Vision Science Society; May 15–19, 2026; St. Petersburg, FL.

eLife Assessment

This valuable study identifies Gcn5 as a regulator of blood cell development in the Drosophila lymph gland, with links to autophagy and nutrient-sensing mTORC1 signalling. While the mechanistic ordering between Gcn5, mTORC1 and TFEB/autophagy remains less fully resolved, the evidence is solid that perturbing Gcn5, autophagy genes and mTORC1 activity affects blood cell homeostasis. The work will be of interest to researchers studying developmental haematopoiesis, metabolic regulation and autophagy.

Reviewer #1 (Public review):

In their manuscript Arjun et al. investigate the role of the histone acetyl transferase Gcn5 in controlling Drosophila blood cell homeostasis in the larval lymph gland. Using gcn5 zygotic mutants as well as targeted knock-down and over-expression of Gcn5 in various lymph gland cell populations, they show that these manipulations impact niche cell number, blood cell progenitor maintenance, plasmatocyte differentiation, crystal cell differentiation, DNA damage accumulation. Their results indicate that Gcn5 controls autophagy and show that reducing the expression of the autophagy machinery affect blood cell differentiation. By using drugs as well as genetic approaches to modulate the mTOR pathway, they found that Gcn5 levels are regulated by mTOR, but that the impact of this pathway on blood cell homeostasis can override Gcn5 function.

Overall, the main conclusions are sound. They bring interesting insights into the regulation of autophagy by Gcn5 and mTOR and lay the ground for further dissections of Gcn5 function and mode of action during Drosophila lymph gland development.

Comments on revised version.

The authors have modified the title, abstract and discussion to tune down some of their conclusion and provide a more nuanced interpretation of their data. They have also modified some figure panels as requested.

Reviewer #2 (Public review):

Summary:

Drosophila haematopoiesis has been shown to be governed by a number of signalling pathways such as JAK/STAT and Dpp. This important study shows a role for nutrient sensing and autophagy in determining blood cell differentiation. The authors show that General control non-derepressible 5 (Gcn5), a histone acetyltransferase affects blood cell differentiation. Gcn5 also negatively regulates autophagy through its effector TFEB which directly regulates autophagy genes. The authors also show that mTORC1 modulates Gcn5 levels and through it TFEB activity thus acting as a fine-tuning mechanism which maintains optimal levels of autophagy.

Strengths:

The main strength of the work lies in the interesting finding that cellular metabolic processes such as autophagy has a direct role in blood cell differentiation and has the potential to be of interest to those working on vertebrate haematopoiesis as well. The report has generated intriguing data, using promoters specific for sub sections of the lymph gland, that different cellular subsets of the lymph gland contribute differently towards haematopoiesis, but this is not followed up in detail and the final conclusions are derived from a combination of whole lymph gland perturbations as well as those from specific promoters.

Weakness:

(1) Gc5 seems to be expressed throughout the lymph gland but modulating it in the subsections do not have the same result. It is very striking that the knockdown of Gcn5 in the prohemocyte population does not have an effect on differentiation whereas overexpression does. And the modulations of Gcn5 in PSC also has variable effects across hemocyte subpopulations which is not explored in the manuscript. Interestingly, also the domain deletion constructs show differential effect on blood cell differentiation when altered solely in the prohemocytes which is not explained. While Gcn5 can be seen in all sections of the lymph gland in the first figure, under the HHLT-Gal4 and Hml-Gal4, Gcn5 looks cytoplasmic and almost completely excluded from the nucleus strikingly unlike Gcn5 expression under the Collier-Gal4 and Dome-Gal4. The rest of the experiments in the manuscript are done with multiple promoters, with autophagy flux measured by modulating Gcn5 with a pan hemocyte promoter, but the mTORC1-Gcn5 axis is explored using chemical modulators which affect the whole of the lymph gland (Fig7) or using two pro-hemocyte promoters (Fig8).

(2) The knockdown of Gcn5 seems to affect the gland size (A compared to B and C). Since mTORC1 is a central regulator of cell size, it is possible that some of the effects seen in these knockdowns are potentially through mTORC1 affecting size suggesting that the signalling axis between mTORC1 and Gcn5 might not be a one-way axis as suggested in Figure 9. Also, this would mean that in experiments where absolute cell counts of crystal cells or niche cells are used to assess blood cell differentiation, further analysis to consider total cell numbers in the lymph gland would strengthen the manuscript.

(3) A genetic manipulation of mTORC1 specifically in the pro hemocytes would strengthen the role of mTORC1 in the pathway rather than the chemical modulation which affects the whole of the lymph gland.

Comments on the revised manuscript:

Overall, the revisions make the narrative more coherent and pertinent to the results presented. The authors have also added new data and the analysis supports their conclusions.

Author response:

The following is the authors’ response to the previous reviews

Public Reviews:

Reviewer #2 (Public Review):

Summary:

Drosophila hematopoiesis has been shown to be governed by a number of signaling pathways such as JAK/STAT and Dpp. This important study shows the role of nutrient sensing and autophagy in determining blood cell differentiation. The authors show that General control non-derepressible 5 (Gcn5), a histone acetyltransferase affects blood cell differentiation. Gcn5 also negatively regulates autophagy through its effector TFEB which directly regulates autophagy genes. The authors also show that mTORC1 modulates Gcn5 levels and through it, TFEB activity thus acting as a fine-tuning mechanism that maintains optimal levels of autophagy.

Strengths:

The main strength of the work lies in the interesting finding that cellular metabolic processes such as autophagy have a direct role in blood cell differentiation and has the potential to be of interest to those working on vertebrate haematopoiesis as well. The report has generated intriguing data, using promoters specific for sub-sections of the lymph gland, that different cellular subsets of the lymph gland contribute differently towards haematopoiesis, but this is not followed up in detail and the final conclusions are derived from a combination of whole lymph gland perturbations as well as those from specific promoters.

Weaknesses:

(1) Gc5 seems to be expressed throughout the lymph gland but modulating it in the subsections does not have the same result. It is very striking that the knockdown of Gcn5 in the prohemocyte population does not have an effect on differentiation whereas overexpression does. The modulations of Gcn5 in PSC also have variable effects across hemocyte subpopulations which is not explored in the manuscript.

We have now explained and discuss why Gcn5 modulation could be affecting the PSC size. Please check Discussion section Paragraph 1 line 10 onwards.

Interestingly, also the domain deletion constructs show a differential effect on blood cell differentiation when altered solely in the prohemocytes which is not explained.

Currently, with our observations all that we can comment about that data is that expression of domain deletion mutants causes aberrant hematopoiesis indicating a dominant-negative phenotype since they are expressed in the wild-type genetic background. Beyond this, we will be exploring mechanistically how these domains are functioning during hematopoiesis in future studies. We have already described the dominant negative effect in the text: Discussion Section Paragraph 3.

While Gcn5 can be seen in all sections of the lymph gland in the first figure, under the HHLT-Gal4 and Hml-Gal4, Gcn5 looks cytoplasmic and almost completely excluded from the nucleus strikingly unlike Gcn5 expression under the Collier-Gal4 and Dome-Gal4.

We have now revised Figure 1 and have only included the images with Collier-Gal4 and Dome-Gal4 which clearly shows both the niche cells, Dome-positive progenitors and Dome-negative cells of the primary LG lobe essentially showing that Gcn5 is expressed throughout the primary LG lobe. In Fig. 1C-F’, Gcn5 expression is both in the nucleus and cytoplasm as this molecule shuttles between cytoplasm and nucleus. The staining pattern with the other Gal4 could be due to problems in the immunofluorescence protocol and acquisition parameters. We have now removed those images from Figure 1. Please check revised Figure 1.

The rest of the experiments in the manuscript are done with multiple promoters, with autophagy flux measured by modulating Gcn5 with a pan hemocyte promoter, but the mTORC1-Gcn5 axis is explored using chemical modulators which affect the whole of the lymph gland (Fig7) or using two pro-hemocyte promoters (Fig8).

We have used a pan-hemocyte promoter for the autophagy analysis to investigate if Gcn5 regulation over autophagy is a hemocyte-specific effect which we indeed see. We have removed the western blot data now in the revised manuscript where we looked at Atg8 and p62 levels in whole larval lysates when Gcn5 was perturbed using hemocyte driver as the results were puzzling and difficult to comprehend given the complete absence of a p62 band in Gcn5 knockdown conditions. Also, it’s worth noting that Hml-Gal4 is also active in the LG hemocytes. We did 2 alternate promoters for prohemocytes to cross-validate some of our results and the chemical modulators experiment was done since effects like mTOR inhibition/nutrient sensing effects are systemic and hence such modalities were employed.

(2) The knockdown of Gcn5 seems to affect the gland size (A compared to B and C). Since mTORC1 is a central regulator of cell size, it is possible that some of the effects seen in these knockdowns are potentially through mTORC1 affecting size suggesting that the signalling axis between mTORC1 and Gcn5 might not be a one-way axis as suggested in Figure 9. Also, this would mean that in experiments where absolute cell counts of crystal cells or niche cells are used to assess blood cell differentiation, further analysis to consider total cell numbers in the lymph gland would strengthen the manuscript.

It is a possibility that Gcn5 perturbation could be affecting lymph gland size although we have not seen any consistent trend that would point towards this phenotype either upon knockdown or over-expression. We believe Gcn5 controls blood cell differentiation phenotypes strongly via mTORC1. But in order to answer reviewer’s comment we have now re-analyzed our crystal cell differentiation data particularly and quantitated it and represented it as crystal cell differentiation index for dome-Gal4 specific Gcn5 modulation and for the data with genetic modulation of mTORC1 pathway. Please see Fig 3P and S10J for the revised analysis.

(3) A genetic manipulation of mTORC1 specifically in the pro hemocytes would strengthen the role of mTORC1 in the pathway rather than the chemical modulation which affects the whole of the lymph gland.

We thank the reviewer for their useful critique. We have now addressed this concern and we have genetically perturbed the mTORC1 pathway in the progenitors using both abrogation of TORC1 via depletion of Tor or Raptor or by activation using over-expression of Rheb. We have now included this data as Supplementary figures – Fig S10 and S11 and have described it in the results section. Please see results section “Chemical or genetic modulation of mTORC1 activity controls blood cell differentiation” in the revised manuscript.

Recommendations for the authors:

Editors Comments:

Dear Rohan,

Thank you for the substantial effort that has gone into the revision. The added data and the removal of some difficult-to-interpret results have made the manuscript more coherent.

After discussing the revised manuscript with the reviewers, our shared view is that the main interpretive issue still remains. The biology remains quite complex: Gcn5 perturbation in different lymph gland compartments still gives non-reciprocal phenotypes; the Gcn5/autophagy/mTORC1 experiments draw on different drivers, cell populations and systemic drug treatments; and the mTORC1 "override" experiment is still stronger as a pharmacological observation than as a clean pathway/epistasis result. As a result, the current framing still makes the relationship between Gcn5, autophagy, mTORC1 and haematopoiesis seem more direct than the data warrant.

We therefore ask that you revise the text to engage more directly with these unresolved points. And we urge you to please revise the title, abstract, model and discussion so that the manuscript frames the data as supporting a broad connection between Gcn5, autophagy, mTORC1 signalling and blood cell homeostasis, rather than as establishing a resolved linear Gcn5-mTORC1-TFEB/autophagy pathway.

Thank you for sending our manuscript for review again. We are glad that the revised data has made the manuscript more coherent along with removal of the data that was difficult to interpret.

We would like to thank the editor for the suggestions. We have now revised the title, abstract, model of our study and also toned down our conclusions and removed extrapolations if any. We now project a broader role of Gcn5 and mTORC1 in regulating autophagy and thereby hematopoiesis in the lymph gland. Our revised discussion section addresses this critical suggestion given by the editor. Please refer to the changes made in the title, abstract, parts of the results and discussion.

Reviewer #1 (Recommendations for the authors):

There are still many typos along the manuscript.

The authors changed their quantification of the crystal cell phenotype to "crystal cell index" in some figures but kept the initial "crystal cell number" in others. They should be consistent. Also, the scale of the axes in Fig. 2O , 2T and 3P is inadequate, as it is impossible to discern the distribution of the data or possible differences between genotypes.

We thank the reviewer for this comment. We have now gone through the manuscript thoroughly and made the required changes. We have taken into consideration and reanalysed the data and changed all quantification to crystal cell differentiation index to make the data uniform and consistent. The scale of axes as suggested for Fig. 2O, 2T and 3P have been modified to ensure the data points are clearly distributed and observable.

eLife Assessment

This important study extends a model of cortical normalization (ORGaNICs) to interacting cortical areas and shows that communication through coherence and communication subspaces can arise from a single set of dynamics. The evidence is convincing, showing analytically that contrast-dependent gamma dynamics and a low-dimensional inter-areal communication subspace arise from one parameter set. The work will interest neuroscientists and theorists concerned with inter-areal communication, cortical oscillations, and divisive normalization.

Reviewer #1 (Public review):

[Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

In this paper, Pal and colleagues propose a mechanistic unification of two influential accounts of inter-areal communication: communication through coherence and communication subspaces. A major strength of the paper is that it does not treat coherence and communication subspaces as independent phenomena, as typically done, but instead derives both from the same circuit with divisive normalization. In this framework, noise-driven fluctuations around the normalized fixed point determine covariance and cross-power structure (which, in retrospect, makes so much sense to be related). Then, they show how these determine linear prediction performance and the effective dimensionality of the communication subspace. They also show (however not very visually, see recommendation below for a figure) how divisive normalization is crucial to shape inter-areal coherence and the dimensionality of communication.

I found this conceptual contribution potentially very influential, but somewhat obscured by the technical complexity of the model. The central intuition (I think) is that recurrent normalization can organize cross-area fluctuations, both frequency-specific correlations and cross-covariances. Took me a while to grasp this insight, mostly because I was stuck with the model details. Note that I have some experience with network dynamics, but not with this particular model.

Reviewer #2 (Public review):

Summary:

The authors extend the ORGaNICs framework (a recurrent circuit that dynamically implements divisive normalization) to connected cortical areas with explicit top-down feedback. Because the network has a known analytical fixed point that coincides with (or closely approximates) the normalization equation, the authors can linearize about that fixed point and derive closed-form expressions for the power spectral density, inter-areal coherence, and communication subspaces. Using a two-area instantiation (V1 & V2) with a single fixed parameter set and no data fitting, they show the model reproduces: (i) contrast-response functions with steeper slope V2; (ii) gamma-band power and coherence peaks that shift to higher frequency with contrast; and (iii) a low-dimensional inter-areal communication subspace that is lower-dimensional than the within-area subspace. They derive parallel predictions of what happens by changing model parameters: feedback gain enhances inter-areal and suppresses within-area communication, and normalization is necessary for both the oscillatory dynamics and the reduced subspace dimensionality. A three-area extension (V1&V4, V1&V5/MT) is used to argue that differential top-down feedback can dynamically route functional connectivity.

Strengths:

(1) Analytical tractability: Deriving power spectra, coherence, and communication-subspace structure in closed form from a known fixed point is genuinely valuable.

(2) Conceptual unification: Framing coherence and communication subspaces as arising from the same normalization-driven dynamics is an elegant and useful contribution.

(3) Breadth from few assumptions: A large range of phenomena (contrast gain, gamma dynamics) emerges from normalization-based model assumptions.

(4) Biological grounding: The mapping of model variables onto identified cell types connects the abstract computation to known cortical microcircuitry.

(5) The prediction that input-gain versus feedback-gain modulation produce distinct spectral signatures gives experimentalists a clear way to test the framework.

Reviewer #3 (Public review):

Summary

The work of Pal and colleagues considers a hierarchical and multi-population version of the "oscillatory recurrent gated neural integrator circuits" (ORGaNICs) model, showing through analytics that the model captures multiple relevant experimental results: first of all, its oscillatory dynamics produce a profile with high resemblance to experimental results, both in terms of decay of power at high frequency and in terms of shifting peak as a function of stimulus contrast. Second, inter-areal communication subspace dimensionality is lower than within-area dimensionality. The authors then proceed to further characterize the model's response properties as a function of input and feedback gain. In particular, they find that frequencies transmitted with higher strength also carry more information, that changing gain modifies the dimensionality of communication subspaces, and that these properties can be used in a three-layer model, where an upstream area can select which downstream area to communicate to, based on the strength of feedback gain.

Strengths

This work demonstrates that a single-circuit model with normalization properties can capture both the oscillatory dynamics and the inter-areal communication properties measured in cortical circuits, matching multiple experimental results. The full analytical tractability of the model is highly advantageous, allowing for easier exploration of parameters, replicability, and effective interpretations of results compared to purely numerical approaches.

The work also makes a useful conceptual link between normalization, coherence-based communication, and subspace-based communication. In particular, it shows how both phenomena can emerge from the same circuit dynamics, where normalization is a key factor.

Interestingly, the model is also extended to multiple areas, showing how attention (in the form of changes in feedback gain) can synchronize the activity of a downstream area with one of two upstream areas, thus effectively selecting which area to communicate with.

In general, this is an interesting computational framework and a useful starting point for future modeling work. A particular strength is that it connects normalization, oscillatory dynamics, coherence, and communication subspaces within one analytically tractable model, making it possible to generate mechanistic hypotheses about when inter-areal communication should be stronger, lower-dimensional, or preferentially routed through feedback.

Author response:

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

eLife Assessment:

This important study extends a model of cortical normalization (ORGaNICs) to interacting cortical areas and shows that communication through coherence and communication subspaces can arise from a single set of dynamics. The evidence is solid, showing analytically that contrast-dependent gamma dynamics and a low-dimensional inter-areal communication subspace arise from one parameter set, though the comparisons to data remain qualitative and the analytics rest on a linearization that is not checked against numerical simulation. The work will interest neuroscientists and theorists concerned with inter-areal communication, cortical oscillations, and divisive normalization.

We thank the editors and reviewers for this assessment. We have addressed both stated qualifications. On the linearization, we now validate the closed-form power and coherence spectra against direct numerical simulation of the full nonlinear stochastic circuit at every contrast (new SI Fig. S6 and SI Sec. II). On the comparison to data, in addition to the unfitted comparison of Fig. 2, we now provide a quantitative comparison in which the model is fitted to the experimental power and coherence spectra, with residuals reported at every contrast (new SI Fig. S7 and SI Sec. III).

Public Reviews:

Reviewer #1 (Public review):

In this paper, Pal and colleagues propose a mechanistic unification of two influential accounts of inter-areal communication: communication through coherence and communication subspaces. A major strength of the paper is that it does not treat coherence and communication subspaces as independent phenomena, as typically done, but instead derives both from the same circuit with divisive normalization. In this framework, noise-driven fluctuations around the normalized fixed point determine covariance and cross-power structure (which, in retrospect, makes so much sense to be related). Then, they show how these determine linear prediction performance and the effective dimensionality of the communication subspace. They also show (however not very visually, see recommendation below for a figure) how divisive normalization is crucial to shape inter-areal coherence and the dimensionality of communication.

I found this conceptual contribution potentially very influential, but somewhat obscured by the technical complexity of the model. The central intuition (I think) is that recurrent normalization can organize cross-area fluctuations, both frequency-specific correlations and cross-covariances. Took me a while to grasp this insight, mostly because I was stuck with the model details. Note that I have some experience with network dynamics, but not with this particular model.

We thank the reviewer for this positive assessment. We agree that recurrent normalization shapes the cross-spectral density of the noise-driven fluctuations about the normalized fixed point, and that power, coherence, and the communication subspace can therefore be viewed as complementary readouts of the same underlying dynamics. To make this central connection explicit from the outset, we have added a new paragraph, “The core mechanism in brief” (Results, end of “Theory”), which states this mechanism before the main results are presented. The dependence of coherence and communication-subspace dimensionality on normalization is shown in SI Figs. S4e–h and S5, which are now cited wherever the role of normalization is discussed. In addition, the new Discussion subsection “Experimental predictions of the model” describes how normalization could be manipulated experimentally and the readouts expected.

Reviewer #2 (Public review):

Summary:

The authors extend the ORGaNICs framework (a recurrent circuit that dynamically implements divisive normalization) to connected cortical areas with explicit top-down feedback. Because the network has a known analytical fixed point that coincides with (or closely approximates) the normalization equation, the authors can linearize about that fixed point and derive closed-form expressions for the power spectral density, inter-areal coherence, and communication subspaces. Using a two-area instantiation (V1 & V2) with a single fixed parameter set and no data fitting, they show the model reproduces: (i) contrast-response functions with steeper slope V2; (ii) gamma-band power and coherence peaks that shift to higher frequency with contrast; and (iii) a low-dimensional inter-areal communication subspace that is lower-dimensional than the within-area subspace. They derive parallel predictions of what happens by changing model parameters: feedback gain enhances inter-areal and suppresses within-area communication, and normalization is necessary for both the oscillatory dynamics and the reduced subspace dimensionality. A three-area extension (V1&V4, V1&V5/MT) is used to argue that differential top-down feedback can dynamically route functional connectivity.

We thank the reviewer for this summary.

Strengths:

(1) Analytical tractability: Deriving power spectra, coherence, and communication-subspace structure in closed form from a known fixed point is genuinely valuable.

(2) Conceptual unification: Framing coherence and communication subspaces as arising from the same normalization-driven dynamics is an elegant and useful contribution.

(3) Breadth from few assumptions: A large range of phenomena (contrast gain, gamma dynamics) emerges from normalization-based model assumptions.

(4) Biological grounding: The mapping of model variables onto identified cell types connects the abstract computation to known cortical microcircuitry.

(5) The prediction that input-gain versus feedback-gain modulation produce distinct spectral signatures gives experimentalists a clear way to test the framework.

We thank the reviewer for the positive assessment.

Weaknesses:

(1) Comparisons are qualitative, not quantitative: The theory/experiment panels are visual side-by-side comparisons. There is no quantitative goodness-of-fit for any predictions.

We thank the reviewer for this comment. We agree, and have added a quantitative comparison. New SI Fig. S7 and SI Sec. III fit the model parameters to the experimental power and coherence spectra and report the residuals at every contrast.

(2) The simulations use τ ≈ 1ms for all cell types, which the authors acknowledge is unrealistically short; realistic values would shift the gamma peaks to lower frequencies.

We thank the reviewer for this comment. We agree. The Limitations (Discussion) state that more realistic values would shift the gamma-band oscillations to lower frequencies, and that when all time constants were released as free parameters in the fit to the experimental spectra (SI Fig. S7) they converged to values between 0.2 and 3ms. The time constants are now described consistently as effective population time constants rather than as intrinsic single-neuron time constants (Limitations).

(3) Divisive normalization is a special case and is recovered exactly only for the identity recurrent matrix (self-normalization). Some statements that the circuit implements divisive normalization exactly need softening.

We thank the reviewer for this comment. We agree, and have softened the statements throughout. The main text now states that exact recovery of Eq. 2 requires W₁₁ = I and β = γ = 1, that for the centre-surround W11 used in all reported results the circuit closely approximates normalization (to within 0.6%), and that the fixed point is obtained numerically, at first mention (Results, “Theory”), in the Methods (“The Model” and the new subsection “Numerical computation of fixed points”) and in the Discussion (“Normalization”).

(4) The element-wise (multiplicative) interaction in the modulator dynamics is not tied to a specific cellular mechanism.

We thank the reviewer for this comment. We agree: The Limitations (Discussion) state that, while weighted sums and gain control may be implemented via balanced excitatory–inhibitory mechanisms, we can only speculate about the cellular mechanisms for the element-wise products. We speculate that the element-wise product between presynaptic firing rates in Eq. 1b might be mediated by coincident spike detection (assuming that each of the spike trains is Poisson and that the synaptic current is proportional to the probability of coincident spikes within a short time window (Srinivasan & Bernard, 1976)), and that the element-wise product between postsynaptic membrane potential and presynaptic firing rate in Eq. 1c might be mediated by NMDA receptors (Polsky, Mel & Schiller, 2004; Silver, 2010) (see Discussion, last paragraph of “Biological Plausibility”).

Reviewer #3 (Public review):

Summary:

The work of Pal and colleagues considers a hierarchical and multi-population version of the “oscillatory recurrent gated neural integrator circuits” (ORGaNICs) model, showing through analytics that the model captures multiple relevant experimental results: first of all, its oscillatory dynamics produce a profile with high resemblance to experimental results, both in terms of decay of power at high frequency and in terms of shifting peak as a function of stimulus contrast. Second, inter-areal communication subspace dimensionality is lower than within-area dimensionality. The authors then proceed to further characterize the model’s response properties as a function of input and feedback gain. In particular, they find that frequencies transmitted with higher strength also carry more information, that changing gain modifies the dimensionality of communication subspaces, and that these properties can be used in a three-layer model, where an upstream area can select which downstream area to communicate to, based on the strength of feedback gain.

We thank the reviewer for the summary.

Strengths:

This work demonstrates that a single-circuit model with normalization properties can capture both the oscillatory dynamics and the inter-areal communication properties measured in cortical circuits, matching multiple experimental results. The full analytical tractability of the model is highly advantageous, allowing for easier exploration of parameters, replicability, and effective interpretations of results compared to purely numerical approaches.

The work also makes a useful conceptual link between normalization, coherence-based communication, and subspace-based communication. In particular, it shows how both phenomena can emerge from the same circuit dynamics, where normalization is a key factor.

Interestingly, the model is also extended to multiple areas, showing how attention (in the form of changes in feedback gain) can synchronize the activity of a downstream area with one of two upstream areas, thus effectively selecting which area to communicate with.

In general, this is an interesting computational framework and a useful starting point for future modeling work. A particular strength is that it connects normalization, oscillatory dynamics, coherence, and communication subspaces within one analytically tractable model, making it possible to generate mechanistic hypotheses about when inter-areal communication should be stronger, lower-dimensional, or preferentially routed through feedback.

We thank the reviewer for the positive assessment.

Weaknesses:

Although I see the analytic approach as a strength, at the same time I regard the lack of any numerical comparison as a big weakness. Circuit simulations would not only confirm the correctness of the analytics, but also offer further insights on the error margins and on the regimes where the analytics are valid. This is because, to my understanding, the analytics are based on a linear approximation around the operating regime, which means deviations might be expected, especially for high gain levels in the input, or in the feedforward and feedback pathways.

We thank the reviewer for this comment. We agree that a numerical check of the linearization is required, and have added one. New SI Fig. S6 and SI Sec. II integrate the full nonlinear stochastic system (SI Eqs. S31–S44, with the firing-rate noise and the filtered synaptic noise of SI Sec. VI) with the Euler–Maruyama scheme (time step 10 µs; 100 independent trials of 16s at each contrast) and compare the V1 power spectra and V1–V2 coherence spectra estimated from the simulated trajectories (Welch’s method) with the closed-form expressions at all eight contrasts of Fig. 2. At the baseline parameters, the root-mean-square deviation is below 0.017 for the normalised power and below 0.018 for the coherence, and the maximum absolute deviation is below 0.072, at every contrast; residuals are shown in the lower panels of Fig. S6. The Methods (“Power and coherence spectra”) now describe this validation. On the regime of validity: the linearization holds for small fluctuations about the fixed point, as stated in the Methods, and the simulations bear this out. When the amplitudes of the noise sources are increased, the deviation between the simulated and the analytical spectra grows, and it does so first at the lowest contrasts. This is because at low contrast < 10% the activities lie near zero, and the noise fluctuations can drive them negative, where the rectification in the model clips them to zero; the linearized dynamics do not capture this clipping. Note that changing the strength of the noise does not alter the analytical results, since the plotted spectra are all normalized. The deviations do not, however, grow with the gains: changing the input or feedback gain changes the fixed point, and the linearization about the new fixed point remains valid, because its accuracy is controlled by the strength of the fluctuations relative to the fixed point, not by the gains themselves. This characterization of the validity regime is now stated in SI Sec. II. All results reported in the paper use the baseline parameters, for which the linearization is accurate at every contrast and for all reported gain values. The stable ranges of the gains are given in SI Table S2.

Another problem is that the analytically tractable model seems to rely on effective connectivity weights that break Dale’s law. Numerical simulations with explicitly modeled excitatory and inhibitory units might give insights into effects due, e.g., to the additional transmission delays mentioned in the Discussion.

We thank the reviewer for this comment. We agree, and have made the distinction between effective and anatomical connectivity explicit. The Results (“Theory”) and the Methods (“The Model”) now state, where the matrices are introduced, that the recurrent, feedforward and feedback matrices are effective (net) connectivity, with negative entries standing for disynaptic inhibition through interneurons that are not modelled explicitly. The Limitations (Discussion) now state that the circuit as written does not respect Dale’s law at the level of individual connections, that the posited inhibitory interneurons would introduce additional delays, and that examining how inter-areal synchronization depends on feedback delay is an important direction for future work.

Another weakness is the use of the term “predictions” to indicate features of the model dynamics that are purely described in the context of the model parameters. Although the model’s response properties may certainly lead to predictions, I think the term requires a better contextualization in terms of neurophysiology and experimental neuroscience. The Discussion draws very interesting and valuable bridges between neuron morphology, interneuron types, and model parameters. But it seems it’s left to the reader to backtrack and figure out which biological mechanisms or experimental manipulations should correspond to changes in input or feedback gain, and how these should be distinguished from possible changes in feedforward gain.

We thank the reviewer for this comment. Each prediction is now grounded in a candidate experimental manipulation and a measurable readout in a new Discussion subsection, “Experimental predictions of the model”, which covers each gain parameter, the frequency-resolved subspace, the routing result and normalization, and states explicitly that the proposed manipulations of the gains are speculative and rest on the hypothesized cell-type mapping. The V2 input gain is addressed in our response to the next point.

Relatedly, the manuscript places substantial emphasis on modulation of feedback gain, but does not comparably explore modulation of the feedforward gain, β₂, which regulates the V1-to-V2 drive.

This seems important because changes in feedforward gain could also influence communication subspace dimensionality and oscillatory dynamics. Therefore, predictions related to top-down feedback modulations should be taken with a grain of salt.

We thank the reviewer for this comment. We agree, and have added the V2 input gain to every analysis in which feedback and V1 input gain are varied. New SI Fig. S8 varies β₂ over the same range as β₁ (0.5–1.5) and reports, side by side with feedback and V1 input gain, the contrast-response functions of V1 and V2, the V1 power spectra and V1–V2 coherence spectra at 3% and 50% contrast, and the inter- and within-area prediction performance at 100% contrast. V2 input gain has a signature distinct from both feedback and V1 input gain:

- Increasing β₂ leaves V1 responses unchanged and enhances only V2 responses (Fig. S8c), whereas feedback and V1 input gain enhance both areas.

- At 3% contrast, β₂ produces neither the alpha peak in V1 power nor the beta-band coherence peak that appear at low feedback gain (Fig. S8f,l).

- At 50% contrast, increasing β₂ shifts the gamma peak in power and coherence to higher frequencies and broadens it, like feedback gain, but also reduces its amplitude (Fig. S8i,o).

- Increasing β₂ enhances inter-areal communication, like feedback gain, but leaves within-area communication approximately unchanged (Fig. S8r), whereas feedback gain reduces within-area communication and V1 input gain reduces both.

These results are reported in the Results (“Contrast response functions”, “Inter-areal coherence”, “Communication subspaces”; the “V1 power spectra” section refers the reader to SI Fig. S8 for the V2 input gain), in the Discussion (opening summary, “Role of feedback” and “Experimental predictions of the model”).

Last but not least, the model dynamics are split among multiple elements and nonlinear interactions, reaching a level of complexity far higher than the other ORGaNICs formulations present in the literature. The authors derive these dynamics in the supplementary material, as a dynamical system that converges to a fixed-point solution that includes “exact divisive normalization”. I wonder, however, if there could be simpler solutions that also produce normalization, either approximate or in a different form than the one proposed by the authors. Note also that the designation of “excitatory neurons” is misleading: despite the presence of two explicitly inhibitory populations, the “excitatory” units also interact with negative effective weights both recurrently and in the inter-areal interactions, thus breaking Dale’s law.

We thank the reviewer for the comment. The complexity is, moreover, not unstructured: as now stated in the Results (“Theory”), each population performs a defined operation — principal cells y sum the gain-modulated input, recurrent and feedback drives, u accumulates the normalization signal, a implements the division, and q provides the sign inversion required by the feedback term in the dynamics of a, and it is through these populations that normalization is implemented dynamically, and simultaneously across the reciprocally coupled areas, rather than imposed as a static input–output relation. Whether a materially simpler circuit could produce normalization, approximately or in a different form, while retaining analytical tractability is an open question that we do not resolve here; we would regard such a circuit as a valuable simplification of the present framework. On terminology: we agree, and have changed it. The y population is now called “principal cells” (or “principal neurons”) throughout the main text, figure captions and SI; “excitatory” is reserved for the anatomically excitatory populations in the hypothesised cell-type mapping (Discussion, “Biological plausibility”); the distinction between effective and anatomical weights is stated where the matrices are introduced (Results, “Theory”); and the Limitations state that the principal cells interact through mixed-sign effective weights both recurrently and across areas.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

First of all, my sincere apologies for my long delay. It has been quite a hectic period.

If I could suggest one improvement, it would be to add an early conceptual figure and/or minimal toy model that illustrates the core mechanism before introducing the full circuit. For the illustration, I would try to get inspiration from illustrations of the original Semedo et al paper. Such a figure could show how divisive normalization shapes the covariance/cross-power structure, how this affects coherence, and how the resulting eigen spectrum gives rise to a low-dimensional communication subspace. Starting from this stripped-down mechanism would make the novel insight more transparent, after which the more biologically motivated model would be easier to appreciate.

We thank the reviewer for this suggestion. We agree that the core mechanism should be stated before the full circuit is introduced, and have done so in the text rather than with a new figure. A new paragraph, “The core mechanism in brief”, closes the “Theory” section (Results), before any result is presented. It follows the progression the reviewer outlines: the noise-driven fluctuations about the normalized fixed point are characterized by the cross-spectral density matrix (Eq. 6); the power spectra are its diagonal entries, the coherence its normalised off-diagonal entries (Eq. 7), and the communication subspace follows from its frequency-resolved covariance through reduced rank regression, so that power, coherence and the communication subspace are complementary readouts of the same underlying dynamics. How normalization shapes this matrix, by creating a resonance and by suppressing weakly driven neurons, which lowers its rank, is stated in the Discussion (“Unified view of inter-areal communication”), and removing normalization removes all three effects (SI Fig. S4e–h). The full circuit is then introduced as the implementation of this mechanism. We have not added a schematic figure, because the corresponding panels already exist: the with/without-normalization comparison (SI Figs. S4 and S5).

Minor comments/suggestions:

The authors forgot to cite other work modelling communication subspaces. Citing them could improve impact. Disclaimer: I am among the authors of one of them.

https://www.nature.com/articles/s41467-023-42519-5

https://www.science.org/doi/full/10.1126/sciadv.adl6120

https://www.pnas.org/doi/10.1073/pnas.2404039122

https://www.biorxiv.org/content/10.64898/2026.05.18.725829v1.abstract

We thank the reviewer for these references. We have added all four (Barbosa et al. 2023; Gozel & Doiron 2024; Clark & Beiran 2025; Mastrogiuseppe et al. 2026) in the Introduction, where communication subspaces are introduced.

It could improve the impact if the authors could flesh out how they would test ‘experimentally testable predictions’ (e.g., lines 162, 329).

We thank the reviewer for this suggestion. We agree, and have revised accordingly. Each prediction is now paired with a candidate manipulation and a readout in a new Discussion subsection, “Experimental predictions of the model”.

Figure 2 c-d and e-f could make the message stronger if a correlation between model and data is shown: how good is the prediction?

We thank the reviewer for this suggestion. We agree that a quantitative measure of agreement is needed, and have provided it as a fit rather than a correlation coefficient. New SI Fig. S7 and SI Sec. III fit the model to the experimental power and coherence spectra of Roberts et al. (2013) at all measured contrasts and show the residuals. The Fig. 2 caption and the Results (“V1 power spectra”, “Inter-areal coherence”) now cite this comparison.

Figure 4c is very interesting and deserves more spotlight (could even be tested with available data?)

We thank the reviewer for this suggestion. We agree, and have given this result more prominence and specified the test. It is stated in the Abstract in its own right. The Methods (“Frequency decomposition of the communication subspace”) now end with a paragraph that states how the analysis may be applied to experimental recordings: from the cross-spectral density matrix of simultaneously recorded V1 and V2 populations, its real part at each frequency is the frequency-specific covariance, whose source, target and cross blocks give the communication subspace, the prediction performance and the dimensionality at that frequency in closed form (Eqs. 20 and 22); no band-pass filtering or per-band regression is needed. The resulting test, that the dimensionality should be smallest in the band of peak V1–V2 coherence and that this band should shift to higher frequency with contrast, is stated at the end of that Methods paragraph and in the Discussion (“Experimental predictions of the model”). We have not performed this analysis, which requires the raw simultaneous recordings.

Reviewer #2 (Recommendations for the authors):

Major:

(1) If possible, try to add at least one quantitative model-data comparison (variance explained) even without fitting if needed, so readers can gauge agreement beyond visual inspection.

We thank the reviewer for this comment. We agree, and have added a quantitative comparison. New SI Fig. S7 and SI Sec. III fit the model parameters to the experimental power and coherence spectra measurements and report the residuals at every contrast.

(2) Show how the gamma-peak predictions change under physiologically realistic time constants (e.g., 10-20 ms), so that readers get an idea.

We thank the reviewer for this suggestion. The Limitations subsection of the Discussion now explains that if every time constant of the model is multiplied by a common factor k then every spectral feature would be shifted to a frequency k times lower with its shape preserved. The Limitations also now state that when all time constants were released as free parameters in the fit to the experimental spectra (SI Fig. S7) they converged to values between 0.2 and 3ms. The uniform 1ms values are now described throughout as effective population time constants. We decided not to add a figure because it would look identical to Figs. 2 and 3 with a relabelled frequency axis.

(3) For the communication-subspace results, report how dimensionality estimates scale with the number of modeled/subsampled neurons, so the comparison to the experimental literature is calibrated.

We thank the reviewer for this suggestion. We agree, and have added the analysis as new SI Fig. S9. Fig. S9 reports the dimensionality of the inter-areal (V1–V2) and within-area (V1–V1) communication subspaces (i) as a function of the number of modeled principal neurons per area, N = 9, 18, 36, 72 and 144, at the sampling fraction of the main text (30/72 of the neurons as source and as target, i.e. 4–60 neurons; the connectivity kernels and normalization pool are defined on the ring and scale with N, with no parameter re-tuned), and (ii) as a function of the number of sampled source and target neurons, 5–30, at N = 72. Each point is the mean ± SD over 200 random draws of source and target populations at 100% contrast and baseline parameters, with dimensionality defined as the rank at which prediction performance reaches 95% of its full-rank value. The dimensionality grows slowly with both quantities (roughly two-fold over the 16-fold change in N), remains far below the number of sampled neurons, and the inter-areal subspace is lower-dimensional than the within-area subspace at every population size.

Minor:

(1) Qualify quantitative-sounding language to reflect that comparisons are qualitative and unfitted.

We thank the reviewer for this comment. We agree, and have revised accordingly. The unfitted comparisons are now described as qualitative: “successfully replicates key experimental observations” now reads “reproduces the main qualitative features” (Results, “V1 power spectra”), the captured trends and observations are described as qualitative in the power-spectra and coherence sections, and “align well with” now reads “qualitatively consistent with the experimental data” (Results, “Communication subspaces”). The fitted comparison of SI Fig. S7 is cited alongside the power-spectra and coherence comparisons.

(2) Clarify in the main text the scope of the “exact normalization” claim

We thank the reviewer for this comment. We agree, and have revised accordingly. At first mention (Results, “Theory”) the main text now states that the implementation is exact for W₁₁ = I with β = γ = 1, that for the centre-surround W₁₁ used in all reported results the circuit closely approximates normalization, and that the fixed point for the parameters used is obtained numerically by integrating the deterministic dynamics to steady state.

Reviewer #3 (Recommendations for the authors):

(1) Abstract: I think the text could be improved in terms of clarity and readability. For example, the (f) for frequency appears only after the 1/f⁴, and similarly for the explanation for “communication”. The term “high-coherence” is not properly defined.

We thank the reviewer for this comment. We agree, and have revised the Abstract. Frequency and its symbol f are introduced together before 1/f⁴ appears; “communication” is defined at first use (“the ability to linearly predict neural activity in one brain area from neural activity in another”); “high-coherence frequencies” is replaced by “the frequencies at which V1–V2 coherence peaks”.

(2) L50 and L56: The term “a priori constraints” seems to refer to two general features of the model. The model, however, is much more constrained in terms of parameters and architecture. So the statement that “these two constraints alone give rise to a rich set of emergent properties” is not really corroborated by the results.

We thank the reviewer for this comment. We agree, and have revised the Introduction. It now states the two assumptions (a hierarchy with feedforward, recurrent and feedback connections; recurrently implemented normalization within each area) and then notes that a concrete model additionally requires choices about the connectivity, the normalization pool, and the noise, which are specified in the Methods and whose influence on the results is examined in SI Fig. S4 and the SI section “Robustness with respect to model parameters”. The sentence “Remarkably, these two constraints alone give rise to...” now reads “Given these choices, and a single set of parameters, the theory gives rise to...”.

(3) L69: “competing hypotheses”, but they were called “complementary hypotheses” in L34.

We thank the reviewer for this comment. We agree, and have made the wording consistent. Communication through coherence and communication subspaces are now described, in the Introduction and in the Discussion (“Unified view of inter-areal communication”), as hypotheses that have been treated as alternatives in the literature and that our results show to be complementary.

(4) L132: “all neurons perform the same computation”; this comment is a bit too generic and dismissive.

We thank the reviewer for this comment. We agree, and have rewritten the sentence as a neutral statement where it first occurred (Results, “Theory”) and removed the duplicate sentence from the Discussion (“Biological plausibility”). In Wilson–Cowan-type excitatory–inhibitory models the same form of rate equation is applied to each population, with different signs and weights, whereas here the four populations perform different roles with structurally different equations.

(5) L134: “unique computations to each cell type” although it is clear that different cell types have different roles in the overall dynamics, it is not clear to me which is the unique computation attributed to each of them.

We thank the reviewer for this comment. The computation performed by each cell type is now stated at that point in the text (Results, “Theory”): principal cells y sum the gain-modulated input, recurrent and feedback drives (squaring output nonlinearity); excitatory modulator cells u accumulate the normalization signal, a weighted sum of principal-cell responses (square-root output nonlinearity); inhibitory modulator cells a implement the division, i.e. the multiplicative gain control of the recurrent and feedback drives; and inhibitory cells q track principal-cell activity and provide the sign inversion required by the feedback term in the dynamics of a (SI Eqs. S21–S22). Additionally, they all have different activation nonlinearities.

(6) L146: “positive and negative weights”: mixed signs for feedforward and feedback weights.

We thank the reviewer for this comment. We agree, and have revised accordingly. The phrase now reads “an embedding matrix of effective weights”, followed by a new sentence stating that all weight matrices describe effective connectivity rather than anatomical synapses and that negative weights may be implemented with disynaptic inhibition through interneurons that are not modelled explicitly, with a cross-reference to the Discussion. The same statement accompanies the definition of the matrices in the Methods (“The Model”).

(7) L160: Again mixed signs for recurrent weights.

We thank the reviewer for this comment. We agree, see the response above.

(8) L189: Normalization is exact, but only for parameters different from those used. I think it should be added there that the fixed point for the parameters in use is found numerically through simulation of the dynamics.

We thank the reviewer for this suggestion. We agree, and have added the statement in those terms. The passage now states that the implementation is exact only for W₁₁ = I with β = γ = 1, that for the centre-surround W₁₁ used in all reported results the circuit closely approximates normalization (deviates from Eq. 2 by less than 0.6%), and that the fixed point for these parameters is obtained numerically by integrating the deterministic dynamics to steady state, after which all spectra and subspaces are computed in closed form by linearising about it. The same statement is made in the Methods, in the new subsection “Numerical computation of fixed points”.

(9) L212: There should be something more explicit to corroborate the “experimentally testable”.

We thank the reviewer for this comment. We agree, and have revised accordingly. Each prediction is now paired with a manipulation and a readout in a new Discussion subsection, “Experimental predictions of the model”, which is cited where the predictions are stated in the Introduction and the Discussion.

(10) L214: I think it’s important to clarify a bit more that even if both areas have input gains, the manipulation is only of the input gain in the upstream area, β₁.

We thank the reviewer for this comment. We agree, and have clarified this at that point in the text (“with the V2 input gain β₂ held at its baseline value”) and in the captions of Fig. 3 and SI Figs. S1–S3, each of which now states that β₂ is held at 1.0. β₂ is varied separately in new SI Fig. S8.

(11) Figure 2: Please consider using a better color scale for the power and the coherence lines. The lower three contrasts are literally indistinguishable. Same for the top two contrasts.

We thank the reviewer for this suggestion. We have retained the palette, for the following reasons. The palette is the “Reds” sequential scheme of matplotlib, derived from the ColorBrewer scheme of the same name (Harrower & Brewer, 2003). The curves encode an ordered variable (contrast), for which a single-hue sequential palette is the recommended choice (Zeileis et al., 2020). This encoding is also colour-vision-deficiency safe: the curves are distinguished by lightness rather than by red-versus-green discrimination, and the ColorBrewer “Reds” scheme is classified as colourblind-safe on the basis of empirical testing with colour-vision-impaired viewers (Olson & Brewer, 1997). The same palette is used in the new SI Figs. S6 and S7 for consistency.

(12) L273: Something that confuses me is how to disambiguate between input gain and input intensity, since the theory considers effects around fixed points. This equivalence is also mentioned in the discussion.

We thank the reviewer for this comment. In the model, the V1 input gain β₁ enters the dynamics only through the product β₁z₁ (Eq. 1a), and the input drive z₁ is proportional to stimulus contrast (Methods). Therefore, V1 input gain and stimulus intensity are mathematically equivalent because they enter the dynamics only through their product (β₁ is scalar). This equivalence would not generally hold for spatially or feature-selective gain modulation, where attention can differentially modulate subsets (β₁ is vector) of neurons relative to the normalization pool (Reynolds & Heeger, 2009).

(13) L375: There seems to be a lot of emphasis on modulation of the feedback gain, but related to L214, I wonder why there has been no manipulation of the feedforward gain, β₂, which might be equally important in influencing the communication subspace dimensionality and in modulating the oscillatory dynamics. In terms of experimental predictions, the two might be very different, so why consider only one and ignore the other? See also statement at L388. To attribute the effect to top-down feedback, one should prove that bottom-up modulation does not produce the same effect.

We thank the reviewer for this comment. We agree with the methodological point in the final sentence, and have acted on it. β₂ is now varied across the same analyses as γ₁ and β₁ (new SI Fig. S8; outcomes listed in our response to Weakness 4 above). Bottom-up modulation does not produce the same effect: increasing β₂ enhances inter-areal communication, as feedback gain does, but leaves within-area communication unchanged (feedback gain reduces it), leaves V1 responses unchanged, produces no low-contrast alpha or beta peaks, and reduces rather than increases the amplitude of the gamma peak. The passage at L375 (Results, “Communication subspaces”) now reports that increasing β₂ enhances inter-areal communication while leaving within-area communication approximately unchanged, with a substantially smaller effect than feedback gain (SI Fig. S8r); the readouts that distinguish the two gains are described in the Discussion (“Experimental predictions of the model”).

(14) Figure 4: See comment for Figure 2.

As for Fig. 2, we have retained the palette; see our response to item (11). The contrast values are given in the legend, the luminance ordering is monotonic in contrast, and the same palette is used across Figs. 2 and 4 and the SI.

(15) L440: I think here it would be more honest to say that the two measures, of coherence and inter-areal communication, are not independent. If two signals are highly correlated at a certain frequency, then the prediction power will also peak for that same frequency. Or would it be possible to have conditions where it would not be the case, showing that the two measures are in fact independent?

We thank the reviewer for this comment. We agree, and now state the dependence explicitly. The passage at L440 (Results, “Frequency decomposition of communication subspaces”), the Discussion (“Unified view of inter-areal communication”) and the Methods (“Interpretation of the results”) now state that coherence (Eq. 7) and frequency-resolved prediction performance (Eq. 22) are functionals of the same cross-spectral density matrix, so that the coincidence of their peaks is a mathematical consequence rather than an independent empirical confirmation. The claim retained is that coherence, prediction performance and the reduction in dimensionality are all generated by normalization; removing it eliminates all three (SI Fig. S4e–h).

(16) L446: This paragraph mentions dimensionality, but only the following paragraph defines it.

We thank the reviewer for this comment. We agree, and have moved the definition (the number of source dimensions required to reach the full-rank prediction performance, dictated by the rank of ) so that it precedes the first use of “dimensionality” (Results, “Frequency decomposition of communication subspaces”); the sentence that previously followed has been removed.

(17) L485: See comment for L375.

We address this together with item (13) and the corresponding point of the public review (Weakness 4). We have added a statement regarding the V2 input gain at the end of the section “Dynamic modulation of functional connectivity through feedback” (Results), and the conclusion of that section now describes the three-area result as a demonstration that differential feedback strength can control functional connectivity.

(18) L575: I find this part a bit too vague. Within the scope of the paper, it is not clear what the network is “processing” and what kind of “balancing” it performs.

We thank the reviewer for this comment. In this passage, “processing” refers to the computation performed by the principal cells: each cell sums its gain-scaled input, recurrent and feedback drives (Eq. 1a). “Balancing” refers to how these drives are combined: the recurrent gain control adjusts them so that the steady-state responses follow the normalization equation (Eq. 2), exactly or approximately depending on the recurrent weights. This is the mechanistic contrast with predictive coding: predictive coding matches top-down predictions against feedforward signals so as to minimize the prediction error, whereas hierarchical ORGaNICs balances the input, recurrent and feedback drives so as to maintain normalization.

(19) L583: Similarly to above, I find these lines very confusing. In the context of this paper, there is no notion of “predictable stimulus”, so I find this part very confusing. Also, the model always produces sustained activity by construction, since it has stable steady-state solutions across a wide range of input patterns.

We thank the reviewer for this comment. The passage refers to a limitation of the predictive coding framework. A constant, continuously presented stimulus is fully predictable, so in predictive coding the forward-propagated activity, which represents the prediction error, should be explained away and decay toward zero; cortical neurons instead maintain sustained responses to constant stimuli, barring adaptation (Burns, Xing & Shapley, 2010). We agree that our model produces sustained responses by construction, because it has a stable fixed point for a wide range of inputs; this behaviour is closer to the observed physiology than the explaining-away account.

(20) L601, 602: I don’t think that not producing mathematically exact normalization is a significant, critical limitation in a model. We can be pretty confident that neural circuits are not, in fact, producing mathematically exact normalization. Or, to say it more constructively, simpler models that produce approximate normalization might actually be closer to the biological reality.

We thank the reviewer for this comment. We agree, and have rewritten the passage (Discussion, “Normalization”). It no longer presents approximate normalization as a critical limitation of prior circuit models; it instead states how the present circuit differs: the normalization equation, with an arbitrary normalization pool, is a fixed point of the recurrent dynamics; the dynamics remain stable even with strong recurrent connections; and the fixed point and the linearization about it yield the power spectra, coherence, and communication subspaces in closed form.

(21) L653: this line states that feedback is essential for network stability. But the paragraph seems to refer to top-down feedback drive, which should not be necessary for dynamical stability in the model.

We thank the reviewer for this comment. The line refers to the feedback drive onto the inhibitory modulator cells (a), which contributes to the local normalization signal; it is this feedback-mediated inhibition, not the top-down drive to the principal cells, that is essential for stability. We have clarified the sentence accordingly (Discussion, “Role of feedback”).

(22) L661: I think this statement is misleading and this part should be stated more cautiously. Although plausible in general, I don’t see these results anywhere in the paper. It would require a more functional connotation of the model, which now has a very homogeneous ring structure.

We thank the reviewer for this comment. We agree, and have stated the claim more cautiously (Discussion, “Role of feedback”): the passage now describes the generative-model regime as a hypothesis (“the network is hypothesized to behave like generative models...”) rather than as a property demonstrated by the present results.

(23) L715, L719: I feel confused by the contrast between spatially global feature-based attention and spatially-selective attention.

We thank the reviewer for this comment. The two forms of attention differ in the domain over which neurons are selected for gain modulation. Feature-based attention is spatially global: attending to a feature boosts the neurons selective for that feature across the entire visual field, regardless of retinotopic location (Saenz, Buracas & Boynton, 2002). Spatially selective attention instead boosts only the neurons whose receptive fields overlap the attended location, regardless of their feature preference. Both are gain modulations that can be implemented in the model through the input gain: applied as a vector rather than a scalar, β₁ raises the gain of a selected subset of neurons relative to the normalization pool (Reynolds & Heeger, 2009), as noted in our response to item (12). On the ring, a β₁ profile over preferred orientation implements feature-selective gain; implementing spatially selective attention in the same way would require adding a spatial (retinotopic) dimension over which β₁ varies, as now noted in the Discussion (“Attention and functional connectivity”). The three-area routing result corresponds to the spatially global case, in which boosting an entire feature-selective area increases its feedback to V1 at all locations.

(24) L729, L730: Here it is important to define exogenous and endogenous attention. If they are defined by the pathways on which they act, then the reasoning becomes circular.

We thank the reviewer for this comment. We agree, and have defined both terms behaviourally at first use, independently of the model (Discussion, “Attention and functional connectivity”): exogenous attention as the fast, involuntary orienting driven by a salient stimulus, and endogenous attention as the slower, voluntary, goal-directed deployment of attention. The mapping onto input gain and feedback gain, respectively, is then stated as a testable hypothesis rather than as a definition, falsifiable through the divergent spectral signatures of the two gains.

(25) Figure 7: I suggest changing the color map into a “diverging” color map centered at zero, so that negative and positive values appear visually distinct and zero has a neutral color. For example, the “seismic” or “PuOr” color maps in matplotlib would be appropriate.

We thank the reviewer for this suggestion. We agree, and have revised Fig. 7 accordingly. It now uses the PuOr colour map, centred at zero with a shared colour bar, so that positive and negative weights in the tuning-curve, recurrent and feedback matrices are visually distinct and zero is neutral.

References cited in this letter

Burns, S.P., Xing, D. & Shapley, R.M. (2010). Comparisons of the dynamics of local field potential and multiunit activity signals in macaque visual cortex. Journal of Neuroscience 30, 13739–13749.

Harrower, M. & Brewer, C.A. (2003). ColorBrewer.org: an online tool for selecting colour schemes for maps. The Cartographic Journal 40, 27–37.

Olson, J.M. & Brewer, C.A. (1997). An evaluation of color selections to accommodate map users with color-vision impairments. Annals of the Association of American Geographers 87, 103–134. Polsky, A., Mel, B.W. & Schiller, J. (2004). Computational subunits in thin dendrites of pyramidal cells. Nature Neuroscience 7, 621–627.

Reynolds, J.H. & Heeger, D.J. (2009). The normalization model of attention. Neuron 61, 168–185.

Saenz, M., Buracas, G.T. & Boynton, G.M. (2002). Global effects of feature-based attention in human visual cortex. Nature Neuroscience 5, 631–632.

Silver, R.A. (2010). Neuronal arithmetic. Nature Reviews Neuroscience 11, 474–489.

Zeileis, A., Fisher, J.C., Hornik, K., Ihaka, R., McWhite, C.D., Murrell, P., Stauffer, R. & Wilke, C.O. (2020). colorspace: a toolbox for manipulating and assessing colors and palettes. Journal of Statistical Software 96, 1–49.

Srinivasan, M.V. & Bernard, G.D. (1976). A proposed mechanism for multiplication of neural signals. Biological Cybernetics 21, 227–236.

eLife Assessment

Using their unique Fish-On-Chips optofluidics platform, the authors make three important findings: the presence of precise coupling between saccades and tail flips can be used to discriminate between turning or gliding behaviours; aversive and appetitive chemosensory cues differentially modulate these behaviours; transformation from cue valence to behaviour is encoded by the pallium. The evidence supporting these findings is solid. The work advances our understanding of the ancient interplay between chemosensation and motor output through the modulation of eye-body coordination.

Reviewer #1 (Public review):

This study was designed to manipulate and analyze the effects of chemosensory cues on visuomotor control. They approach this by analyzing how eye-body coordination and brain wide activity are altered with specific chemosensation in larval zebrafish. After analyzing the dynamics of coupled saccade-tail coordination sequences - directionally linked and typically coupled to body turns - the authors investigated the effects of sensory cues shown to be either aversive or appetitive on freely swimming zebrafish on the eye-body coordination. Aversive chemicals lead to an increase of saccade-tail sequences in both number and dynamics, seemingly facilitating behaviors like escape. Brain-wide imaging led the authors to neurons in the telencephalic pallium as a target to study eye-body coordination. Pallium neuron activity correlated with both aversive chemicals and coupled saccade-tail movements.

Comments on revised version.

The revised manuscript addresses my previous concerns.

Reviewer #3 (Public review):

The manuscript investigates the coupling of saccadic eye-movements (S) with directed tail flips (T), and the remarkable discovery is that tail flips that are preceded by a conjugate sacced (S-T) can be credibly classified as specific "volitional" turns that are distinguished from the standard tail movements that seem to be more of a spontaneous and "impulsive" nature.

They show that 'turning intent', as indicated by a small increase in S, is elevated by aversive odors, while 'gliding intent', as indicated by a decrease in S and an increase in undulation cycles, is elevated by appetitive odors.

This is a very important finding, which is backed up by a thorough behavioral analysis, and the identification of neural populations in the pallium and sub-pallium that clearly distinguish between these kinds of turns is very promising. Here they identify neuronal populations that are preferentially active during - and predictive of - coupled (S-T) versus isolated (T) tail flips.

Especially the fact that S-T turns (but not T turns) can be predicted already by pre-event, ramping, pallial activity is intriguing.

The authors then go on and demonstrate that the frequency of (S-T) turns is modulated in fish exposed to appetitive or aversive odors.<br /> Specifically, they quantify the aversiveness and appetitive-ness of several odors in a free swimming assay, they select a couple of these odors based on their valence and they demonstrate that these odors induce moderate modulation in the frequency of eye saccades (S) and tail flips (T) and (S-T) turns.

The study is rigorous and thorough and the findings are informative and novel.

In important controls they confirm that brain wide imaging can distinguish between appetitive and aversive context, and they show that pallial activation by aversive odors is consistent with neural activity in the rhombencephalon that correlates with turning activity, whereas sub-pallial activation by appetitive odors correlates with rhombencephalic activity related to gliding.

It's all very good.

Comments on revised version.

The authors have addressed all my concerns.

Author response:

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

We were appreciative of the quality of the work presented in this manuscript and had some points that the authors could address in a revision:

(1) The pallium interpretation: The data clearly show that pallial activity encodes chemosensory context and predicts upcoming S-T turns. At the same time, in the absence of perturbations, one cannot conclude that the pallium is necessary for transforming chemosensory input into motor output. To align the text with the current evidence, we recommend reframing the pallium's role as correlative/predictive (and avoiding terms like "key region," "hub," or "transformation" that imply causality). The authors could also add a brief limitations note on causality.

(2) Statistics: There were a few points related to the statistics that needed clarification/tightening. We felt that the animal-level replication was limited and with mixed ages, and the data related to events/neurons were treated as independent observations. It was unclear if the hierarchical structure of the data (events/neurons in different fish) was taken into account, how age differences were handled, and what the exclusion criteria were for 'robustly behaving fish'. We recommend reanalysing the data with fish as the unit of inference (or use clustered/mixed models with fish-level random effects), reporting effect sizes with CIs, and clarifying age (5-6 dpf) handling and any exclusion criteria.

(3) Figure clarity/data reduction: Some of the figures were crowded, obscuring the main message. To make the main effects easy to see, we recommend that the authors aggregate across doses and move detailed, condition-by-condition results to tables/SI. We also recommend that they include per-fish distributions where helpful, standardize axes/legends, and define onset metrics in captions.

Summary of Revision

We thank the reviewers and editors for their overall positive evaluations and encouraging remarks, as well as for their suggestions and constructive criticisms. We have performed additional experiments and analyses, and have substantially revised the manuscript based on the reviewers’ comments and in places where we felt additional clarity was needed. Major additions and changes include:

- We reframed the pallium’s role as correlative/predictive and avoided all the terms (e.g., “key region”, “hub”, or “transformation”) that imply causality throughout the manuscript. In Discussion, we added a brief limitation note on causality, and incorporated five recent independent OB/pallium-related studies to better contextualize our neuronal findings.

- To improve animal-level replication, we performed new partially immobilized behavioral experiments in the microfluidic chip (now n = 8 larvae in total; i.e., increased from 5), and reanalyzed the data using linear mixed-effects models with larva as a random intercept (Figures 2, 3, Supplementary Figures 3, 5). For the presentation, IQRs are used to visualize the spontaneous structure of S-T events (Figures 2d, e, i, j, Supplementary Figures 3e, f), the freely swimming plots (Supplementary Figure 4), and chemical-elicited bout kinematic changes (Figures 3e, f, Supplementary Figures 5g, h), whereas 95% CIs are used to visualize time-dependent chemical-elicited rate changes across larvae (Figures 3b–d) and in regression plots (Figures 3g, h). We also included per-larva distributions (Figures 1e, 2b, c, f, 3b–f, Supplementary Figures 3b, c, 5a–h). These additions strengthen the behavioral findings and their presentation. For neuronal analyses (Figures 4, 5, Supplementary Figures 7–11), statistical tests were performed at the individual-larva level, as explicitly clarified in the corresponding figure legends and Methods.

- We rearranged the figures to highlight findings, simplified analyses and their presentation (Figures 1–3, Supplementary Figures 1, 3–6). We now present the most representative concentration for the valence determination assay in Figure 1c, while leaving the rest in Supplementary Figure 1c. We added a new PCA analysis to present the valence-dependent relationship of temporal proximity and turn magnitude in chemical-evoked S-T events (Figures 3g, h, Supplementary Figure 6). We also added within-trial circular saccade-time shuffle analyses to test whether increased aversive saccade occurrence (Figure 3b) could explain the structured temporal and directional properties of chemical-evoked S-T events (Supplementary Figures 5i, j). In addition, we clarified the handling of larval age and explicitly stated in Methods that there were no additional exclusion criteria (page 19, paragraph 2). We also defined onset metrics and referenced the illustrative schematic (Figure 2a) in captions (Figure 5, Supplementary Figure 10).

Public review:

Reviewer #1 (Public review):

Summary:

This study was designed to manipulate and analyze the effects of chemosensory cues on visuomotor control. They approach this by analyzing how eye-body coordination and brain-wide activity are altered with specific chemosensation in larval zebrafish. After analyzing the dynamics of coupled saccade-tail coordination sequences - directionally linked and typically coupled to body turns - the authors investigated the effects of sensory cues shown to be either aversive or appetitive on freely swimming zebrafish on the eye-body coordination. Aversive chemicals lead to an increase in saccade-tail sequences in both number and dynamics, seemingly facilitating behaviors like escape. Brain-wide imaging led the authors to neurons in the telencephalic pallium as a target to study eye-body coordination. Pallium neuron activity correlated with both aversive chemicals and coupled saccade-tail movements.

Recommendations for improvement are minimal. So much of the data is ultimately tabular, and the figures are an impenetrable wall of datapoints. 1c is an excellent example: three concentrations are presented, but it's rare for the three averages to trend appropriately. The key point, which is that aversive odors are repulsive and attractive odors (sometimes) attractive just gets lost in showing the three concentrations individually; it also makes direct comparisons impossible. There are similar challenges abound in the violin plots in 4e-4h, the error bars on the "fits" in 4i-4m, and so on. We recommend selecting an illustrative subset of data to present to permit interpretation and putting the rest in a supplemental table. (Presenting) less is more (effective).

We are very grateful to the reviewer for their positive assessment and important comments. We have revised the figures to improve clarity and interpretability in line with this suggestion. Specifically, for the valence determination assay, we now present only the most representative one in Figure 1c, while the remaining concentrations are moved to Supplementary Figure 1c. In addition, the new PCA analysis reduced multiple regression plots into summary plots to show the valence-dependent relationship of temporal proximity and turn magnitude in chemical-evoked S-T events (Figures 3g, h, Supplementary Figure 6). For improved presentations, we selected a subset of the most important panels for the behavioral findings (Figures 2, 3, Supplementary Figures 3–6), added the per-larva data points and larva-level summary statistics (Figures 1e, 2b, c, f, 3b–f, Supplementary Figures 3b, c, 5a–h), and adjusted the optical transparency of the error bands (Figures 3b–h, Supplementary Figures 5g, h).

Reviewer #2 (Public review):

Summary:

The manuscript by Sy SKH. et al. on pallium encoded chemosensory impact of eye-body coordination describes how the valence of chemosensory stimuli can affect the coordination of eye saccades with tail flips. They show that aversive valence stimuli can increase both the strength and frequency of tail flips through a pallium-mediated circuit.

Overall, the manuscript is well-written and easy to follow, although the figures are quite dense, the methodology is mostly sound, and the improvement to the fish on chips system is very interesting. The methods description is thorough and welcome, making the experiments clear. The limited number of animals, and the spread between 5 and 6dpf is a concern as most of the statistics seem to have been done on the individual events, and not the number of biological samples.

The initial behavioural experiments are very promising. However, the conclusions surrounding the role of the pallium are a lot more speculative and not supported by the results.

We thank the reviewer for their positive assessment of our manuscript, including the clarity of the writing, methodological description, improvements to the Fish‑on‑Chips system, and very promising behavioral experiments. We also appreciate the important concerns raised about the limited number of animals and the spread between 5 and 6 dpf, as well as the strength of our original phrasing regarding the role of the pallium. As described in the above Summary of Revision, we have substantially addressed these points through additional behavioral experiments and refined statistical analyses, a more cautious framing of our neuronal findings throughout the manuscript and contextualization of our findings in Discussion, and clarifications in the revised manuscript, including the explicit specification of age and statistics in Methods (page 19, paragraph 2). In addition, we now clarify that 5 and 6 dpf larvae were intermixed across all stimulus conditions in a within‑subject design, so that age is not confounded with chemosensory valence and condition effects are estimated within animals rather than across different groups of larvae (Methods, page 18, paragraph 2).

Comments:

(1) The fish on chips 2.0 methods show a lot of promise for future studies of chemosensory stimuli, combined with whole-brain imaging. This will provide new avenues of research for zebrafish neuroscientists.

We thank the reviewer for the positive assessment of the Fish‑on‑Chips 2.0 methods and their potential for future chemosensory whole‑brain imaging studies.

(2) Chemosensory cues would have a very different timing than visual cues; timing is very important for multisensory integration. How do the authors suggest those are integrated? How would they differentiate between an integration of various cues or a different arousal state, as they describe in the introduction?

We thank the reviewer for these important comments. In Discussion, we now explicitly consider how visual and chemosensory cues with different temporal dynamics might be integrated. We propose that visual cues may provide rapid guidance for individual movements, whereas chemosensory input may bias the selection and kinematic expression of broader orientation and gaze-shift strategies, including updating gaze after larger reorientations (page 10, paragraph 4).

We also clarify the alternative explanation of altered arousal. The chemosensory effects were largely confined to the 10-s stimulus window and returned to baseline thereafter, which is more consistent with a rapid cue-locked adjustment rather than a sustained global state change. However, this temporal profile does not exclude a transient arousal-related contribution. We therefore state that our data establish an association between chemosensory context and enhanced eye–body coordination, while distinguishing a specific integration mechanism from transient arousal will require further experiments (page 11, paragraph 1).

(3) Studies have looked at chemosensation in Drosophila, including multisensory integration, which should be discussed by the authors (see the work of Mark Frye, amongst others).

We thank the reviewer for pointing out the importance of prior work on chemosensation and multisensory integration in Drosophila. In Discussion, we now relate our findings to studies showing that visual and olfactory cues jointly shape steering and odor‑tracking in flies, including work from the Frye lab, and highlight that reorientation and active visual sampling are common multisensory strategies in small animals (page 10, paragraph 4).

(4) In the brain imaging methods, there is a mention of robustly behaving larvae. Does that mean that an exclusion criterion was used to select only 5 larvae? If so, this should be stated clearly. The authors also do not mention how they avoid the switch to a passive state that one of the coauthors has observed in closed closed-loop setup. The authors should comment on this point.

We thank the reviewer for this clarification request. The mention of “robustly behaving” larvae was not intended to introduce any additional exclusion criterion beyond the leveling requirement already described in Methods. In the revised manuscript, we now state explicitly that there were no further exclusions of animals beyond this positioning criterion, so that the basis for the final sample size is clear (page 19, paragraph 2). We also clarify in Methods that appetitive and aversive cues were presented in brief (10-s), balanced, pseudo‑random trials, rather than as a continuous aversive exposure (page 18, paragraph 2).

(5) Were the statistics in Figure 2 done with an n of 5, or do they assume that each tail flip and saccade is an independent event? I would imagine the latter would have inflated p-values and should be avoided.

We thank the reviewer for highlighting this important issue. In the original analyses for Figures 2, 3, Supplementary Figures 3, 5, individual tail flips and saccades were treated as independent observations, which could underestimate variability across larvae. In the revised manuscript, we have addressed this in two ways. First, we collected data from three additional larvae and now base our analyses on eight larvae in total. Second, we re‑analyzed these data using linear mixed‑effects models including larva as a random intercept, so that multiple events from the same larva are explicitly modeled as non‑independent. The mixed‑effects analyses yields qualitatively similar conclusions to our original analysis, while avoiding inflated p‑values due to event‑level pseudoreplication. We have updated Methods and legends (Figures 2, 3, Supplementary Figures 3, 5) to describe this model specification and to clarify that larva identity was included as a random effect to account for the non-independence of multiple behavioral events recorded from the same larva.

(7) Page 7: Why do the authors think that the cumulative effect of these minor differences could lead to very different behavioural goals? Especially when comparing to actual startle responses, which are extremely strong and stereotypical. How do their observations compare to the thermosensory navigation of larval zebrafish observed by Martin Haesemeyer, for example, or the work of the RoLi lab?

We thank the reviewer for this important point. While the observed changes in individual motor parameters are modest compared to the large and stereotyped responses seen in startle behaviors, our interpretation is based on the principle that larval zebrafish navigation is governed by the accumulation of small, probabilistic changes in bout statistics rather than single, large motor events. This has been well demonstrated in thermosensory navigation, where subtle biases are sufficient to drive reliable temperature preference over time (Haesemeyer et al.). We have now clarified this comparison in the manuscript (page 7, paragraph 2) to better present how the cumulative effects of these small changes can give rise to distinct behavioral outcomes such as approach or avoidance.

(8) Page 8: Figure 5, I am confused by the y-axis of g, in e and f, the values are capped at 2, whereas in g they go up to 6, with apparently a number of cells whose preference is out of the y-axis limit (especially in Q2). Having the number of cells in each quadrant would also help to assess if indeed there is some preference in the pallium towards Q1.

We thank the reviewer for highlighting this potential source of confusion regarding Figure 4g. The difference in the y-axis caps between Figures 4f, g arises because these panels display related but distinct quantities. Figure 4f shows regional mean valence preference, averaged across ROIs and across larvae, which lies within a relatively narrow range and is therefore naturally confined to approximately −2 to 2. In contrast, Figure 4g plots the valence preference of individual motor- and valence‑encoding ROIs, which spans a broader dynamic range; for this reason, the y-axis in Figure 4g extends up to ±6. We have also made this distinction more explicit in the Figure 4 legend.

To address the concern that a substantial number of cells might lie outside the visible axis limits in Figure 4g, we have now quantified the fraction of ROIs whose motor or valence preference values exceed the plotted range. We find that only 0.47% of ROIs have values beyond the chosen axis limits on either axis. These ROIs are plotted at the respective axis limits (i.e. values are clipped to the shown range) and therefore remain visible in the scatter plot. Thus, the plot is representative of virtually all sensorimotor‑tuned neurons included in this analysis.

In addition, in line with the reviewer’s suggestion, we have clarified the distribution of cells across quadrants by modifying Figure 4g. Specifically, we have added a lower subpanel that shows a grouped bar plot of the proportion of ROIs in each sensorimotor quadrant (I–IV) for pallial versus non‑pallial brain regions. To make the underlying numbers explicit, the corresponding absolute ROI counts for pallial and non‑pallial regions in each quadrant are now reported in the Figure 4 legend. These additions demonstrate that pallial ROIs are strongly enriched in quadrant I relative to other quadrants and to non‑pallial regions, and they directly address the reviewer’s request for the number of cells in each quadrant.

(9) Figure 6: How is the onset of neuronal activity determined compared to the motor stimulus? Looking at Supplementary Figure 8, it is quite unclear how the pallium is different from the OB or subpallium. The label of onset delay is also confusing in this figure.

In our analysis, neuronal activity is aligned to tail flip onset for each S-T or T event (dotted vertical lines in neuronal traces in Figures 5a, b, Supplementary Figures 8a, b, 9–11). At each of the two time points preceding the motor event, we compare the fluorescence signal of the motor‑encoding ROIs to their pre‑event baseline; activity exceeding baseline by ≥ 2.5 S.D. is considered significantly elevated (Z‑test with Benjamini–Hochberg correction; see Methods and Figure 5 legend). In Supplementary Figure 9, pallial motor‑encoding neurons exhibit earlier statistically significant elevations relative to tail flip onset during aversive, cue‑associated S-T events, whereas OB and subpallial neurons primarily encode chemosensory input and valence and do not show the same pattern of early significant activation.

Additionally, we have clarified in the legends of Figure 5 and Supplementary Figure 10 that ‘saccade–tail flip onset delay’ is defined as the time difference between tail flip onset and saccade onset within an S-T event (tail flip onset minus saccade onset), and we indicate how this metric is used to group “more-coupled” versus “less-coupled” S-T events. We note that this “saccade–tail flip onset delay” captures how tightly the eye and body components of S-T events are temporally coupled, and thus serves as a quantitative link between motor coordination (Figures 3g, h, Supplementary Figure 6) and the pallial premotor activation observed during aversive cues (Figure 5a, Supplementary Figure 9).

We hope these clarifications resolve the ambiguity the reviewer noted.

(10) Page 9: I do not think that the small differences observed in the pallium are as clear-cut as the authors make them out to be, or that they provide such strong evidence of their importance. As there are no interventions showing any causality in the presence of these pallium responses and the sensorimotor responses, these could represent different arousal states rather than any integration of sensory information.

We thank the reviewer for this thoughtful comment. We agree that, without targeted perturbations, our pallium signals cannot be interpreted as causal for the sensorimotor responses, and we have accordingly reframed them throughout the manuscript (in Title, Abstract, Results and Discussion) as correlative/predictive rather than causal, with an explicit limitation note in Discussion. In addition, we added five related recent studies in Discussion to contextualize our pallial findings. While the pallial differences are moderate in magnitude, they are quantified at the larval level and consistently show a relative bias toward S-T events together with aversive valence, distinguishing the pallium from other regions. Please see also the response to comment 2 regarding the altered arousal point.

Reviewer #3 (Public review):

The manuscript investigates the coupling of saccadic eye movements (S) with directed tail flips (T). The remarkable discovery is that tail flips that are preceded by a conjugate sacced (S-T) can be credibly classified as specific "volitional" turns that are distinguished from the standard tail movements that seem to be more of a spontaneous and "impulsive" nature.

They show that 'turning intent', as indicated by a small increase in S, is elevated by aversive odors, while 'gliding intent', as indicated by a decrease in S and an increase in undulation cycles, is elevated by appetitive odors.

This is a very important finding, which is backed up by a thorough behavioral analysis, and the identification of neural populations in the pallium and sub-pallium that clearly distinguish between these kinds of turns is very promising. Here they identify neuronal populations that are preferentially active during - and predictive of - coupled (S-T) versus isolated (T) tail flips.

Especially the fact that S-T turns (but not T turns) can be predicted already by pre-event, ramping, pallial activity is intriguing.

The authors then go on and demonstrate that the frequency of (S-T) turns is modulated in fish exposed to appetitive or aversive odors.

Specifically, they quantify the aversiveness and appetitive-ness of several odors in a free swimming assay. They select a couple of these odors based on their valence, and they demonstrate that these odors induce moderate modulation in the frequency of eye saccades (S) and tail flips (T) and (S-T) turns.

The study is rigorous and thorough, and the findings are informative and novel.

In important controls, they confirm that brain-wide imaging can distinguish between appetitive and aversive contexts, and they show that pallial activation by aversive odors is consistent with neural activity in the rhombencephalon that correlates with turning activity, whereas sub-pallial activation by appetitive odors correlates with rhombencephalic activity related to gliding.

Overall, this manuscript is very good.

We are very grateful to the reviewer for their positive and thoughtful assessment of our work. We are pleased that the distinction between saccade‑coupled (S‑T) turns and independent tail flips, their modulation by appetitive versus aversive odors, and the identification of pallial populations predictive of these events are viewed as rigorous, informative, and novel.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Full disclosure: We reviewed the manuscript previously for Cell Reports, and we're happy to see that it is much improved. As such, our comments are pretty minimal.

We thank the reviewer for revisiting the manuscript and for their positive assessment of the revisions. In the present revision, we also performed within-trial saccade-time shuffle analyses to directly examine whether the increased occurrence of aversive saccades could account for the temporal and directional organization of S-T events (Supplementary Figures 5i, j). These analyses showed that altered saccade occurrence alone cannot account for this organization.

Reviewer #2 (Recommendations for the authors):

(1) The introduction would benefit from more discussion of the existing disruption of coordination beyond Parkinson. Especially in terms of the circuits involved, as these have been well-studied, and it would provide a better understanding of the system as studied in the manuscript.

We thank the reviewer for this suggestion. In Introduction, we now place Parkinson’s disease within the broader context of movement disorders that disrupt eye–body coordination and expand the circuit‑level discussion accordingly, emphasizing the shared involvement of cortical eye fields, superior colliculus, basal ganglia, cerebellum, and brainstem locomotor centers (page 2, paragraph 2).

(2) "For example, a sudden surge of noxious chemical signals can trigger fleeing" - in what model and condition, what type of escape? There are a number of options to discuss here.

We thank the reviewer for this helpful suggestion. In response, we have revised the sentence to include a concrete and well-characterized example from teleost fish (page 3, paragraph 1). Specifically, we now refer to the response to injury-released alarm substances (“Schreckstoff”), which reliably elicit rapid escape bouts. This provides a clear example of the model system, stimulus condition (alarm cue), and behavioral output (escape locomotion).

(3) The introduction mostly mentions the methodological advantages of zebrafish, but the authors should discuss the relevance of the model to the research question. Especially with regard to the differences in chemosensory spread underwater and in the air.

We thank the reviewer for this suggestion. We have revised the Introduction to better highlight the biological relevance of zebrafish beyond their technical advantages, emphasizing that larval zebrafish naturally use waterborne chemosensory cues to guide locomotion (page 3, paragraph 2). We also address differences between chemosensory signal propagation in water and air, noting that despite these physical differences, animals in both environments must integrate chemical cues with other sensory inputs to select appropriate behavioral responses. We believe this revision clarifies the relevance of the model to our research question.

(4) The GitHub is missing some of the registration code used for calcium imaging. It would be useful to make sure that all the code mentioned in the manuscript is shared.

We thank the reviewer for this thoughtful comment and for drawing attention to code availability. The calcium-imaging processing pipeline in this study used established, publicly available tools, including NoRMCorre for motion correction, as cited and described in Methods. In addition, we used custom MATLAB scripts to implement and link the affine and non-rigid registration workflow for aligning anatomical stacks to the Z-Brain Atlas. These custom scripts, together with the analysis code, have now been uploaded to the GitHub repository.

(5) Figure 4: In subpanel c, it seems that the blank stimulus has a similar, but opposed, effect to aversive stimuli. Would it be possible to see the full 30s pre-stimulus to judge how variable the baseline is?

We thank the reviewer for this helpful comment. In the analyses, the first 10s of each behavioral recording are discarded to avoid transient responses to the onset of light‑sheet illumination, and we normalize the time courses to the mean pre‑stimulus activity in the 10s immediately preceding stimulus onset. In the extended 40s plot provided in Figure 3b, we additionally display 10s of earlier baseline and, for visualization, normalize the traces to the mean activity in these first two 5s bins, so that baseline fluctuations across blank, appetitive, and aversive conditions can be directly compared. All conditions show modest pre-stimulus fluctuations, but the largest and most sustained increase is observed for the aversive condition and is confined to the stimulus window, whereas blank and appetitive traces remain comparatively close to baseline throughout. Thus, although some baseline variability is present, the strong aversive effect remains stimulus-locked and is not explained by pre-stimulus drift in the blank condition.

Because this extended plot was used only to assess baseline variability in response to the reviewer’s question and does not alter the quantitative analyses or figure panels in the manuscript, we have included it in the rebuttal rather than in the revised manuscript.

(6) Figure 5: I assume the legend is flipped between subpanels a and b? Subpanel d is used to show that the pallium has a preference for S-T events, but the result is not convincing; the motor preference is still below zero, very similar to PO, pTec or even Rh3. Furthermore, I assume the statistics have been done on all the neurons across the 5 fish, not treating them as individual samples.

We thank the reviewer for their important comments. The legend on Figure 4 has been corrected so that the labels for subpanels a and b match the plotted data. Although the pallial motor preference is below zero, indicating a net bias toward independent tail flips, we note that the pallium is nonetheless the least T‑preferring region and shows the strongest relative preference for S-T events when motor preference is compared across all brain regions. We have clarified this relative, contrast-based interpretation in the revised manuscript. Importantly, motor‑preference statistics are computed at the individual-larva level (n = 5), averaging neurons within each region for each larva and then testing across larvae rather than treating individual neurons as independent samples. This approach is now described more clearly in Methods, and the Figure 4 legend.

(6) Supplementary Figure 7: What are the ROI numbers for the different graphs in a and b? It is difficult to judge if the "significant" increase in activation level is due to a low level of neurons or a genuine biological result.

We thank the reviewer for this helpful comment. To address the concern, we have revised the legend of Supplementary Figure 8 to clarify that mean ± S.E.M. values are computed across larvae using per‑larva averages and to report the ROI counts for each preference range. These changes make clear that each curve is supported by on the order of 10²–10³ ROIs pooled across the five larvae, rather than by a very small number of neurons.

Reviewer #3 (Recommendations for the authors):

(1) Shouldn't the ipsilateral tendency relative to each preceding flip across a series of tail flips decay to chance over repeated events? In Figure 2h, the teal curve appears to be flat over twenty seconds. Isn't the laterality index calculated relative to the event at n=0? Are the individual fish laterally biased (for spontaneous flips) over 20 bouts?

We thank the reviewer for this insightful comment. In Figure 2h, the laterality index is computed relative to the tail flip at n = 0, and the sustained ipsilateral tendency (please see also Supplementary Figure 3c) over subsequent bouts reflects that spontaneous locomotion in larval zebrafish is not a memoryless random walk. To clarify this point, we now complement the citation on history‑dependent chaining of turns with an additional reference highlighting stable individual left–right motor asymmetries in larval zebrafish (page 5, paragraph 4). This added citation provides stronger context for our observation that tail flips can remain biased ipsilaterally to the initial event over the bout window we analyze.

(2) Do zebrafish initiate a 'turn bout' by a preceding saccade also in free swimming conditions? Are there any data available with sufficient resolution to extract saccades in freely swimming larvae?

We thank the reviewer for this important question. In the prior free‑swimming assay, we used a large field of view (60 mm × 30 mm) at approximately 0.1 mm per pixel to track multiple larvae and quantify their position and heading. Given this combination of field of view and spatial sampling, the larval eyes occupy only a few pixels, and the small angular displacements associated with saccades are very challenging to resolve with sufficient precision. For this reason, we did not extract eye movements or define turn bouts based on preceding saccades in the free‑swimming condition, and instead focused our saccade analyses on the head‑fixed imaging setup, where magnification and resolution are optimized for quantifying eye rotations.

We have added a brief note in Methods (page 22, paragraph 2) to clarify this limitation.

eLife Assessment

This study presents a valuable contribution to comparative cognitive neuroscience by directly mapping functional homologues of the human multiple-demand network in macaques using a matched spatial maze task. It provides solid evidence for functional correspondence between human and macaque cognitive control networks in dorsomedial frontal cortex and insula. However, given only two animals and no cross-animal convergence in parietal or ventrolateral prefrontal regions, claims that these regions form part of a macaque MD network should be read as provisional. The work will be of interest to researchers studying the evolution of cognitive control and cross-species neuroimaging.

Reviewer #2 (Public review):

Summary:

Mione et al. aim to resolve a long-standing question in comparative neuroscience: whether the macaque brain contains a functional analogue to the distributed human multiple-demand (MD) network. To address this, they employ a direct cross-species fMRI comparison using a multi-step saccadic maze task in humans and a simplified two-step version in macaques. By contrasting goal-directed navigation against a control condition that requires similar motor responses but no strategic planning, the study isolates the neural signatures of cognitive control across species.

Strengths:

The most compelling aspect of this work is its methodological alignment. Previous attempts to compare these systems often relied on comparisons of human BOLD signals and macaque single-unit recordings. By running parallel fMRI protocols, the authors establish a shared measurement basis that allows for a more direct comparison. The resulting activation maps demonstrate conserved network topology in dorsolateral and dorsomedial frontal cortex and insula. Individual-level data are presented alongside a conjunction map across the two animals, with the threshold used to define common activations stated explicitly, and supplementary analyses restricted to correctly completed trials or using a smaller smoothing kernel leave the main features intact. This strengthens confidence in the frontal correspondences that appear robust.

Weaknesses:

A design asymmetry between the human and monkey tasks remains. In humans, 2-, 4- and 6-step problems were intermixed within blocks, whereas macaques solved only two-step trials. Because participants never know when a harder problem will occur, they may stay in a state of proactive cognitive control (Braver, 2012) throughout the block, which could inflate MD activation even during simpler trials and complicate direct comparison with the monkey data. The authors acknowledge this limitation but do not quantify or correct for the effect.

Individual-level consistency for some claimed MD regions is limited. The two monkeys differ considerably in parietal cortex, and a conjunction analysis shows no common activation there even at a permissive threshold of t > 1.5. Strong dorsal/medial parietal and ventrolateral prefrontal activations were each seen in only one animal, and the abstract presents them accordingly. This is an honest presentation, but it leaves the evidence for these regions resting on one subject, so whether they can be regarded as part of a macaque MD network remains uncertain.

In the Discussion, the authors conclude that results from single animals "tell against" genuine species differences in these regions. The implied reasoning appears to be that if macaques truly lacked parietal and ventrolateral prefrontal nodes, neither animal would have shown activation there. Since one monkey did show it, they take this as evidence that the regions are not genuinely different between species, attributing the other animal's absence to limited power or insufficient task demands. This inference is difficult to justify with only two animals and no common activation across them at any meaningful threshold. A single positive result does not establish that a region reliably belongs to the macaque MD network.

Overall, this manuscript provides solid evidence for functional correspondence between human and macaque cognitive control networks in dorsomedial frontal cortex and insula. Individual-level data are presented transparently, with conjunction thresholds stated explicitly and limitations acknowledged throughout. What remains is largely interpretive. Given only two animals and no cross-animal convergence in parietal or ventrolateral prefrontal regions, claims that these form part of a macaque MD network should be read as provisional.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

The "multiple-demand" (MD) system is a well-known finding of human brain imaging and is thought to play a central role in cognitive control. To directly compare the MD system in humans and monkeys, Mione et al. used functional magnetic resonance imaging to measure whole brain activation in a multi-step saccadic maze task. In humans, the authors found a distributed pattern of brain activity close match to the canonical MD network and extending to adjacent regions of dorsal attention and other networks. While there was good correspondence between monkey and human data, differences were also notable in lateral frontal cortex, dorsal parietal cortex, and sensorimotor cortex.

Strengths:

Though previous data hint at a corresponding network in the macaque, there has been no direct comparison to human data. This study provides a direct cross-species comparison with whole-brain data of fMRI, and the findings suggest an extended and strongly interconnected brain network recruited by increased cognitive challenge.

Weaknesses:

In previous human imaging, the MD system is defined by overlapping activation for many kinds of cognitive demand. In the present work, however, the authors used just a single task. Although there is some overlap between putative monkey MD network and canonical MD network identified in human imaging, it should be cautious to link current findings to MD system based on limited task events.

We agree that present conclusions are limited by data from a single task. As we note, however, our human data show that this contrast works well for identifying the usual components of the MD network, suggesting it is a suitable candidate for use in the monkey.

Reviewer #2 (Public review):

Summary:

Mione et al. aim to resolve a long-standing question in comparative neuroscience: whether the macaque brain contains a functional analogue to the distributed human multiple-demand (MD) network. To address this, the authors employ a direct cross-species fMRI comparison using a multi-step saccadic maze task in humans and a simplified two-step version in macaques. By contrasting goal-directed navigation against a control condition that requires similar motor responses but no strategic planning, the study isolates the neural signatures of cognitive control across species.

Strengths:

The most compelling aspect of this work is its methodological alignment. Previous attempts to compare these systems often relied on comparisons of human BOLD signals and macaque single-unit recordings. By running parallel fMRI protocols, the authors establish a shared measurement basis that allows for a more direct comparison. The resulting activation maps demonstrate conserved network topology in dorsolateral and dorsomedial frontal cortex and insula.

Weaknesses:

However, there is concerning inter-individual variability in the macaque data, along with a notable design asymmetry between the human and monkey tasks. In the human experiment, 2-, 4-, and 6-step trials were mixed within the same blocks, whereas macaques performed only 2-step problems. This design difference likely places human participants in a state of sustained proactive cognitive control (Braver, 2012), as they must remain prepared for more demanding trials at any moment. This elevated baseline arousal may potentially inflate MD network activation even during the simpler 2-step trials in humans, making direct comparisons with the macaque data difficult.

The authors acknowledge this concern and note that similarities between species survive despite this procedural difference. However, their own individual-level data raise doubts about how robust these "similarities" really are. A defining feature of the MD system is consistent recruitment across individuals performing the same task. In this study, however, the two monkeys show markedly different activation maps in parietal cortex. The conjunction map (Supplementary Figure 6) does not support the claim that lateral and medial parietal regions are commonly activated across both animals at any meaningful statistical threshold, even when using a permissive threshold of |t| > 1.5, corresponding to approximately p ≈ 0.13 - 0.14 (two-tailed).

This suggests that the group-level parietal activations are largely driven by a single animal, raising questions about whether these regions can reliably be considered as part of the monkey MD network. Notably, in Premereur et al. (2018), the only other direct macaque fMRI study of a cognitively demanding task, their conjunction map between the two animals also failed to show common parietal activation during task switching. Taken together, these findings suggest that either the monkey MD network is not as similar to its human counterpart as claimed, or the monkey version of the maze task was simply not sufficiently challenging to fully engage parietal MD nodes.

Because of the design asymmetry described above, it remains difficult to determine which of these possibilities is correct (a genuine species difference versus insufficient task demands in macaques). The authors' current claims treat the group-level activations, which appear largely driven by one animal, as reliable evidence for homology, but without consistent individual-level support, such conclusions remain tentative.

In summary, the authors demonstrate functional correspondence between human and macaque cognitive control networks in dorsolateral and dorsomedial frontal regions. However, assertions that this extends to lateral and medial parietal cortex are not consistently supported at the individual level. While the overall conclusions are plausible, they would be significantly strengthened by demonstrating consistent activation across both animals in all claimed MD regions, ideally with a properly thresholded conjunction map.

These are all good points, and in the revision, we have placed greater emphasis on differences as well as similarities between animals. We have also pointed out the limitations of the conjunction map. In particular, two strong components of the human MD network (vlPFC and parietal cortex) are each seen in only one of the two animals. As we note, limited task difficulty, long training, and limited statistical power may all contribute to negative results. For these regions we agree that between-animal differences limit conclusions, though provisionally, results for single animals tell against genuine species differences.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

(1) Although the whole-brain data obtained by fMRI can provide a direct comparison between species, a single cognitive task might be insufficient to link the findings with MD system. A cognitively challenging task likely activates multiple regions of the MD system; however, it may also recruit some task-specific regions, which do not belong to the canonical MD network.

Though this is a very reasonable concern, the results suggest little influence on the present findings. In the human, strong MD activation was accompanied by activation of the DA network, perhaps reflecting the spatial content of the task. In the monkey, however, there was little evidence of DA activation, likely reflecting limited statistical power.

(2) Based on the current work, the title seems to be too general.

For the above reason, we think the title is reasonable – acknowledging limitations of data from single animals (see below), the cortical and subcortical distribution of monkey activations indeed suggests correspondence to the human MD network.

Reviewer #2 (Recommendations for the authors):

(1) Please clearly state the statistical threshold used for your conjunction analysis in Supplementary Figure 6. The current display appears to rely on |t| > 0, which is not a meaningful criterion: it effectively includes any voxel with positive activation regardless of significance. Even at a more permissive but still unjustified threshold of t = 1.5 (corresponding to roughly p ≈ 0.13-0.14, two-tailed), lateral and posterior medial parietal regions do not appear in the conjunction map. Without a properly justified threshold, it is difficult to accept the claim that these regions are commonly activated across both animals.

The threshold and limitations of the conjunction map are now described (p. 13), and differences between animals given stronger emphasis.

(2) Given that lateral and medial parietal activations do not survive even a low uncorrected threshold in the conjunction analysis and are therefore largely driven by one animal at the individual level, I recommend explicitly acknowledging this limitation in the Discussion rather than treating these regions as reliable components of the macaque MD network. Alternatively, acquiring data from a third monkey would help determine whether this reflects limited statistical power or genuine inter-individual variability.

We have reworded our conclusions concerning parietal activation throughout (beginning with Abstract), emphasising their dependence on data from a single animal.

(3) The design asymmetry between humans and monkeys makes it difficult to interpret cross-species differences in activation. For parietal and ventral prefrontal regions specifically, the monkey version of the task may simply have been insufficiently demanding to fully engage these MD nodes, which might require higher cognitive load before they are reliably recruited. Alternatively, this could reflect a genuine species difference. I recommend discussing this possibility explicitly when interpreting missing or inconsistent activations in these regions.

These possibilities are now considered in a new paragraph of the Discussion (p. 18).

(4) Please specify exactly how Figure 7 was thresholded and state whether FDR correction was applied separately for cortical and subcortical analyses. The colorbar suggests t > 0 as a starting point, but large empty regions within the slices imply that some higher masking or threshold was applied.

See revised figure legend.

eLife Assessment

This work has much novelty, supportive data, and the potential for changing thinking about melanoma biology with the discovery of MRGPRX4 as a relatively melanoma-specific G protein-coupled receptor. MRGPRX4 is normally restricted to peripheral sensory neurons, but it is aberrantly expressed in human melanomas; melanomas with neural-crest-like/invasive features were most likely affected. When ectopically expressed in murine melanocytes, MRGPRX4 drives melanoma development with metastasis; the mechanism appears to involve ligand-independent receptor signaling. Overall, the text presents this as a novel model of oncogenesis driven by re-expression of a sensory GPCR - this is an important study with compelling data.

Reviewer #1 (Public review):

[Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the comments raised in the previous round of review.]

Summary:

Naina Gour and colleagues provide a detailed observational study in which they demonstrate that MRGPRX4, a human G-protein coupled receptor (GPCR), is expressed exclusively in human melanomas and, when expressed in mouse melanocytes, drives the development of melanomas in mice. These findings provide evidence that MRGPRX4 has the properties of an oncogene, at least in certain cellular environments.

Strengths:

A strength of this work is the nice historical note in which overexpression of MAS1, a GPCR, led to classic studies of transformed fibroblasts in culture and tumors in nude mice. Cloning of MAS1 led to the identification of the MRGPR family of receptors, now known to be key players in neuroimmune and neurosensory phenomena. Here, the story comes full circle with a member of the MRGPR family being linked to a tumor, specifically melanoma. Perhaps the story is not entirely surprising given that the neural crest serves as a precursor for both nerves and melanocytes. But it is nice to see.

Additional strengths include the vast array of tools and techniques employed, from public databases to engineered mice, to establish firmly that MRGPRX4 is expressed in melanomas, although not in every malignant cell.

Reviewer #2 (Public review):

Summary:

This study presents a fundamental new finding - the identification of a sensory-neuron itch receptor, MRGPRX4, as an unexpected melanoma oncogene through a mechanism of lineage-inappropriate expression rather than mutation. The evidence supporting the core observation (tumor-specific upregulation, restriction to invasive transcriptional states, and sufficiency to drive fully penetrant metastatic melanoma in vivo) is compelling, drawing on convergent human genomic datasets and a well-controlled genetic mouse model. However, several of the mechanistic and translational claims - particularly regarding causal drivers of invasion, the immunosuppressive tumor microenvironment, and in vivo pharmacological efficacy - rely on correlative evidence.

Strengths:

The authors propose that MRGPRX4, normally restricted to a subset of peripheral sensory neurons, is aberrantly re-expressed in melanoma rather than through mutational mechanisms, and that this re-expression is sufficient to drive tumorigenesis through basal, ligand-independent GPCR signaling. This is a genuinely novel model for oncogenesis, and the manuscript deploys an impressive range of approaches - bulk and single-cell transcriptomics, spatial transcriptomics, proteomics, phosphoproteomics, and pharmacology - to support it.

The claim that MRGPRX4 is selectively upregulated in melanoma and confined to neural-crest-like/invasive transcriptional states is well supported, with consistent results across multiple independent human scRNA-seq datasets. The claim that ectopic MRGPRX4 is sufficient to drive melanoma is convincingly demonstrated by the fully penetrant, metastatic phenotype in the Tyr-CreER;MRGPRX4-LSL model, with appropriate specificity controls showing that MRGPRX1, MRGPRX2, and MRGPRX3 do not phenocopy this effect.

The claim that MRGPRX4 signals through basal, ligand-independent activity is reasonably well supported by bile-acid quantification showing endogenous ligand concentrations well below the EC50 required for activation.

Overall assessment:

The manuscript's central, most novel claim - that lineage-inappropriate expression of a sensory GPCR is sufficient to drive melanoma - is compellingly supported. The secondary mechanistic and translational claims built around this finding are convincing and consistent with the broader literature but are currently supported by correlative rather than causal or functional evidence.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

Naina Gour and colleagues provide a detailed observational study in which they demonstrate that MRGPRX4, a human G-protein coupled receptor (GPCR), is expressed exclusively in human melanomas and, when expressed in mouse melanocytes, drives the development of melanomas in mice. These findings provide evidence that MRGPRX4 has the properties of an oncogene, at least in certain cellular environments.

Strengths:

A strength of this work is the nice historical note in which overexpression of MAS1, a GPCR, led to classic studies of transformed fibroblasts in culture and tumors in nude mice. Cloning of MAS1 led to the identification of the MRGPR family of receptors, now known to be key players in neuroimmune and neurosensory phenomena. Here, the story comes full circle with a member of the MRGPR family being linked to a tumor, specifically melanoma. Perhaps the story is not entirely surprising given that the neural crest serves as a precursor for both nerves and melanocytes. But it is nice to see.

Additional strengths include the vast array of tools and techniques employed, from public databases to engineered mice, to establish firmly that MRGPRX4 is expressed in melanomas, although not in every malignant cell.

We thank the reviewer for the positive assessment of our work and for noting the arc back to the original MAS1 studies, which we agree makes for a satisfying full-circle story. We would add one nuance: while the shared neural crest origin of melanocytes and nociceptors offers a plausible explanation for why an MRGPR family member could be co-opted in melanoma, we were nonetheless surprised that this property is specific to MRGPRX4. When we tested the other MRGPRX family members, some of which are also known to be expressed in sensory neurons, using the same genetic strategy (MRGPRX1, MRGPRX2, MRGPRX3; Supplementary Fig. 2B), none induced melanoma, despite arising from the same family of receptors. This suggests the oncogenic property is not simply a generic consequence of neural crest ancestry, but reflects a biology specific to MRGPRX4 itself, which we find is one of the more intriguing open questions this work raises.

Weaknesses:

Given the power of the strengths of the data and story, the following comment is only sort of a weakness, as the topic is addressed while being saved for future studies. Specifically, what leads to the expression of MRGPRX4? The authors posit that it is an epigenetic phenomenon, look briefly at methylation, and rather than going down the proverbial rabbit hole of what comes first, have reasonably decided to punt.

We agree with the reviewer that understanding the precise mechanisms through which MRGPRX4 gets activated during oncogenesis is important, and we had noted this as a limitation of current findings. As we mentioned in the discussion, we hypothesize that one or more environmental modulators - UV exposure, inflammatory cues, or an aging skin microenvironment could initiate the epigenetic reprogramming underlying this expression. This will require dedicated studies and is suited for future work.

Another concern is that given what comes across as the initial observation of MRGPRX4 being expressed in melanoma, what do all of the additional studies add?

The expression data in human melanoma are correlative and serve only as the entry point. The subsequent studies establish MRGPRX4 as a causal, druggable driver of melanoma: it is sufficient to induce 100% penetrant metastatic melanoma in vivo, required for melanoma proliferation and invasion, signals through ligand-independent basal PI3K-AKT/MAPK activity, and is pharmacologically targetable.

For the non-cognoscenti, and to make the manuscript more accessible, the abbreviation NC/EMT, which is also inverted to EMT/NC, should be spelled out periodically as neural crest/epithelial-mesenchymal transition.

This is corrected in the revised version.

Please explain how this study came about. Was it a result of someone deciding to look at expression in the GTEx project and compare it to a tumor database?

This was a serendipitous discovery. As MRGPRs can modulate immune cell function, we were driving expression of various human MRGPRX genes in immune lineages in mice. We noticed that mice overexpressing MRGPRX4 developed spontaneous melanoma-like growths on the ear and tail. Upon investigation, we found that while the Cre driver we had used is appropriate for immune cells, it is also expressed at low levels in melanocytes. We hypothesized that our initial "immune-knock-in" strategy likely resulted in low-level knock-in in mouse melanocytes, thereby allowing MRGPRX4 to directly drive melanocytic transformation. This led us to explore MRGPRX4 expression in human malignancies, where we found MRGPRX4 was highly expressed in melanoma. Next, to directly test our hypothesis that MRGPRX4 transforms melanocytes, we overexpressed MRGPRX4 using Tyrosinase-CreER, the standard Cre driver for melanocytes. This led to a fully penetrant spontaneous melanoma phenotype (Fig. 2).

A comment could be made to explain that while NSG and normal mice were used, the former are immunocompromised, and drawing conclusions without specifying these differences is a weakness.

This is a fair point, and we agree the distinction deserves clarification. The two mouse systems in this study serve different purposes and are not interchangeable. The transgenic TyrCreER+; MRGPRX4LSL+/- model, in which melanoma arises spontaneously from endogenous mouse melanocytes, was studied in immunocompetent mice, allowing us to fully characterize the tumor, including the tumor immune microenvironment (Fig. 2L–O), in an intact immune setting. In contrast, the human A2058 melanoma cell studies (proliferation and metastatic seeding; Fig. 4B–J) required an immunodeficient host, since human cells would otherwise be rejected by a competent murine immune system. For these experiments, we used NSG mice, the standard immunodeficient strain used in human xenograft studies. We have updated the text to reflect this.

In Figure 1A, the p-value of -145 begs for a little explanation. I don't recall seeing such a p-value.

The small p-value reflects two features of the comparison: (1) both GTEx normal skin and TCGA-SKCM contain large sample sizes (hundreds of samples per group), which gives the Mann-Whitney U test (also known as the Wilcoxon rank-sum test) statistical power, and (2) MRGPRX4 expression shows minimal overlap between the two groups; it's essentially undetectable in normal skin but broadly expressed across melanoma samples. Together, these yield such a p-value. This is expected for rank-based tests applied to large, cleanly separated datasets.

Have you considered treating the murine melanomas with murine via PD-L1? I appreciate that this comment is somewhat superfluous given the inhibition of MRGPRX4 with compound 31-2, but given the human therapeutics combined with the fact that you have done 'everything else', I wonder what might happen.

We agree with the reviewer that testing checkpoint blockade in the MRGPRX4-driven model is a relevant direction. Our finding that MRGPRX4-driven tumors develop a PD-L1hi, neutrophil-rich microenvironment (Fig. 2M–O) makes anti-PD-L1 treatment a natural next experiment, and we would predict it to be informative both on its own and in combination with an MRGPRX4 inhibitor, given that the two target distinct compartments- the tumor-immune microenvironment versus tumor-intrinsic proliferative/invasive signaling. We consider it an important direction for future work.

Given the basal ligand-independent signaling, might engineering variants of MRGPRX4 that do not signal be of value?

This is a great point. A variant that disrupts basal (ligand-independent) signaling specifically would be informative. Identifying and validating such a basal-activity-disrupting variant, including confirming normal receptor expression and trafficking, is an important direction for future work.

Reviewer #2 (Public review):

Summary:

This study presents a fundamental new finding - the identification of a sensory-neuron itch receptor, MRGPRX4, as an unexpected melanoma oncogene through a mechanism of lineage-inappropriate expression rather than mutation. The evidence supporting the core observation (tumor-specific upregulation, restriction to invasive transcriptional states, and sufficiency to drive fully penetrant metastatic melanoma in vivo) is compelling, drawing on convergent human genomic datasets and a well-controlled genetic mouse model. However, several of the mechanistic and translational claims - particularly regarding causal drivers of invasion, the immunosuppressive tumor microenvironment, and in vivo pharmacological efficacy - remain incomplete, relying on correlative evidence.

Strengths

The authors propose that MRGPRX4, normally restricted to a subset of peripheral sensory neurons, is aberrantly re-expressed in melanoma rather than through mutational mechanisms, and that this re-expression is sufficient to drive tumorigenesis through basal, ligand-independent GPCR signaling. This is a genuinely novel model for oncogenesis, and the manuscript deploys an impressive range of approaches - bulk and single-cell transcriptomics, spatial transcriptomics, proteomics, phosphoproteomics, and pharmacology - to support it.

Strengths:

The claim that MRGPRX4 is selectively upregulated in melanoma and confined to neural-crest-like/invasive transcriptional states is well supported, with consistent results across multiple independent human scRNA-seq datasets. The claim that ectopic MRGPRX4 is sufficient to drive melanoma is convincingly demonstrated by the fully penetrant, metastatic phenotype in the Tyr-CreER; MRGPRX4-LSL model, with appropriate specificity controls showing that MRGPRX1, MRGPRX2, and MRGPRX3 do not phenocopy this effect.

The claim that MRGPRX4 signals through basal, ligand-independent activity is reasonably well supported by bile-acid quantification showing endogenous ligand concentrations well below the EC50 required for activation.

We are thankful to the reviewer for the positive feedback.

Weaknesses:

The claim that MRGPRX4 remodels the tumor microenvironment toward an immunosuppressive state rests on flow cytometric frequency data (altered neutrophil/eosinophil ratios, increased PD-L1+ myeloid populations) but lacks any functional immune assay to demonstrate that this remodeling actually impairs anti-tumor immune responses.

We acknowledge the reviewer’s suggestion that functional immune assays will demonstrate that MRGPRX4-driven tumor remodeling impairs checkpoint-driven anti-tumor immune response. We consider this an important direction for future work.

The claim that the two MRGPRX4-enriched tumor subpopulations (ECM-rich and NC-like/invasive) underlie the observed invasive and metastatic phenotype is not directly tested; the authors appropriately acknowledge this as an open question, but it is worth noting explicitly that this leaves the mechanistic link between the identified cell states and the functional phenotype (proliferation, invasion, metastasis shown in Figure 4) unresolved.

We agree with the reviewer. To test the direct role of these subpopulations in driving metastasis in our model, we need to ablate them specifically. This is possible with an intersectional genetics strategy, wherein we could use a state-specific Dre or Flp driver (e.g., a Prrx1-DreER or Prrx1-FlpO knock-in) crossed to a dual-recombinase-dependent effector allele (e.g., a Frt- or Rox-gated diphtheria toxin receptor), and then crossed to our TyrCreER-MRGPRX4-LSL model. This quadruple-transgenic strategy would allow ablation restricted specifically to MRGPRX4-driven tumor cells occupying the target mesenchymal state, for example. These genetic lines can be created but require generating or sourcing new dual-recombinase-dependent alleles and a substantially longer breeding and validation timeline (approximately 18-24 months). Together, these represent possible, if long-term, strategies for directly testing the causal contribution of these MRGPRX4-enriched subpopulations to melanoma invasion and metastasis.

Finally, the comparison with BRAF- and NRAS-driven GEMMs (Figure 3K-L) establishes overlap in transcriptional cell states but does not report whether these canonical models themselves upregulate endogenous Mrgprx4. This omission leaves unclear whether MRGPRX4 acts as a convergent node downstream of canonical oncogenic signaling, or represents an independent, parallel route to a similar phenotypic endpoint - a distinction that matters considerably for how broadly the finding should be interpreted.

MRGPRX4 is a primate-specific receptor with no mouse ortholog in melanocytes; thus, we cannot assess endogenous expression of MRGPRX4 in BRAF/NRAS-driven GEMMs. Nevertheless, the following observations argue against MRGPRX4 functioning solely downstream of canonical oncogenes. First, TyrCreER⁺; MRGPRX4LSL mice develop fully penetrant melanoma without engineered BRAF or NRAS activation or tumor-suppressor loss. Second, MRGPRX4 loss in BRAF V600E mutant A2058 cells reduces pERK1/2, pAKT, and pS6K, showing that MRGPRX4 sustains these signaling outputs even in the presence of activated BRAF. Together, these findings support a model in which MRGPRX4 could provide a distinct oncogenic input.

Overall assessment:

The manuscript's central, most novel claim - that lineage-inappropriate expression of a sensory GPCR is sufficient to drive melanoma - is compellingly supported. The secondary mechanistic and translational claims built around this finding are convincing and consistent with the broader literature but are currently supported by correlative rather than causal or functional evidence.<br />

Recommendations for the authors:

Reviewing Editor Comments:

Both reviewers found much novelty and potential for changing thinking about melanoma biology with your discovery of MRGPRX4 as a relatively melanoma-specific G protein-coupled receptor. Particularly in neural-crest-like/invasive states, MRGPRX4 was aberrantly re-expressed in human melanomas. With the Tyr-CreER; MRGPRX4-LSL mouse model, there were compelling data in transcriptomics and proteomics. As one reviewer noted, your work closes a circle from early MAS1 (a related GPCR) oncogene studies to the MRGPR family, consistent with the shared neural-crest origin of melanocytes and sensory neurons.

There were several minor suggestions and clarifications suggested by the reviewers. Please address them in a revised manuscript, which we will review without sending it back to the reviewers.

Reviewer #1 (Recommendations for the authors):

There is a number of grammatical errors and spacing issues that could be sorted out via built-in editing tools.

We have checked the manuscript and are resubmitting updated version.

Reviewer #2 (Recommendations for the authors):

Examine whether Mrgprx4 is detectable (even at low levels) in the BRAF^V600E; Pten^fl/fl and NRAS^Q61K; Ink4a−/− datasets used for integration in Figures 3K-L.

Levels of MRGPRX4 in BRAF- and NRAS-driven GEMMs cannot be interrogated, as MRGPRX4 is a primate-specific gene and hence not expressed in mice.

In vivo antagonist efficacy: Given the well-established transgenic model, in vivo testing of compound 31-2 (or a more potent/optimized MRGPRX4 antagonist) on tumor growth or metastatic burden would considerably strengthen the translational claims.

We appreciate the reviewer's recommendation and agree that in vivo testing of an MRGPRX4 antagonist would meaningfully strengthen the translational relevance of our findings. However, we do not believe compound 31-2 is a viable candidate for in vivo studies. While the reported biochemical IC50 on the company's (MedChemExpress) website is ~100 nM, in our hands, using a cell-based assay, we required a substantially higher concentration (5 µM) to block receptor activity. Additionally, using our GPCR (calcium-based) activation assays, we found that the compound is no longer active after approximately 6 hours in solution, necessitating the addition of freshly prepared compound at least twice daily in our in vitro culture system.

Compound 31-2 is a tool compound with no available in vivo pharmacokinetic characterization (bioavailability, clearance, half-life, or protein binding), and translating our in vitro dosing requirement to an in vivo regimen would demand achieving sustained micromolar free-drug exposure, likely requiring impractically high and frequent dosing, or specialized delivery methods (e.g., continuous infusion). We therefore consider compound 31-2 unsuitable for in vivo efficacy testing in its current form. Development or identification of a more potent, metabolically stable, and orally bioavailable MRGPRX4 antagonist with an established in vivo PK profile will be necessary before in vivo antitumor efficacy can be meaningfully evaluated.

eLife Assessment

This important study presents a cell-based screen for small-molecule activators of GCN2, an eIF2α kinase that regulates the Integrated Stress Response under diverse stress conditions. The authors identify a compound as a potent GCN2 activator with GCN1-independent activity under the tested conditions, providing a new pharmacological tool for probing GCN2 regulation. The revised manuscript provides compelling support for the central conclusions through a clearer description of the screening workflow, extended kinetic analyses, and demonstration that the identified compound causes a GCN2-dependent reduction in protein synthesis.

Reviewer #1 (Public review):

Summary:

This manuscript describes a chemical screen for activators of the eIF2 kinase GCN2 (EIF2AK4) in the integrated stress response (ISR). Recently, reported inhibitors of GCN2 and other protein kinases have been shown at certain concentrations to paradoxically activate GCN2. The study uses CHO cells and ISR reporter screens to identify a number of GCN2 activator compounds, including a potent "compound 20." These activators have implications for the development of new therapies for ISR-related diseases. For example, although not directly pursued in this study, these GCN2 activators could be helpful for the treatment of PVOD, which is reported for patients with certain GCN2 loss-of-function mutations. The identified activators are also suggested to engage with the GCN2 directly and can function devoid of GCN1, a co-activator of GCN2.

Strengths:

The manuscript appears to be a largely rigorous study that flows in a logical manner. The topic is interesting and significant.

Weaknesses:

Portions of the manuscript are not fully clear. There are some experimental presentation and design concerns that should be addressed to support the stated conclusions.

Reviewer #2 (Public review):

Summary:

In this manuscript Zhu, Emanuelli and colleagues describe a novel pharmacological activator of the Integrated Stress Response kinase GCN2. The work is conclusive and biochemically solid. This work significantly adds to the pharmacological arsenal targeting the ISR and in particular GCN2.

Strengths:

Strong biochemistry, novel molecular activator of GCN2 (GCN1 independent).

Weaknesses:

Rationale for the screen not exploited in the results (e.g. pathogenic GCN2 mutants), lots of cell-based read-outs not endogenous.

Comments on revised version.

The authors did a great job at addressing my initial critique on their manuscript and consequently I have no further comment.

Reviewer #3 (Public review):

Summary:

In this manuscript, the authors describe the results of a high throughput screen for small molecule activators of GCN2. Ultimately, they find 3 promising compounds. One of these three, compound 20 (C20) is of the most interest both for its potency and specificity. The major new finding is that this molecule appears to activate GCN2 independent of GCN1, which suggests that it works by a potentially novel mechanism. Biochemical analysis suggests that each bind in the ATP binding pocket of GCN2, and that at least in vitro C20 is a potent agonist. Structural modeling provides insight into how the three compounds might dock in the pocket and generates testable hypotheses as to why C20 perhaps acts through a different mechanism than other molecules.

Strengths:

Of the 3 compounds identified by the authors, C20 is of the most interest, not just for its intriguing mechanistic distinction as being GCN1-independent (shown genetically in two distinct cell lines, CHO and 293T, and in contrast to other GCN2 activators) but also for its potency. Ultimately, C20 might be a tool for providing mechanistic insight into the details of GCN2 activation and regulation and could be exploited therapeutically.

Weaknesses:

The chief limitation of this work is that the experiments exploring the effects of C20 on ISR output in cells are limited, so how useful these compounds are both experimentally and therapeutically remains to be determined.

Comments on revised version.

The authors have satisfactorily addressed my comments. A more extensive analysis of UPR signaling in cells (transcription and cell death in particular) would have further strengthened the paper, but that can be left to future work.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

This manuscript describes a chemical screen for activators of the eIF2 kinase GCN2 (EIF2AK4) in the integrated stress response (ISR). Recently, reported inhibitors of GCN2 and other protein kinases have been shown at certain concentrations to paradoxically activate GCN2. The study uses CHO cells and ISR reporter screens to identify a number of GCN2 activator compounds, including a potent "compound 20." These activators have implications for the development of new therapies for ISR-related diseases. For example, although not directly pursued in this study, these GCN2 activators could be helpful for the treatment of PVOD, which is reported for patients with certain GCN2 loss-of-function mutations. The identified activators are also suggested to engage with the GCN2 directly and can function while devoid of GCN1, a co-activator of GCN2.

Strengths:

The manuscript appears to be a largely rigorous study that flows in a logical manner. The topic is interesting and significant.

Weaknesses:

Portions of the manuscript are not fully clear. Some experimental presentation and design concerns should be addressed to support the stated conclusions.

We thank the reviewer for their supportive comments. We agree that portions of the manuscript were not fully clear and that some aspects of the experimental presentation and design required clarification.

To address this, we have revised the manuscript to make the experimental logic more transparent. In particular, we now explain more clearly the rationale for the screening strategy, including the use of histidinol as a canonical GCN2 activator, latrunculin A as a modulator of PPP1R15A-mediated eIF2α dephosphorylation, and tunicamycin as a PERK-dependent ER-stress control. We also clarify why a submaximal concentration of histidinol was used: this was intended to reveal compounds that enhance ISR signalling when GCN2 is partially activated.

We have clarified the use of the two ATF4 reporter systems. The ATF4–NanoLuc reporter was used for sensitive primary screening, whereas the ATF4–luc2 reporter was used as a more stringent orthogonal assay to prioritise robust ISR activators. We now state explicitly why some initial hits were not retained after testing in the second reporter line, and why the NanoLuc system was subsequently used again for mechanistic experiments.

Finally, we have revised the presentation of the orthogonal validation steps to make clearer how they support the stated conclusions. These include assays designed to distinguish GCN2-dependent ISR activation from indirect activation through ER stress, additional analysis of GCN2 dependence, and clearer interpretation of biochemical and docking data.

We hope that these revisions address the reviewer’s concern that the experimental design and data presentation needed to be made clearer in order to support the conclusions.

Reviewer #2 (Public review):

Summary:

In this manuscript, Zhu, Emanuelli, and colleagues describe a novel pharmacological activator of the Integrated Stress Response kinase GCN2. The work is conclusive and biochemically solid. This work significantly adds to the pharmacological arsenal targeting the ISR and, in particular, GCN2.

Strengths:

Strong biochemistry, novel molecular activator of GCN2 (GCN1 independent).

Weaknesses:

The rationale for the screen is not exploited in the results (e.g., pathogenic GCN2 mutants), and lots of cell-based read-outs are not endogenous.

We thank this reviewer for their positive assessment of the work. We address the three major concerns in turn below.

Major points

(1) Regarding the justification of the work. Since the authors justify the screen for GCN2 activators with loss-of-function mutants associated with diseases, it would be of interest to evaluate whether the best compounds identified in the study are indeed able to prompt activation of those mutants (or at least of the most prevalent). This approach could actually go in parallel with the docking experiments carried out in the last figure of the manuscript, where mutants could be modelized as well.

To address this point, we tested whether the lead compounds could activate disease-associated GCN2 variants linked to pulmonary veno-occlusive disease. In contrast to GCN2iB, the new compounds did not activate these variants. We now state this explicitly in the manuscript, thereby clarifying that although the compounds identify a new mode of GCN2 activation, they do not rescue the pathogenic GCN2 variants tested here.

Results

“Moreover, in contrast to GCN2iB (17), the current compounds did not activate disease-associated GCN2 variants linked to PVOD [data not shown].”

(2) The compounds are only tested using « artificial » proximal signaling outputs. It would be interesting to evaluate whether the best identified compounds are capable of prompting endogenous eIF2alpha phosphorylation in cellular models.

We thank the reviewer for this suggestion. Detecting eIF2α phosphorylation following activation of GCN2 is technically challenging and typically produces weaker signals compared to activation of other ISR kinases, such as PERK (e.g. by thapsigargin). For this reason, many studies rely on downstream reporter assays to monitor GCN2 activity. To address the reviewer’s concern, we have now included an orthogonal readout of ISR activation by assessing global translation using a puromycin incorporation assay. Using this approach, we show that compound 20 significantly reduces translation, and importantly, this effect is attenuated in GCN2-deficient cells, supporting a GCN2-dependent mechanism.

Results

“Studies with compound 18 were limited by poor aqueous solubility; therefore, time‑course analyses focused on compounds 20 and 21. To assess ISR activation over an extended period, live‑cell luciferase measurements were performed using CHO cells stably expressing an ATF4::Nanoluc-PEST reporter. Both compounds elicited maximal reporter activation between 6 and 8 h (Figure S1A&B). Compound 20, but not 21, induced a significant GCN2‑dependent reduction in mRNA translation, as measured by puromycin incorporation, with a progressive effect observed up to 7 h (Figure S1C-F).”

(3) Other GCN2 activators (other than GCN2iB, e.g., HC-7366) were recently identified. In this context, it would be of interest to carry out a small benchmarking study to evaluate how the compounds identified in the current study perform against the previously identified molecules.

We thank the reviewer for this suggestion. In response, we obtained HC-7366 and assessed its activity alongside our compounds in the CHO ATF4::NanoLuc reporter assay. In this system, compound 20 demonstrated greater potency than HC-7366 (see reviewer figure below). However, we note that HC-7366 showed relatively limited activity in CHO cells in our hands, despite previously reported strong effects in other cellular systems and in vivo models. This context-dependent activity makes direct benchmarking difficult. Accordingly, we have included this comparison in the revised manuscript and discuss this limitation in the Discussion.

Discussion

“We also evaluated the reported GCN2 activator HC-7366 in our CHO ATF4::NanoLuc reporter system. In this context, HC-7366 showed limited activity relative to compound 20, despite its reported efficacy in other cellular systems and in vivo (data not shown). This highlights potential context dependence in small‑molecule activation of GCN2 and limits direct cross-study comparison.”

Author response image 1.

ISR activation by compound 20 and GC-7366 in CHO cells

Normalised fold-change in ATF4 signal in CHO ATF4::NanoLuc reporter cells treated for 19 hours with Compound 20 or HC-7366. DMSO was used as vehicle control. (representative experiment, mean ± SEM, n=3 technical replicates).

Reviewer #3 (Public review):

Summary:

In this manuscript, the authors describe the results of a high-throughput screen for small-molecule activators of GCN2. Ultimately, they find 3 promising compounds. One of these three, compound 20 (C20), is of the most interest both for its potency and specificity. The major new finding is that this molecule appears to activate GCN2 independent of GCN1, which suggests that it works by a potentially novel mechanism. Biochemical analysis suggests that each binds in the ATP-binding pocket of GCN2, and that at least in vitro, C20 is a potent agonist. Structural modeling provides insight into how the three compounds might dock in the pocket and generates testable hypotheses as to why C20 perhaps acts through a different mechanism than other molecules.

We agree that GCN1-independent activation suggests a potentially distinct mechanism of action. While we are currently unable to define the mechanistic basis underlying the GCN1-independence of compound 20, prior work provides some relevant context. Recent studies have shown that the ATP-competitive modulator GCN2iB can activate GCN2 independently of GCN1 under specific conditions, notably in the context of the GCN2 E26A mutant [Carlson, 2023]. This observation raises the possibility that, under certain conditions, engagement of the kinase domain, potentially via the ATP-binding pocket, may bypass the requirement for GCN1. However, in our system, we did not observe GCN1-independent activation with GCN2iB at the concentrations tested. This discrepancy may reflect a narrow or context-dependent window for such activity, or differences between wild-type and mutant GCN2. These findings suggest that GCN1-independent activation of GCN2 may occur under specific conditions or with distinct classes of compounds, although further work will be required to define the underlying mechanism for compound 20. We have added the following to the main text:

Discussion

“Recent work suggests that the ATP-competitive modulator GCN2iB can activate GCN2 independently of GCN1 under specific conditions using a GCN2 E26A mutant (9). In our hands, we did not observe GCN1‑independent activation with GCN2iB at the concentrations tested. This discrepancy may reflect a narrow concentration window for GCN1‑independent activation or context‑dependent effects of the E26A mutation. These findings raise the possibility that GCN1‑independent activation of GCN2 may occur under specific conditions or with distinct classes of compounds.”

Strengths:

Of the 3 compounds identified by the authors, C20 is the most interesting, not just for its intriguing mechanistic distinction as being GCN1-independent (shown genetically in two distinct cell lines, CHO and 293T in Figure 4, and in contrast to other GCN2 activators) but also for its potency. In in-cellulo assays, compound 21 appears as more of an ISR enhancer than an activator per se, and although compound 18 and compound 21 lead to upregulation of the ISR targets (Figure 2), that degree of upregulation is probably not significantly different from that induced by those compounds in Gcn2-/- cells. For C20, the effect appears stronger (although it is unclear whether the authors performed statistical analysis comparing the two genotypes in Figure 2D). In Figure 3, only C20 activates the ISR robustly in both CHO and 293T. Ultimately, C20 might be a tool for providing mechanistic insight into the details of GCN2 activation and regulation, and could be exploited therapeutically.

Prompted by this suggestion, we assessed C20 in two additional commonly used cell lines: human colon carcinoma HCT116 cells and African green monkey COS7 cells. C20 showed no activity in these models. In contrast, primary mesothelioma cells (Mesobank T12) exhibited robust PPP1R15A induction in response to the compound. The following text has been added to the manuscript.

Results

“We went on to examine downstream cellular consequences of GCN2 activation in multiple models. While compounds did not induce detectable ISR signalling in HCT116 or COS‑7 cells under the conditions tested, induction of PPP1R15A was observed in Mesobank T12 primary mesothelioma cells, indicating context-dependent biological responses… [data not shown].”

Weaknesses:

There are some limitations to the existing work. As the authors acknowledge, they do not use any of the compounds in animals; their in vivo efficacy, toxicity, and pharmacokinetics are unknown. But even in the context of the in cellulo experiments, it is puzzling that none of the three compounds, including C20, has any effects in HeLa cells when Neratinib does. It's beyond the scope of this paper to address definitively why that is, but it would at least be reassuring to know that C20 activates the ISR in a wider range of cells, including ideally some primary, non-immortalized cells. In addition, the ISR is a complex, feedback-regulated response whose output varies depending on the time point examined. The in cellulo analysis in this paper is limited to reporter assays at 18 hours and qRT-PCR assays at 4 and 8 hours. A more extensive examination of the behaviour of the relevant ISR mRNAs and proteins (eIF2, ATF4, CHOP, cell viability, etc.) for C20 across a more extensive time course would give the reader a clearer sense of how this molecule affects ISR output.

We thank the reviewer for this insightful suggestion. To address the need for a more comprehensive assessment of ISR signalling, we have extended our analysis across a broader time course and incorporated additional functional readouts. Using the ATF4-NanoLuc reporter, compounds 20 and 21 exhibit peak ISR activation at approximately 6-8 h in wild-type cells. In parallel, we assessed global mRNA translation using puromycin incorporation and found that compound 20, but not compound 21, induces a progressive reduction in translation over this period, which is dependent on GCN2. While we agree that direct measurement of upstream ISR markers such as eIF2α phosphorylation can be informative, detection of GCN2-mediated eIF2α phosphorylation is technically challenging and often less robust than activation of other ISR kinases (e.g. PERK). For this reason, we have prioritised orthogonal downstream functional readouts, including reporter activity and translational output, to capture ISR pathway engagement. These additional data provide a clearer picture of the kinetics and functional consequences of compound-induced ISR activation and have been incorporated into the revised manuscript.

Results

“Studies with compound 18 were limited by poor aqueous solubility; therefore, time‑course analyses focused on compounds 20 and 21. To assess ISR activation over an extended period, live‑cell luciferase measurements were performed using CHO cells stably expressing an ATF4::Nanoluc-PEST reporter. Both compounds elicited maximal reporter activation between 6 and 8 h (Figure S1A&B). Compound 20, but not 21, induced a significant GCN2‑dependent reduction in mRNA translation, as measured by puromycin incorporation, with a progressive effect observed up to 7 h (Figure S1C-F).”

I also find it a bit strange that the authors describe C20 as "demonstrat(ing) weak inhibition of ... PKR" - the measured IC50 is ~4 μM, which is right around its EC50 for GCN2 activation. This raises the confounding possibility that C20 would simultaneously activate GCN2 while inhibiting PKR. While perhaps inhibition of PKR is not relevant under the conditions when GCN2 would be activated either experimentally or therapeutically, examining in cells the effects of C20 on GCN2 and PKR across a dose range would shed light on whether this cross-reactivity is likely to be of concern.

We thank the reviewer for highlighting compound 20 as the most interesting lead compound and for recognising its apparent ability to activate GCN2 independently of GCN1. The reviewer identified several limitations relating to cell-type specificity, the temporal behaviour of ISR activation, and possible PKR cross-reactivity.

In response to the concern about cell-type specificity, we tested compound 20 in additional cellular models. Compound 20 did not induce detectable ISR signalling in HCT116 or COS-7 cells under the conditions tested, consistent with the reviewer’s observation that activity is not universal across cell types. However, we observed induction of PPP1R15A in primary Mesobank T12 mesothelioma cells. We have therefore revised the manuscript to present compound 20 activity as cell-context dependent rather than broadly generalisable across all cell types.

To address the reviewer’s concern that the ISR output was examined only at limited time points, we extended the time-course analysis for compounds 20 and 21. Using live-cell ATF4::NanoLuc reporter measurements, both compounds showed maximal reporter activation at approximately 6-8 hours. We also measured translational output by puromycin incorporation and found that compound 20, but not compound 21, caused a progressive GCN2-dependent reduction in translation. These data provide a clearer view of the kinetics and functional consequences of compound 20-mediated ISR activation.

The reviewer also noted that compound 20 inhibits PKR in vitro at concentrations close to those required for GCN2 activation in cells. We agree that this is an important potential liability. Because PKR signalling was not robustly or reproducibly inducible in our CHO-based reporter system, we were unable to perform a reliable cellular dose-response analysis of PKR engagement in the present study. We have therefore revised the Discussion to acknowledge kinase cross-reactivity, including possible PKR inhibition, as an important limitation and an issue for future development of this chemical series.

Finally, we have moderated our mechanistic interpretation of compound 20. Although the data support direct engagement of GCN2 and suggest a mechanism distinct from canonical GCN1-dependent activation, we now discuss GCN1-independent activation more cautiously and in the context of prior reports that GCN2iB can display GCN1-independent activity under specific experimental conditions.

Discussion

“While the functional relevance of PKR inhibition in our cellular systems is uncertain, these observations highlight the potential for kinase cross-reactivity, which will be important to address in future studies.”

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):<br /> (1) The description of the chemical screen for Gcn2 activators is not sufficiently clear and detailed. a) Briefly and early on provide the rationales (modes of action) for using histindinol and latruculin A. Explain further the rationale in Figure 2, outlining the purpose for the combined compound + submaximal dose of histindinol.

The text has been amended.

Results

“Histidinol activates GCN2 by inhibiting histidyl‑tRNA synthetase, leading to the accumulation of uncharged tRNAHis. This uncharged tRNA binds to GCN2, relieving its autoinhibition and activating the kinase (35). Latrunculin A sequesters G‑actin, thereby inhibiting PPP1R15A activity (36, 37). Tunicamycin inhibits protein glycosylation in the endoplasmic reticulum (ER), resulting in activation of PERK (38).”

“This submaximal concentration was used to allow detection of compounds that enhance ISR signalling when GCN2 is partially activated.”

b) What is unique about the second CHO: ATF4-luc2 reporter line? Why do only 89 out of the original 130 compounds induce the ISR in this line versus the original CHO: ATF4-Nanoluc cell line? This is confusing for the reader about how compounds were triaged for characterization.

The ATF4‑Nanoluc and ATF4‑luc2 reporter lines differ only in the luciferase used, but this has important practical consequences. The Nanoluc reporter is substantially more sensitive, so it was used for the primary screen to detect even weak ISR activation. The luc2 reporter has lower sensitivity and a narrower dynamic range, making it a more stringent orthogonal assay. As a result, not all hits from the Nanoluc screen (130 compounds) reproduced in the luc2 line; the 89 compounds retained are those that robustly activate the ISR under these more stringent conditions. This step was therefore used to prioritise stronger, more reproducible activators for downstream characterisation.

Results

“While primary screening was performed in ATF4‑Nanoluc lines for maximal sensitivity, hits were subsequently re-tested in a second CHO ATF4::luc2 reporter line as a more stringent orthogonal assay to prioritise robust ISR activators. Of the 130 hits identified in the sensitive Nanoluc screen and passing early toxicity assessment, 89 were confirmed in the luc2 assay, consistent with enrichment for higher-amplitude ISR activators under more stringent detection conditions.”

c) The study uses a second CHO reporter line in the flow scheme (CHO:ATF4-luc2) and then switches back to an ATF4-Nanoluc line to establish GCN2 dependence. What is the rationale for switching back to the original reporter line?

The luc2 reporter line was used as a more stringent, orthogonal validation step to prioritise robust ISR activators. For subsequent mechanistic studies, including assessment of GCN2 dependence, we returned to the ATF4‑Nanoluc line because its higher sensitivity and simpler single‑reagent assay format are better suited to multi‑point measurements and comparative analyses. In effect, the luc2 reporter was used for triage, whereas the Nanoluc system was retained for mechanistic characterisation and downstream screening.

Results

“In subsequent mechanistic studies, the ATF4‑Nanoluc reporter was again used to take advantage of its higher sensitivity and simpler assay format for multi‑condition comparisons.”

d) The rationale for the first orthogonal screen described in the results section to identify inducers of ER stress is not clearly explained. The compounds were already determined to be dependent on GCN2 prior to this test, and one would have thought that this criterion would have covered ER stress and alternative eIF2 kinase activators.

We agree with the reviewer that, in principle, establishing GCN2 dependence should reduce the likelihood of capturing compounds acting through alternative eIF2α kinases. However, we performed this orthogonal ER stress screen to address two practical considerations. First, high‑throughput screening is inherently prone to false positives, as it is typically conducted at a single concentration and time point, and compound libraries may contain degraded or chemically inconsistent material. We therefore used a lower‑throughput, more controlled ER stress assay with freshly sourced compounds and additional readouts (e.g. CHOP and XBP1) to improve confidence in the hits. Second, despite prior evidence of GCN2 dependence, ER stress signalling via PERK converges on the same downstream endpoints: eIF2α phosphorylation and ATF4 induction. We therefore wished to explicitly exclude compounds that activate the ISR indirectly via ER stress. In practice, this proved important, as the orthogonal assay did identify compounds that induced ER stress, which we subsequently excluded from the lead set.

Results

“Although hits were prioritised for GCN2 dependence, we performed an additional orthogonal screen to exclude compounds that activate the ISR indirectly via ER stress, which converges on the same downstream outputs.”

(2) A major point of the manuscript is that there is GCN1 independence for the small molecule activation of GCN2, and this has not yet been reported. One report for this GCN1 independence is reference 9 [Carlson … Wek 2023] (Figure 5). In this report, low doses of GCN2iB that can activate GCN2 (although by the present manuscript at much lower levels than the identified new compounds) induce ATF4 expression in cells expressing an E26A mutant of GCN2 that is suggested to negate GCN1 binding and enhancement of GCN2 activity. Halofuginone induction of ATF4 expression was thwarted by the GCN2 E26A mutant.

We thank the reviewer for highlighting this important point. We agree that Carlson et al. (2023) suggest that, under certain conditions, GCN2iB can activate GCN2 independently of GCN1 using the E26A mutant. In our experiments, however, we did not observe GCN1‑independent activation with GCN2iB under the conditions tested, i.e. similar low doses. One possible explanation is that the GCN1‑independent activity reported by Carlson et al. occurs only within a narrow concentration range; their observations were made at very low compound concentrations, whereas higher concentrations may engage additional regulatory mechanisms. In our study, we used concentrations optimised for robust ISR activation, which may mask such effects. We also note that the E26A mutation (E18A in yeast), originally identified by two‑hybrid analysis, disrupts the GCN2-GCN1 interaction but may not completely eliminate all modes of functional coupling under all conditions. Taken together, these observations raise the possibility that GCN1‑independent activation represents a context‑dependent mechanism that may be unmasked only under specific experimental conditions or by particular classes of compounds.

We have revised the Discussion to acknowledge this prior report explicitly and to clarify how our findings relate to it.

Discussion

“Recent work suggests that the ATP-competitive modulator GCN2iB can activate GCN2 independently of GCN1 under specific conditions using a GCN2 E26A mutant (9). In our hands, we did not observe GCN1‑independent activation with GCN2iB at the concentrations tested. This discrepancy may reflect a narrow concentration window for GCN1‑independent activation or context‑dependent effects of the E26A mutation. These findings raise the possibility that GCN1‑independent activation of GCN2 may occur under specific conditions or with distinct classes of compounds.”

(3) The authors state that the ISR was exaggerated in Ppp1r15a KO cells. It would be helpful to include statistical analyses to support this statement.

Thank you for enabling us to be more precise. New text added:

Results

“Activation of the ISR by tunicamycin was exaggerated in the Ppp1r15a^-/- cells owing to their defective dephosphorylation of eIF2a (wild type vs Ppp1r15a^-/-, p<0.05).”

(4) The results state that Chop and Ppp1r15a mRNAs were measured following 4 hours of treatment with compound 18, 20, or 21, but Figure 2D shows treatment from 0 to 8 hours? It appears that the compound still induces these mRNAs in GCN2 KO cells, possibly with delayed kinetics. A lengthened time course study would help determine if this is indeed the case.

We thank the reviewer for this careful observation. To address the reviewer’s point regarding delayed or GCN2‑independent signalling, we have extended our analysis using compounds 20 and 21, which are more tractable experimentally (solubility). Using the ATF4‑Nanoluc reporter, both compounds show peak ISR activation at ~6–8 h in wild-type cells over an extended time course. In parallel, functional readouts of mRNA translation (puromycin incorporation) demonstrate that compound 20, but not 21, induces a progressive, GCN2‑dependent reduction in translation over this period. These clarify the temporal aspects of signalling by these two compounds.

Results

“Studies with compound 18 were limited by poor aqueous solubility; therefore, time‑course analyses focused on compounds 20 and 21. To assess ISR activation over an extended period, live‑cell luciferase measurements were performed using CHO cells stably expressing an ATF4::Nanoluc-PEST reporter. Both compounds elicited maximal reporter activation between 6 and 8 h (Figure S1A&B). Compound 20, but not 21, induced a significant GCN2‑dependent reduction in mRNA translation, as measured by puromycin incorporation, with a progressive effect observed up to 7 h (Figure S1C-F).”

Legend

“Supplementary Figure S1. Kinetics of responses to compounds 20 and 21

(A-B) Wild-type CHO cells stably expressing the ATF4::nanoLuc-PEST reporter were treated with Nano-Glo and either (A) 13mM compound 20 or (B) 13mM compound 21. Median bioluminescence (fold change normalised to DMSO control) ± 95% confidence. Representative experiment (n=4 technical repeats). (C-F) Representative immunoblot of lysates from wild-type or Eif2ak4^-/- CHO cells treated with 10μM compound 20 or 7.5μM 21 for the indicated times. Immediately before harvesting, cells were treated with 10μg/mL puromycin to label newly synthesised polypeptides. “-“ indicates cells not incubated with puromycin. “U” cells were treated with puromycin but without test compound. “CHX” represents the cycloheximide control (100μg/mL). Molecular size in kDa. (E-F) Quantification of puromycinylated proteins normalised to GAPDH. Mean ± SEM. CHO WT (black) and Eif2ak4^-/- cells (turquoise. N = 4 independent experiments. Two-way ANOVA with Šídák's multiple comparisons test; ***: p ≤ 0.001.”

(5) In the section describing the differences between cell lines in the ability of compounds to induce the ISR, this is difficult for the reader to interpret, as no controls are included. How does histidinol (or other canonical inducers of the ISR) behave in the three reporter assays (CHO, 293T, and HeLa)?

As requested, we now provide ATF4::Nanoluc reporter activation (3mM, 20 hours because of this drug’s slow kinetics)

Results

“To benchmark ISR activation in these models, each cell type was treated with 3mM histidinol (Figure S2). Reporter activation was most robust in CHO cells, followed by 293T cells, then HeLa cells.”

Discussion

“Moreover, histidinol-induced ISR activation showed a clear hierarchy across cell lines, with CHO cells being the most responsive and HeLa cells the least.”

Legend

“Supplementary Figure S2. Cell-type differences in response to histidinol

Fold-change of ATF4::NanoLuc reporter signal in HEK293T, HeLa and CHO cells transiently transfected with reporter and treated for 20 hours with 3mM histidinol. Fold-change calculated relative to vehicle control. Mean ± SEM).”

We thank the reviewer for this important question. However, we respectfully disagree that a direct correspondence between the concentrations required for target engagement in the BRET assay and for ISR activation in functional assays should necessarily be expected. BRET (including NanoBRET) is a target engagement assay that measures compound binding to the protein in intact cells, typically by competition with a labelled tracer, and thus reports on apparent intracellular affinity and occupancy rather than downstream biological effect (Robers 2019, PMID 30519940). By contrast, ISR activation is a functional readout that reflects amplification through signalling networks, and can be influenced by multiple additional variables including pathway non-linearity, feedback, and kinase regulation. Consequently, it is well established that potencies derived from target engagement assays do not always align with those measured in functional assays. For example, intracellular kinase profiling studies using NanoBRET have demonstrated systematic potency offsets between binding/engagement measurements and downstream cellular activity, arising from factors such as intracellular ATP competition and pathway context (PMID Capener 2026, PMID 41495225). More generally, target engagement assays provide a quantitative measure of binding, whereas functional assays measure biological outcome, and these readouts need not coincide because they capture distinct aspects of a compound’s mechanism of action. Accordingly, we interpret our BRET data as evidence of direct interaction with GCN2 in cells, rather than as a predictor of the concentration required to activate the ISR. The observation that higher concentrations are required in the BRET assay is therefore not unexpected and does not argue against a requirement for kinase-domain engagement in ISR activation. Instead, it reflects the different mechanistic endpoints captured by the two assay formats.

We will clarify this point explicitly in the revised manuscript.

Results

“The concentrations required to detect target engagement in NanoBRET assays did not directly mirror those required for ISR activation, reflecting the distinction between ligand binding and downstream pathway output.”

(7) In Figure 5E, the authors suggest that compounds 18 and 20 are non-competitive inhibitors of GCN2 since the Vmax increases with increasing ATP concentration. What is the Km for ATP in the absence or presence of compound 18 or 20? It would be helpful to include progress curves as supplementary data to support the Vmax plots in Fig. 5E. Consider providing more specific units (currently arbitrary units) for the y-axis.

We thank the reviewer for this insightful comment and agree that our original wording overstated the mechanistic interpretation of these data. In particular, the use of the term “non‑competitive” is not well supported by the current analysis and may be misleading, especially given that our data are consistent with binding within or proximal to the ATP-binding pocket. We have therefore revised the text to remove this designation and instead describe the data more conservatively in terms of changes in apparent Vmax, without assigning a specific inhibition mechanism. With respect to kinetic analysis, we agree that full determination of Km values and inclusion of progress curves would provide a more rigorous mechanistic interpretation. However, given the primary focus of this manuscript on identifying and characterising small‑molecule activators of GCN2 in cells, we believe that a detailed steady‑state kinetic analysis would be beyond the scope of the current study. We have therefore moderated our conclusions accordingly and now present these data as preliminary kinetic observations rather than definitive evidence of inhibition modality.

Results

“Compounds 18 and 20 altered the apparent kinetic parameters of GCN2, including an increase in the observed V_max; however, these data do not allow assignment of a specific inhibition modality.”

(8) The full-length GCN2 assay presented in Figure 5G appears to be unresponsive to uncharged tRNA, a known regulator of GCN2. The statement that compound 20 induces eIF2 phosphorylation to a greater extent than tRNAs is true for the in vitro assay, but arguably is because the in vitro assays do not recapitulate the in vivo arrangement.

We thank the reviewer for this comment. We respectfully disagree that the assay is unresponsive to uncharged tRNA. In our hands, GCN2 does exhibit activation in response to tRNA; however, the magnitude of this effect is modest (~4‑fold) compared to the substantially stronger activation observed with compound 20 (~40‑fold). As a result, the tRNA response can appear compressed when both are plotted on the same scale. We agree with the reviewer that the in vitro assay does not fully recapitulate the in vivo regulatory environment, where factors such as GCN1 and ribosome association are known to potentiate GCN2 activation. This limitation likely explains the relatively weaker response to tRNA under our assay conditions and was a key motivation for incorporating cellular assays in our study. Interestingly, the marked difference in activation magnitude between tRNA and compound 20 in vitro raises the possibility that compound-mediated activation may, at least in part, bypass regulatory features that normally constrain GCN2 activity in a GCN1‑dependent manner. While we have not directly tested this hypothesis, we will temper the wording and include this as a speculative point in the Discussion.

Results

“Of note, uncharged tRNA produced a modest (~4‑fold) activation of GCN2 under these conditions, whereas compound 20 induced substantially greater (~40‑fold) activation.”

Discussion

“The markedly greater activation observed with compound 20 compared with uncharged tRNA in vitro raises the possibility that such compounds may partially bypass regulatory constraints on GCN2 activation, including those normally mediated by GCN1.”

(9) Using purified GCN2 kinase domain at low ATP concentrations (10 μM), compounds 18 and 20 were shown not to inhibit GCN2 up to concentrations of 3 μM. In previous assays, much higher concentrations of compounds 18 and 20 were used to inhibit GCN2. Why were different concentrations used? This makes this interpretation of this data difficult for the reader to draw conclusions.

We thank the reviewer for this comment and agree that the use of different concentration ranges across assays may not have been sufficiently clear. The kinase‑domain assay performed at low ATP (10 μM) was specifically designed to assess whether compounds 18 and 20 have a propensity to inhibit GCN2 under conditions that sensitise detection of ATP‑competitive effects and facilitate comparison with related eIF2α kinases. This assay was therefore optimised for detecting inhibition, rather than activation. In contrast, the higher concentrations used in other experiments were selected to robustly measure ISR activation in cellular or full‑length protein contexts, where higher compound exposure is required to observe downstream signalling outputs. These two assay systems therefore address distinct mechanistic questions—targeting inhibition under controlled biochemical conditions versus activation in more complex functional settings—and are not directly comparable in terms of concentration–response relationships. We will revise the manuscript to clarify this distinction and to emphasise that the kinase‑domain assay was not intended to define the activation potency of the compounds.

Results

“This assay was performed at low ATP concentrations to sensitise detection of ATP-competitive inhibition and was not optimised to detect compound-mediated activation of GCN2.”

(10) In silico docking studies support the binding of compound 20 in the ATP-binding pocket of the GCN2 kinase domain. How is this compatible with the stated ATP non-competitive mechanism?

We agree with the reviewer that our previous description was misleading. The designation of compounds 18 and 20 as “ATP non‑competitive” is not supported by the available data and is inconsistent with the docking results suggesting binding within the ATP‑binding pocket. We have therefore revised the manuscript to remove this terminology and to describe the kinetic behaviour more cautiously, without assigning a specific mode of inhibition.

(11) The study does not appear to feature biological assays demonstrating the effects of GCN2 activation. For example, does compound 20 reduce translation or growth of cells in a GCN1/GCN2-dependent manner, or do the compounds overcome PVOD mutations akin to the authors' Hum Mol Genet 2024 Aug 18;33(17):1495-1505 article?

We thank the reviewer for this suggestion. We agree that defining downstream biological consequences of GCN2 activation is an important goal. We did explore this using several cellular systems; however, these effects were context-dependent and not consistently observed across models. Specifically, compounds did not induce a detectable ISR in HCT116 or COS‑7 cells under the conditions tested, despite responsiveness of these systems to canonical activators such as histidinol. In contrast, we did observe induction of PPP1R15A in primary mesothelioma cells, indicating that biological responses can be elicited in certain cellular contexts. We also tested whether these compounds could rescue disease-associated GCN2 variants linked to PVOD, as previously reported for GCN2iB, but did not observe activation of these mutants. These findings suggest that while the compounds robustly activate GCN2 signalling in reporter assays, downstream biological outputs are context-dependent and may require specific cellular conditions or co-factors. Given the variability across systems, we have limited our conclusions to ISR activation and have not generalised broader biological effects. We will clarify this point in the revised manuscript.

Results

“We went on to examine downstream cellular consequences of GCN2 activation in multiple models. While compounds did not induce detectable ISR signalling in HCT116 or COS‑7 cells under the conditions tested, induction of PPP1R15A was observed in Mesobank T12 primary mesothelioma cells, indicating context-dependent biological responses. Moreover, in contrast to GCN2iB (17), the current compounds did not activate disease-associated GCN2 variants linked to PVOD [data not shown].”

(12) For the control of neratinib and other activators linked with ATP binding that are suggested to be dependent on GCN1 in Fig. 4, include reporter induction by drug treatment in Gcn2-/- cells. It would be helpful to be clear in the list of compounds between those suggested to be direct activators versus those that may create stress that leads to GCN2 activation.

We thank the reviewer for this suggestion. In response, we have performed additional reporter assays in both WT and GCN2^-/- cells to assess the dependence of drug-induced ISR activation on GCN2. These experiments reveal clear GCN2 dependence for reporter induction upon treatment with sunitinib, NXP800, WEE1-in-4, Debio0123, gefitinib, and erlotinib, as signal is markedly reduced in GCN2⁻/⁻ cells compared to WT.

In contrast, dovitinib and AZD1775 show less clear dependence, with relatively low reporter signal even in WT cells (notably lower than observed in Fig. 4C), limiting interpretation. Interestingly, dabrafenib induces stronger reporter activity in GCN2^-/- cells than in WT, indicating that its effects are independent of GCN2 and may reflect activation of alternative stress or signalling pathways.

Results

“To further assess the mechanism of compound-induced ISR activation, we evaluated reporter responses in GCN2-deleted cells (Supplementary Figure S3). Several compounds, including sunitinib, NXP800, WEE1-in-4, Debio0123, gefitinib, and erlotinib, showed reduced reporter activity in GCN2-deficient cells, consistent with GCN2-dependent activation. In contrast, dovitinib, AZD1775 and dabrafenib produced weaker or inconclusive responses even in the paired wild-type lines, limiting analysis. These data allow us to distinguish compounds consistent with direct or GCN2-dependent activation from those more likely to induce ISR indirectly through cellular stress upstream of GCN2. These findings support a distinction between compounds that activate the ISR through GCN2-dependent mechanisms and those that likely act indirectly via alternative stress pathways.”

Legend

“Supplementary Figure S3. GCN2-dependence of ISR activation by putative GCN2 agonists

Normalised fold-change in ATF4 signal in CHO WT (purple) and Gcn2-/- (blue) ATF4::NanoLuc reporter cells treated for 19 hours with a panel of ATP-competitive kinase inhibitors reported to activate GCN2 (at 1 and 3µM, sunitinib used at 3 and 10µM, AZD1175 used at 0.3 and 1µM, gefitinib and erlotinib used at 3 and 10µM). DMSO was used as vehicle control. (n=3; mean ± SEM).”

Minor Comments:

(1) In the introduction, the authors state that "Several type 1 and 1.5 kinase inhibitors can activate GCN2 at low concentrations while inhibiting at higher concentrations". What is the evidence that type I inhibitors can activate GCN2?

Thank you for this query. We believe our initial phrasing was open to misinterpretation. The new wording is

Introduction

“Several kinase inhibitors classified as type 1 or type 1.5 with respect to their canonical targets can activate GCN2 at low concentrations while inhibiting it at higher concentrations; however, their binding mode to GCN2 remains undefined.”

(2) A CHO:ATF4-Nanoluc translation reporter screen was used to screen 123K compounds and divide them into a pool of inhibitors and a pool of activators. The criteria used to make this distinction are not sufficiently described in the manuscript, and screen results are not supplied as supplementary data.

We thank the reviewer for highlighting the need for greater clarity regarding the screening criteria and reproducibility. Compounds from the primary screen (BioAscent library of 123,222 drug-like compounds performed by the ALBORADA Drug Discovery Institute) were classified based on Z-score thresholds, with activators defined as those with Z-score > 3. To assess robustness, the primary screen data were re-analysed independently. In an initial analysis of 121,000 compounds (excluding plates failing quality control), 6,521 compounds met the activator threshold. Of these, 6,461 overlapped with the original hit list. The small number of discrepancies included: (i) compounds absent from the analysed dataset (e.g. originating from failed plates), and (ii) compounds with Z-scores close to the threshold (typically between −3 and −3.02), consistent with minor analytical variation (e.g. rounding). Overall, these analyses show a high degree of concordance in hit identification, with differences restricted to borderline cases near the selection threshold. We have clarified these criteria below.

Results

“Compounds were classified based on Z-score thresholds derived from the primary screen, with activators defined as those with Z-score > 3. An initial analysis of 121,000 compounds, excluding plates failing quality control, identified 6,521 activators, of which 6,461 overlapped with the subset selected for follow-up screening. Minor discrepancies were restricted to compounds absent from the analysed dataset (e.g. originating from failed plates) or those with Z-scores close to the threshold (3 to 3.02), consistent with limited analytical variation. Using this approach, we assembled a subset of 6,461 compounds enriched for potential ISR activators and screened these in 384-well format at 10 µM for 16 hours.”

Reviewer #3 (Recommendations for the authors):

(1) The legend to Figure 4 should read "Compound 20 displays GCN1 independence", not "dependence".

Thank you for spotting this error. We have made the correction.

(2) The text describes Figure 2D as examining 4h, but it also examines 8h.

We have amended the text to:

“After validating their effects using the assays described above, we next confirmed activation of the ISR by these compounds at the transcriptional level at 4 and 8 hours.”

(3) Unless I'm misreading Table S1, compound Z134826202 is listed as activating the ISR, but the authors describe it in the text as inactive.

Thank you. We have corrected this error.

Reviewer #2 (Public review):

Summary:

This manuscript, by Kang and colleagues, examines the activity of the Gasdermin (GSDM) family of proteins in epidermal skin cells called keratinocytes. Gasdermins are a relatively recently discovered family of pore-forming proteins that are generally studied for their role in inflammasome activation and pyroptosis. The role and functions of GSDMs in keratinocytes are not fully understood, but programmed cell death is important to keratinocytes as part of the cornification process. Keratinocyte pyroptosis associated with GSDMD and GSDME has been linked to the pathogenesis of psoriasis, and GSDMA-mediated pyroptosis in keratinocytes is responsive to Staphylococcus aureus infection. These data suggest that defining the role of GSDM in keratinocytes could provide valuable insight into the treatment of skin pathologies, particularly those involving inflammation.

Strengths:

In this study, the authors use two human keratinocyte cell lines to focus primarily on the biology of GSDM in keratinocytes, particularly A, B and C, both during homeostatic conditions and in response to some pathogenic stimuli. The data indicate that the full-length GSDMA and B proteins are colocalizing not to the plasma membrane, but to the cytoskeletal, cytoplasmic or membrane fractions, in direct contrast to a number of other studies. Further, there are data showing the GSDM are not cleaved by the same proteases as in other studies, and that GSDMA is not involved control of Vaccinia virus skin infection. The controls for cleavage done in HeLa cells with a number of inhibitors were particularly strong, and the overall finding is interesting and potentially important, providing novel insight in the biology of GSDM proteins in keratinocytes. The second half of the paper focuses on treatment with calyculin A, a protein phosphatase inhibitor that the authors report recapitulates acantholysis, i.e. the loss of cell-cell junctions between keratinocytes. Calyculin A altered the localization of GSDM but increased cell death and keratin release irrespective of GSDM activity, suggesting that acantholytic cell death does not involve GSDM. Overall, these data are an interesting study showing differences in GSDM biology in keratinocytes and indicating that GSDM are not involved in a number of pathogenic processes associated with skin diseases.

Weaknesses:

Unfortunately, the data supporting these findings are not particularly strong and are somewhat confusing, and the study is not well organized. One of the major weaknesses is that most work was done in two human keratinocyte cell lines, and subcellular localization subcellular localization results were not consistent between the two lines. This is not hugely surprising, as cell lines even from the same cell type are very different, and the authors note that there may be some variability associated with the model used. However, the inconsistency between the two lines, alongside the lack of comparison to primary cells or patient data, makes it very difficult to determine which, if any, result accurately recapitulates the biology in human cells. The authors do suggest that the N-TERT line is more similar to primary cells but rely on a single study and no other research or data to support this. Several other experiments compare the keratinocyte lines to a colorectal adenocarcinoma cell line, showing the keratinocytes are also different from these cells, but it is not clear that they support the information about GSDM in either keratinocyte line other than showing that keratinocytes are distinct from colorectal cells.

There are several specific concerns regarding the data, particularly the western blotting. In a number of the blots, there are bands which are not mentioned. For example, in Figure 3A, "GSDMA and GSDMB translocated from the cytoskeleton equally to both the cytoplasmic and membrane fractions upon calyculin A treatment". But in the western blot shown, GSDMA moves from being entirely in the cytoskeleton to having lighter bands in the cytoskeleton and chromatin-bound regions and heavier bands in the membrane and cytoplasm. GSDMB shows heavier bands in the membrane and cytoplasm after Calyculin A, but the heaviest band remains in the chromatin-bound region. In addition, the western blots have no associated N or pooled data graph with statistics, so it is not clear how many studies it represents or if there were significant changes. A lack of statistical analysis was also seen in the Yoyo-1, membrane ballooning and cell death analyses in 3A, 3C, 4B and 4C, with the membrane ballooning finding being entirely qualitative and not using any image analysis - comparison of P2A or cell area could resolve this. The methods do note that unpaired T-tests and 2-way ANOVAs were performed in some places, so the results of these should be clearly shown on the graphs. Beyond this, all of the localization studies rely on a single modality, western blotting. This finding should be supported by orthogonal approaches, ideally immunofluorescence. Finally, the presence of GAPDH in the supernatant (5A) raises an important issue regarding nonspecific intracellular release. If the GAPDH is present in the sup that is representative of cell damage and does not correlate with the absence of other intracellular markers - such as those found in lysates - in the supernatant.

The final weakness of the paper was that it is not really clear what the study is trying to say, and it was difficult to follow the logic and data, with several figures being out of order. There are a number of statements that feel overinterpreted, such as stating that clathrin-mediated endocytosis is required for calyculin A mediated acantholysis. While it is clear that Calyculin A recapitulates features similar to acantholysis, it is a broad perturbation of the cell, and it is not clear that it is the same process. Further, the pharmacological evidence is not strong enough to make such a clear claim about the role of clathrin-mediated endocytosis. Similarly, the statement that "GSDMs can induce pyroptosis not only in their cleaved forms but also as full-length proteins through specific post-translational modifications" is much too strong for these data. The terminology surrounding "pyroptosis" should be used more carefully, as yoyo-1 uptake and membrane ballooning demonstrate membrane permeability and morphological changes but are not sufficient for cell death.

Reviewer #1 (Public review):

The manuscript by Kang et al. examines Gasdermin family member localization in keratinocytes and further reports that the PP2A inhibitor caliculyn A induces a pyroptosis-like morphology, presumably due to cytoskeletal disruption. Gasdermin A and B then translocate from the cytoskeleton to the membrane, where they enable keratin secretion without cleavage. The manuscript is well-controlled, and the data are convincing. The findings could be triangulated more effectively with additional microscopic studies to complement the fractionation experiments.

My specific comments are as follows:

(1) Nitrogen cavitation experiments and fluorescence microscopy have shown that the N-terminus of Gasdermin A first localizes to the mitochondrial membrane and then to the plasma membrane (Kondolf et al., JBC, 2024). The same was shown for GSDMD by Judy Lieberman's group in a 2024 Immunity paper. It is surprising that the fractionation experiments didn't include mitochondrial localization.

(2) It appears that the blots in Figure 2 showing purity of fractionation (Lamin, LDHA, KRT19) are the same as those in Figure 1. I am assuming the experiment was performed at the same time. I think this should be made crystal clear, and it also raises the question of why not just include GSDMC in Figure 1, since they already have the other Gasdermins in that figure.

(3) There's negative data in this manuscript, but I consider it important. Gasdermin antibodies are problematic, and caspase cleavage of GSDMC has been inconsistent. This negative data is important to share.

(4) In Figure 1, why do GSDMD and GSDME show cleavage products without stimulation (minor point for curiosity)?

(5) The fractionation experiments would be more convincing if they were complemented by microscopic localization. Antibodies might be an issue, but others have used fluorescent tags successfully (need to tag the Gasdermins internally).

(6) PP2A is a broad serine-threonine phosphatase with hundreds of substrates. It's hard to rule in a cytoskeletal role to the exclusion of all the other changes caliculyn will cause - especially after 3 hours of exposure. A paragraph discussing this could be added to the Discussion. It's surprising that the authors don't disrupt the cytoskeleton in other ways.

(7) I'm confused by the data in Figure 5C. Knocking down GSDMA and GSDMB causes no change in keratin release (and may even increase it). If the point of the manuscript is that these cause protein secretion, shouldn't there be a change? Do they affect secretion?

(8) Along these lines, which GSDMB isoforms are present in the cell? It's well established that exon structure affects function.

(9) Minor point for curiosity: N/TERT cells can be differentiated and can form organoids. Have the authors attempted to determine if differentiation affects their findings? This would be especially interesting given the different localizations of GSDMA and GSDMB in the layers of the skin.

Overall, this manuscript provides valuable data for the field. Although it includes negative results, I believe sharing them is crucial. The experiments are well-executed and mostly convincing, though some aspects, particularly Figure 5C, appear contradictory. The manuscript would benefit from this reputable lab addressing the earlier points, especially the microscopic testing of the hypothesis.

eLife Assessment

This valuable study investigates the cleavage-independent functions of Gasdermins in keratinocytes. The data collected using advanced microscopy methods and Western blotting are solid, but the main conclusions from these experiments remain partially supported. This paper is relevant to those studying cell-type specific inflammatory responses and client/cargo specificities of Gasdermins.

eLife Assessment

This study provides valuable insights into the role of thalamic nuclei in associative threat and extinction learning, underpinned by a large dataset and rigorous, multipronged analyses. The evidence provided is solid, supporting the main conclusions. The manuscript will be of broad interest to researchers in learning and memory, fear, thalamic circuitry, and related mental health conditions.

Reviewer #1 (Public review):

Summary:

Badarnee and colleagues analyse fMRI data collected during an associative threat-learning task. They find evidence for parallel processes mediated by the mediodorsal, LGn and pulvinar nuclei of the thalamus. The evidence for these conclusions is promising, but limited by a lack of clarity regarding the preprocessing and statistical methods.

Strengths:

The approach is inventive and novel, providing information about thalamocortical interactions that are scant in the current literature.

Comments on revised version.

No further comments. The authors have adequately addressed my concerns.

Reviewer #3 (Public review):

Summary:

The present work was aimed at investigating the specific contributions of thalamic nuclei to associative threat learning and extinction. Using fMRI, it examined activation patterns across pulvinar divisions, the lateral geniculate nucleus (LGN), and the mediodorsal thalamus (MD) during threat acquisition, extinction, and recall. It goals was to uncover whether distinct thalamic systems support different modes of learning-automatic survival mechanisms versus more deliberate processes-and to propose a hierarchical pulvinar model of fear conditioning. The manuscript also tried to refine current neuroanatomical models of threat learning and memory, highlighting the role of thalamic nuclei in it.

Strengths:

(1) Valuable theoretical elaboration and modeling regarding the differential role of pulvinar subdivisions on feedforward (inferior, lateral) and higher-order integration (anterior), and their functional interplay with other relevant subcortical and cortical structures in associative threat and extinction learning.

(2) Large sample sizes and multipronged analytical approaches were used for hypothesis testing.

(3) Exhaustive literature review in the field of associative threat, as well as regarding the role of thalamic nuclei and other brain structures in it.

Comments on revised version:

My three main concerns remain insufficiently addressed or only partially addressed in the study limitations. First, using 3 mm isotropic voxels limits anatomical and functional specificity when investigating small thalamic nuclei. Higher-resolution fMRI protocols and probabilistic thalamic atlases that allow subject-specific segmentation have been available for several years and could have been used to improve anatomical precision. Second, although analyses in normalized space are widely accepted for group-level analyses, normalization can distort the size and boundaries of small thalamic nuclei. Analyses in native space, using subject-specific segmentation, would help address this limitation, particularly given the relatively low functional resolution. Finally, standard preprocessing procedures and the inclusion of motion parameters as covariates do not fully exclude residual motion-related effects or potential differences in motion across experimental conditions. The authors do not report additional motion quality-control procedures, such as excluding participants with excessive motion or substantial translational or rotational drift. Taken together, these concerns prevent me from upgrading my assessment.

Author response:

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

eLife Assessment:

This study provides valuable insights into the role of thalamic nuclei in associative threat and extinction learning, underpinned by a large dataset and rigorous, multipronged analyses. The evidence provided is solid, supporting the main conclusions. Minor analytical refinements notwithstanding, the manuscript will be of broad interest to researchers in learning and memory, fear, thalamic circuitry, and related mental health conditions.

We sincerely thank the editors and the reviewer for their careful evaluation of the revised manuscript and for their constructive comments. We greatly appreciate the time and effort invested in reviewing our work, and we have carefully considered each suggestion to further improve the clarity, presentation, and interpretation of the manuscript.

Author’s note:

We note that several comments refer to figure numbers and manuscript text from an earlier version of the manuscript (e.g., Figure 7 in the earlier version corresponds to Figure 8 in the current revised manuscript). For clarity, our responses below refer to the corresponding figures and sections in the current revised manuscript while addressing each comment individually.

Public Reviews:

Reviewer #1 (Public review):

Summary:

Badarnee and colleagues analyse fMRI data collected during an associative threat-learning task. They find evidence for parallel processes mediated by the mediodorsal, LGn and pulvinar nuclei of the thalamus. The evidence for these conclusions is promising, but limited by a lack of clarity regarding the preprocessing and statistical methods.

We thank the reviewer for these encouraging comments and for recognizing the novelty of our approach and its potential contribution to advancing our understanding of thalamocortical interactions during associative threat learning. We also appreciate the reviewer’s constructive suggestions regarding the methodological description, which have helped us improve the clarity of the manuscript.

Strengths:

The approach is inventive and novel, providing information about thalamocortical interactions that are scant in the current literature.

Weaknesses:

(1) There are not sufficient details present to allow for the direct interrogation of the methods used in the study.

Thank you for raising this point. As described in our previous revision, we substantially expanded the Experimental Procedure section by adding three dedicated subsections describing the fear learning procedure, shock calibration, and participant instructions. These additions included details regarding the stimuli, experimental context, stimulus and shock durations, average inter-trial interval, pseudorandomized stimulus presentation, and the procedure used to determine each participant’s shock intensity.

In the current revision, we have further clarified the Methods by specifying electrical stimulator that we used and specified the usage of the gel and placement of the electrodes. We believe these additions provide sufficient methodological detail to allow readers to understand and replicate the experimental procedure (page 40):

“Electrical stimulation was delivered using a BIOPAC stimulation system with BIOPAC electrodes and conductive gel. The electrodes were attached to the palm of the participant's right hand and positioned in close proximity without making contact with one another. The electric stimulation began…”

(2) The figures do not contain sufficiently granular details, making it challenging to determine whether the observed effects were robust to individual differences.

Thank you for this suggestion. To better illustrate the underlying data distribution and between-participant variability, we revised Fig. 1 and 4-6 by adding individual participant data points for the block-level analyses and kernel density estimate (KDE) distributions for the trial-level analyses. These additions allow readers to better assess the distribution of the data and the extent to which the reported effects are consistent across participants while maintaining figure readability. An example of the revised Fig. 1.

Comments on revisions:

I continue to recommend the plotting of individual data points. While there may be individual variance, it is important to quantify this in publication so that future studies can appreciate the uncertainty surrounding test statistics.

Thank you for this recommendation. As described in our response to Weakness 2 above, we revised the figures to better illustrate the underlying data distribution and between-participant variability. Specifically, we added individual participant data points for the block-level analyses and kernel density estimate (KDE) distributions for the trial-level analyses. We believe these revisions improve the transparency of the underlying data while maintaining figure readability. An example of a revised figure is included above.

Reviewer #2 (Public review):

Summary:

The authors quantify human fMRI BOLD responses in pulvinar and mediodorsal thalamic nuclei during a fear conditioning and extinction task across two days, in a large sample size (hundreds of participants). They show that the BOLD responses in these areas differentiate the conditioned (CS+) and safety (CS-) stimulus. Additionally this changes with repeated trials which could be a neural correlate of fear learning. They show that the anterior pulvinar is most correlated with the MD, and that this is not due to anatomical proximity. They perform graph analysis on the pulvinar sub nuclei which suggests that the medial pulvinar is a hub between the sensory (lateral/inferior) and associative (anterior) pulvinar. They show different patterns of thalamic activity across conditioning, extinction, recall, and renewal.

We thank the reviewer for this thoughtful summary of our work and for recognizing the scope of the study and the value of the findings. We appreciate the reviewer’s careful evaluation and constructive feedback, which helped improve the clarity and interpretation of the manuscript throughout the revision process. Below, we respond to each of the points raised under the Weaknesses and Comments on revisions sections and describe the corresponding revisions made to the manuscript.

Strengths:

The data has a large sample size (n=293 in some measures, n=412 in others). This is a validated human fear conditioning/extinction task that Dr Milad's group has been working with for several years. Few labs have investigated the thalamus activity during fear conditioning and extinction, particularly with a large sample size. There is an independent replication of the pulvinar network structure (Fig. 3), which suggests that the processing in the more sensory-related inferior and lateral pulvinar is relayed to the anterior pulvinar (and possibly thereby to more action-related prefrontal areas) via an intermediate step in the medial pulvinar - potentially a novel discovery but that needs more validation.

Weaknesses:

(1) The authors cannot make causal claims about their results based on correlational neuroimaging evidence. Causal claims should be pared back. E.g. Sentence 1 in results "The anterior pulvinar and MD contribute to early associative threat learning, as evidenced by increased functional activation in response to CS+ compared to CS- at the block level (Fig. 1b-c)." needs to be reworded to something like 'the anterior pulvinar and MD have increased functional activation... This suggests that these areas may contribute to early associate threat learning"

We agree with the reviewer that causal claims cannot be made based on correlational neuroimaging evidence. Accordingly, we carefully revised the manuscript throughout to avoid causal language and instead use wording that reflects statistical associations and interpretations. Representative examples of these revisions are provided below.

Representative examples of these revisions are shown below:

Page 7: “At the block level, both the anterior pulvinar and MD showed increased activation to CS+ vs. CS− (anterior pulvinar: t₍₂₉₂₎ = 4.41, p = 0.00001, d = 0.25; MD: t₍₂₉₂₎ = 6.41, p = 5.83x10^-10, d = 0.37; Fig. 1b–c), suggesting a possible involvement of these regions in early associative threat learning.”

Page 13: “A data-driven approach suggests a hierarchical computational organization of pulvinar divisions during associative threat learning.”

Page 13: “we integrated findings from complementary analyses to propose a computational framework describing the functional relationships among pulvinar divisions during associative threat learning.”

Page 13-14: “The network plot identified a statistical association between the lateral and anterior pulvinar, together with pathways linking the lateral and inferior pulvinar to the anterior pulvinar through the medial pulvinar (Fig. 3b).”

Page 14: “The medial pulvinar statistically mediates the functional relationships within the pulvinar network and is associated with elevated centrality values. Together, these findings are consistent with a possible hub-like functional role of the medial pulvinar, consistent with higher-order integration of CS+ information, relative to the more sensory-related functional profiles in the inferior and lateral pulvinar. This possible hierarchical organization is consistent with previous evidence showing that the inferior and lateral pulvinar are associated with processing basic sensory information (Berman and Wurtz, 2010, 2008; Cortes et al., 2024), while the medial pulvinar is implicated in higher-order processing, including attention and working memory (Homman-Ludiye and Bourne, 2019).”

Page 15: “This data-driven analyses provide a possible new perspective on the functional specialization of pulvinar divisions during associative threat learning. Based on these observations, we hypothesized a computational model in which activation in the medial pulvinar statistically mediates the functional relationships between the inferior and lateral pulvinar and the anterior pulvinar (Fig. 3e). To evaluate this hypothesis, we performed a mediation analysis and assessed the robustness of the model using k-fold cross-validation.”

Page 25: “The vmPFC connectivity may reflect involvement of prefrontal processes related to evaluation and regulation of threat-related information.”

Page 27: “The medial pulvinar may statistically mediate the functional relationships between the inferior and lateral divisions and the anterior pulvinar, consistent with a hierarchical computational model.”

Page 29: “Together, these findings are consistent with the possibility that the anterior pulvinar and MD contribute in parallel to associative threat learning, extending the two-system.”

Page 32: “Using statistical mediation analysis, we found that activation in the medial pulvinar statistically mediated the relationship between activation in the inferior/lateral pulvinar and activation in the anterior pulvinar. This statistical mediation is consistent with, but does not establish, a hierarchical computational organization.”

(2) Fig .1 The fact that the difference in BOLD activity between CS+ and CS- goes away on the third trial is not addressed. This is a very large effect in the data.

Thank you for raising this important point. We have addressed this issue in both the Discussion and the Limitations sections, as follows:

Discussion (page 31):

“The attenuated CS+ vs. CS− differentiation on the third trial specifically in the pulvinar may reflect a decreased requirement for differential thalamic engagement once the initial association has been acquired. Notably, because the MD sustained the BOLD response to the CS+ in the third trial which may indicate involvement of this nucleus in continued processing or maintenance of the learned association. This aligns with the well-established MD-PFC circuit involved in cognitive processes (Wolff and Halassa, 2024). Additionally, in a previous study using a similar paradigm, we observed sustained CS+ vs. CS− differentiation on the third trial in the nucleus reuniens, as well (Tuna et al., 2025). These findings suggest that trial-dependent learning dynamics may vary across thalamic nuclei rather than reflecting a uniform thalamic learning signal. Together, while our paradigm does not inherently distinguish between different stages of learning, such as early acquisition and stabilization, our findings are consistent with stronger associative learning-related engagement during the first two trials, with a reduced differential response by the third trial that may reflect the involvement of different neural processes.”

Limitations (page 36):

“The reduced CS+ vs. CS− differentiation observed in the pulvinar subdivisions during the third trial may reflect a transition from the rapid associative learning processes engaged during the first two trials to a different stage of threat processing. However, the present experimental design does not allow us to determine the specific psychological or neural mechanisms underlying this change. Importantly, this attenuation was not observed uniformly across thalamic nuclei, suggesting that it does not reflect a general thalamic response but may instead indicate nucleus-specific functional dynamics. Future studies specifically designed to dissociate different stages of threat learning will be required to determine the functional significance of this trial-dependent effect.”

(3) Fig. 3 Could the observed network structure be due to anatomical proximity? Perhaps the authors should do an analogous analysis to what they did in Fig. 2 for this intra-pulvinar analysis. This analysis doesn't take into account the indirect connections through corticothalamic and thalamocortical connections with visual cortex and the pulvinar. There is an implicit assumption that there are interconnections between the pulvinar sub nuclei, but there are few strong excitatory projections between these sub nuclei to my knowledge. If visual areas are included in the graph, it would make things more complex, but would probably dramatically change the story. In this way, the message is somewhat constructed or arbitrary.

We thank the reviewer for this thoughtful suggestion. We performed an analysis analogous to that presented in Figure 2. Specifically, we computed partial correlations between pulvinar divisions while controlling for activation in the other thalamic nuclei included in the study (LGN and MD). The relationships between pulvinar divisions remained significant after accounting for shared variance with these nuclei (Supplementary Figure 4), suggesting that the observed associations are not solely explained by activation in neighboring thalamic regions. This analysis is particularly relevant for the LGN because of its close anatomical proximity to the pulvinar and its central role within the visual thalamic system.

Additionally, we statistically compared the strength of the partial correlations among all pairs of pulvinar divisions (Author response table 1). These analyses demonstrated substantial heterogeneity in the correlation strengths, with several pairwise differences remaining significant after FDR correction, indicating that the functional relationships among pulvinar divisions are heterogeneous rather than uniformly correlated, making a simple anatomical-proximity explanation less likely.

Author response table 1.

Pairwise comparisons demonstrate heterogeneous functional relationships among pulvinar subdivisions after controlling for MD and LGN activation.

Note. All values are partial correlations controlling for MD and LGN activation (N = 293). Differences between correlations were tested using Steiger’s test for dependent correlations measured in the same participants. The sample size entered into the test was 293 - 2 = 291 as two variables (MD and LGN) were partialled out. The test itself then uses 291 - 3 = 288, the standard adjustment applied whenever correlations are converted to Fisher z scores. Steiger provides two versions of the same equation: one for comparisons in which the two correlations share a subdivision, and one for comparisons in which they share none (marked †). Analyses were conducted using the cocor package in R.

The p-values for the 15 comparisons were corrected using the Benjamini-Hochberg false discovery rate (FDR) procedure; 9 of 15 comparisons remained significant after correction. Correlations ranged from 0.182 to 0.718. PuA = anterior pulvinar; PuI = inferior pulvinar; PuM = medial pulvinar; PuL = lateral pulvinar.

The two correlations being compared share no subdivision.

Finally, we agree that pulvinar function is embedded within broader corticothalamic and thalamocortical circuits involving visual cortical regions. However, a comprehensive characterization of pulvinar-visual cortical interactions at the level of specific pulvinar divisions was beyond the scope of the current study. To acknowledge this limitation and avoid implying that the observed relationships represent isolated direct interactions among pulvinar divisions, we revised the conceptual model in Fig. 8 by adding bidirectional dashed gray arrows between the pulvinar divisions and visual cortex (see Fig 8). Additionally, we clarified this point in the Figure notes and the study limitations:

Limitations (page 36):

“The observed relationships among pulvinar divisions during conditioning are purely functional and do not distinguish direct inter-nuclear interactions from indirect coupling mediated by corticothalamic and thalamocortical pathways, including visual cortical regions. Thus, the proposed functional pulvinar model may reflect indirect corticothalamic and thalamocortical loops, currently undocumented internuclear interactions, or a combination of these mechanisms.”

(4) In the results section describing Fig. 4-7, there are no statistics supporting the claims made. There needs to be a set of graphs comparing the results across the study sessions and days, with statistical comparisons between the different experiments to confirm differences.