generics is a language feature that is available in many programming languages. It allows to write code that can be used with different types. The type is not specified in the code but is passed as an argument when the code is used. YouTube: How to use generics in Typescript typescriptlang: generics

Examples of generics usage in move: sui move intro course

• three boundaries that industry should have abided by but have been violated:
  • don't put them on the open internet until you solve the control problem
  • don't teach them to code because that enables them to learn and develop on their own
  • Don't allow other AI's prompting them, other AI agents working with them

LinkedList = singlyLinkedList (onlyForwards) DoubleLinkedLIsts = LinkedDeques

When resizing an array, should we always resize it to be double the current capacity?

In what occasion, asymptotic analysis is useful. Why do we need to evaluate the efficiency of the code?

Whenever a developer implements code, will they always have to manually create tests themselves to test implementation? Is there a more time efficient way to handle testing or is there no other way around it?

At what point (or number of element) does the code shown above start struggling and becoming inefficient?

This is Salesforce's second iteration of the CodeT5+ model. The first one already obtained SOA results, but apparently the second one has been trained with an instruct-based dataset, which allows it to be conversational on top of its adeptness for the purposes of code generation.

opposite of "code is law"

This is an error/defect on the part of the author.

This misses the point of LP—the true "fundamental theorem of LP" is basically that the compiler should be made to accept the preferred form.

A generator is a function-like iterator object.

The key concept for whole unit 1

it might be worth comparing this with the ordinary versions

Reviewer #2 (Public Review):

In this work, Hänisch and colleagues investigate the relationship between neurotransmitter transporter and receptor's spatial heterogeneity and well-studied functional and structural brain gradients in the human brain. They calculate the spatial similarity between the distribution of the neurotransmitter transporters and receptors for each parcel, thus obtaining a new brain distribution comprising a similarity index of all neurotransmitters mapped to each brain area. They employ a nonlinear dimensionality reduction on this neurotransmitter similarity map to reveal three spatial gradients for cortical and subcortical levels, respectively. Based on this, they characterize their significance by comparing them with functional fMRI meta-analytic activations, MRI microstructure, architectural contextualization, MRI-based structural and functional connectivity, and gray matter atrophy-derived disease maps.

The claim of the work is broad, and the motivation is general, but the data presented is specific and biologically diverse. The neurotransmitter system operates at different pre- and post-synaptic synaptic levels, and the general assumption that transporters are equivalent to receptors lacks appropriate discussion for supporting this claim. The motivations of the work are very broad, and the analysis used is sufficient for the general claims, but the data presented is specific and biologically diverse.

Besides these conceptual issues, I find this work interesting as it jointly characterizes the cortical and subcortical PET neurotransmitter's distribution maps and their structural and functional meaning for the first time. In essence, the study presents several arguments to consider the organization of the characterized maps as an additional layer of brain organization. The results are convincing and clearly presented. Although this is a correlative study using unconnected datasets, I appreciate the use of multiple brain maps. I also appreciate that the authors made the data and code available for reproducibility. The data and analysis used in the current draft enable a powerful set of tools for hypothesis testing in the human brain's natural distribution of neurotransmitters beyond the usual pharmacological intervention strategy traditionally used in neurotransmitters' brain mapping area.

Code for data processing and model training should be separated as different modules.

Reviewer #3 (Public Review):

In this paper, the authors analyze a large previously published deep mutational scanning data set using a reference-free regression approach. They extract the contributions of single locus and epistatic effects to the functionality of the sequence (no, weak or strong transcription activation of two response elements). They find that pairwise epistasis plays a crucial and dominant role at creating functional sequences and at connecting the functional sequence space.

I enjoyed reading the paper and the topic (role of epistasis at creating and connecting functional sequences; development of measures of epistasis) is very exciting to me. However, I found it difficult to judge the strength of the paper both because it is written in a rather dense and yet potentially redundant fashion (see comment 1) and because I was left with a number of questions upon reading. I will focus on conceptual questions in the following comments, since I am not able to judge the statistical approach in detail.

1/ Regarding the biological result (importance of pairwise epistasis) I was wondering how potentially redundant the consecutive sections of the paper are. In which situation would the authors expect that pairwise epistasis does *not* play a crucial role for mutational steps, trajectories, or space connectedness, if it is dominant in the genotype-phenotype landscape? I would also appreciate an explanation of how much new biological results this paper delivers as compared with the paper in which the data were published (which I, unfortunately, cannot access at the moment of writing this report).

2a/ Regarding the regression approach: I very much appreciate a reference-free approach to the estimation of epistasis. However, I would enjoy an explanation of how the results would have been (potentially) different if a reference-based approach was used, and how it compares with other reference-free approaches to estimating epistasis (e.g., linear regression or the gamma statistics of Ferretti et al. 2015).

2b/ When comparing the outcomes with and without epistasis, I understood that the authors compare the estimated "full model" with the outcome if epistatic effects were ignored - but without a new estimation of main effects if epistasis is ignored. Wouldn't that be a more fair comparison?

2c/ Where do the authors see the applicability of their approach to data beyond those analyzed in the present study? What are the requirements to use it? Does it only work for combinatorially complete landscapes? I did not have a chance to look at the code - how easily could other researchers apply the approach to their data?

Reviewer #3 (Public Review):

In this manuscript, Rossato and colleagues present a method for real-time decoding of EMG into putative single motor units. Their manuscript details a variety of decision points in their code and data collection pipeline that lead to a final result of recording on the order of ~10 putative motor units per muscle in human males. Overall the manuscript is highly restricted in its potential utility but may be of interest to aficionados. For those outside the field of human or nonhuman primate EMG, these methods will be of limited interest.

Notes<br /> 1. Artificial data should be used with this method to provide ground truth performance evaluations. Without it, the study assumptions are unchallenged and could be seriously flawed.

2. From the point of view of a motor control neuroscientist studying movement in animals other than humans or non-human primates, the title was misleadingly hopeful. The use case presented in this study requires human participants to perform isometric contractions, facilitating spatially redundant recordings across the muscle for the algorithm to work. It is unclear whether these methods will be of utility to use cases under more physiological conditions (ie. dynamic movement).

3. The text states that "EMG signals recorded with an array of electrodes can be considered and instantaneous mixture of the original motor unit spike trains and their delayed versions." While this may be a true statement, it is not a complete statement, since motor units at distal sites may be shared, not shared, or novel. It was not clear to me whether the diversity of these scenarios would affect the performance of the software or introduce artifacts. In other words, if at site 1 you can pick up the bulk signal of units 1,2,3,4; at site two you pick up the signals of units 2,3,4,5 and site three you pick up the signal of units 3,4,5,6, what does the algorithm assume is happening and what does it report and why?

4. I could not fully appreciate the performance gap solved by the current methods. What was not achievable before that is now achievable? The 125 ms speed of deconvolution? What was achievable before? Intro text around ln 85 states that 'most of the current implementations of this approach rely on offline processing, which restricts its ability to be used..." but no reference is provided here about what the non 'most' of can achieve.

5. Relatedly, it would have been nice to see a proof of concept using real-time feedback for some kind of biofeedback signal. If that is the objective here, why not show us this? I found the actual readout metrics of performance rather esoteric. They may be of interest to very close experts so I will defer to them for input.

6. I was disappointed to see that only male participants are used because of some vague statement that 'it is widely known in the field' that more motor units can be resolved in males, without thorough referencing. It seems that the objective of the algorithm is the speed of analysis, not the number of units, which makes the elimination of female participants not justified.

7. Human curation is often used in spike sorting, but the description of criteria used in this step or how the human curation choices are documented is missing.

8. The authors might try to add text to be more circumspect about the contributions of this method. I would recommend emphasizing the conceptual advances over the specifics of the performance of the algorithm since processor speed and implementation of the ideas in a faster environment (Matlab can be slow) will change those outcomes in a trivial way. Yet, much of the results section is very focused on these metrics.<br /> Minor<br /> Ln 115, "inversing" is not a word. "inverse" is not a verb<br /> Ln 186, typo, bioadhesive<br /> MVC should be defined on first use. It is currently defined on 3rd use or so.<br /> The term rate is used in a variety of places without units. Eg line 465 but not limited to that

Reviewer #1 (Public Review):

The main objective of this paper is to report the development of a new intramuscular probe that the authors have named Myomatrix arrays. The goal of the Myomatrix probe is to significantly advance the current technological ability to record the motor output of the nervous system, namely fine-wire electromyography (EMG). Myomatrix arrays aim to provide large-scale recordings of multiple motor units in awake animals under dynamic conditions without undue movement artifacts and maintain long-term stability of chronically implanted probes. Animal motor behavior occurs through muscle contraction, and the ultimate neural output in vertebrates is at the scale of motor units, which are bundles of muscle fibers (muscle cells) that are innervated by a single motor neuron. The authors have combined multiple advanced manufacturing techniques, including lithography, to fabricate large and dense electrode arrays with mechanical features such as barbs and suture methods that would stabilize the probe's location within the muscle without creating undue wiring burden or tissue trauma. Importantly, the fabrication process they have developed allows for rapid iteration from design conception to a physical device, which allows for design optimization of the probes for specific muscle locations and organisms. The electrical output of these arrays are processed through a variety of means to try to identify single motor unit activity. At the simplest, the approach is to use thresholds to identify motor unit activity. Of intermediate data analysis complexity is the use of principal component analysis (PCA, a linear second-order regression technique) to disambiguate individual motor units from the wide field recordings of the arrays, which benefits from the density and numerous recording electrodes. At the highest complexity, they use spike sorting techniques that were developed for Neuropixels, a large-scale electrophysiology probe for cortical neural recordings. Specifically, they use an estimation code called kilosort, which ultimately relies on clustering techniques to separate the multi-electrode recordings into individual spike waveforms.

An account of the major strengths and weaknesses of the methods and results.<br /> The biggest strength of this work is the design and implementation of the hardware technology. It is undoubtedly a major leap forward in our ability to record the electrical activity of motor units. The myomatrix arrays trounce fine-wire EMGs when it comes to the quality of recordings, the number of simultaneous channels that can be recorded, their long-term stability, and resistance to movement artifacts.

The primary weakness of this work is its reliance on kilosort in circumstances where most of the channels end up picking up the signal from multiple motor units. As the authors quite convincingly show, this setting is a major weakness for fine-wire EMG. They argue that the myomatrix array succeeds in isolating individual motor unit waveforms even in that challenging setting through the application of kilosort.

Although the authors call the estimated signals as well-isolated waveforms, there is no independent evidence of the accuracy of the spike sorting algorithm. The additional step (spike sorting algorithms like kilosort) to estimate individual motor unit spikes is the part of the work in question. Although the estimation algorithms may be standard practice, the large number of heuristic parameters associated with the estimation procedure are currently tuned for cortical recordings to estimate neural spikes. Even within the limited context of Neuropixels, for which kilosort has been extensively tested, basic questions like issues of observability, linear or nonlinear, remain open. By observability, I mean in the mathematical sense of well-posedness or conditioning of the inverse problem of estimating single motor unit spikes given multi-channel recordings of the summation of multiple motor units. This disambiguation is not always possible. kilosort's validation relies on a forward simulation of the spike field generation, which is then truth-tested against the sorting algorithm. The empirical evidence is that kilosort does better than other algorithms for the test simulations that were performed in the context of cortical recordings using the Neuropixels probe. But this work has adopted kilosort without comparable truth-tests to build some confidence in the application of kilosort with myomatrix arrays? Furthermore, as the paper on the latest version of kilosort, namely v4, discusses, differences in the clustering algorithm is the likely reason for kilosort4 performing more robustly than kilosort2.5 (used in the myomatrix paper). Given such dependence on details of the implementation and the use of an older kilosort version in this paper, the evidence that the myomatrix arrays truly record individual motor units under all the types of data obtained is under question.

There is an older paper with a similar goal to use multi-channel recording to perform source-localization that the authors have failed to discuss. Given the striking similarity of goals and the divergence of approaches (the older paper uses a surface electrode array), it is important to know the relationship of the myomatrix array to the previous work. Like myomatrix arrays, the previous work also derives inspiration from cortical recordings, in that case it uses the approach of source localization in large-scale EEG recordings using skull caps, but applies it to surface EMG arrays. Ref: van den Doel, K., Ascher, U. M., & Pai, D. K. (2008). Computed myography: three-dimensional reconstruction of motor functions from surface EMG data. Inverse Problems, 24(6), 065010.

The incompleteness of the evidence that the myomatrix array truly measures individual motor units is limited to the setting where multiple motor units have similar magnitude of signal in most of the channels. In the simpler data setting where one motor dominates in some channel (this seems to occur with some regularity), the myomatrix array is a major advance in our ability to understand the motor output of the nervous system. The paper is a trove of innovations in manufacturing technique, array design, suture and other fixation devices for long-term signal stability, and customization for different muscle sizes, locations, and organisms. The technology presented here is likely to achieve rapid adoption in multiple groups that study motor behavior, and would probably lead to new insights into the spatiotemporal distribution of the motor output under more naturally behaving animals than is the current state of the field.

Reviewer #2 (Public Review):

The authors provide a comprehensive investigation of self-citation rates in the field of Neuroscience, filling a significant gap in existing research. They analyze a large dataset of over 150,000 articles and eight million citations from 63 journals published between 2000 and 2020. The study reveals several findings. First, they state that there is an increasing trend of self-citation rates among first authors compared to last authors, indicating potential strategic manipulation of citation metrics. Second, they find that the Americas show higher odds of self-citation rates compared to other continents, suggesting regional variations in citation practices. Third, they show that there are gender differences in early-career self-citation rates, with men exhibiting higher rates than women. Lastly, they find that self-citation rates vary across different subfields of Neuroscience, highlighting the influence of research specialization. They believe that these findings have implications for the perception of author influence, research focus, and career trajectories in Neuroscience.

Overall, this paper is well written, and the breadth of analysis conducted by authors, with various interactions between variables (eg. gender vs. seniority), shows that the authors have spent a lot of time thinking about different angles. The discussion section is also quite thorough. The authors should also be commended for their efforts in the provision of code for the public to evaluate their own self-citations. That said, here are some concerns and comments that, if addressed, could potentially enhance the paper:

1. There are concerns regarding the data used in this study, specifically its bias towards top journals in Neuroscience, which limits the generalizability of the findings to the broader field. More specifically, the top 63 journals in neuroscience are based on impact factor (IF), which raises a potential issue of selection bias. While the paper acknowledges this as a limitation, it lacks a clear justification for why authors made this choice. It is also unclear how the "top" journals were identified as whether it was based on the top 5% in terms of impact factor? Or 10%? Or some other metric? The authors also do not provide the (computed) impact factors of the journals in the supplementary.

By exclusively focusing on high impact journals, your analysis may not be representative of the broader landscape of self-citation patterns across the neuroscience literature, which is what the title of the article claims to do.

2. One other concern pertains to the possibility that a significant number of authors involved in the paper may not be neuroscientists. It is plausible that the paper is a product of interdisciplinary collaboration involving scientists from diverse disciplines. Neuroscientists amongst the authors should be identified.

3. When calculating self-citation rate, it is important to consider the number of papers the authors have published to date. One plausible explanation for the lower self-citation rates among first authors could be attributed to their relatively junior status and short publication record. As such, it would also be beneficial to assess self-citation rate as a percentage relative to the author's publication history. This number would be more accurate if we look at it as a percentage of their publication history. My suspicion is that first authors (who are more junior) might be more likely to self-cite than their senior counterparts. My suspicion was further raised by looking at Figures 2a and 3. Considering the nature of the self-citation metric employed in the study, it is expected that authors with a higher level of seniority would have a greater number of publications. Consequently, these senior authors' papers are more likely to be included in the pool of references cited within the paper, hence the higher rate.

While the authors acknowledge the importance of the number of past publications in their gender analysis, it is just as important to include the interplay of seniority in (1) their first and last author self-citation rates and (2) their geographic analysis.

4. Because your analysis is limited to high impact journals, it would be beneficial to see the distribution of the impact factors across the different countries. Otherwise, your analysis on geographic differences in self-citation rates is hard to interpret. Are these differences really differences in self-citation rates, or differences in journal impact factor? It would be useful to look at the representation of authors from different countries for different impact factors.

5. The presence of self-citations is not inherently problematic, and I appreciate the fact that authors omit any explicit judgment on this matter. That said, without appropriate context, self-citations are also not the best scholarly practice. In the analysis on gender differences in self-citations, it appears that authors imply an expectation of women's self-citation rates to align with those of men. While this is not explicitly stated, use of the word "disparity", and also presentation of self-citation as an example of self-promotion in discussion suggest such a perspective. Without knowing the context in which the self-citation was made, it is hard to ascertain whether women are less inclined to self-promote or that men are more inclined to engage in strategic self-citation practices.

این Routing در واقع برای map کردن یک HTTP Request به یک Function است. مثل اینکه این Routing میتونن به انواع و اقسام HTTP Requst جواب بدن مثل GET و POST و غیره. در واقع در نهایت یک Routing Table درست میشه که هر URL به یک تیکه کدی وصل می کنه و به این Routing Table یا Routing Engine کفته می شود.(یاد مفهوم Routing Table در لایه 3 یا network افتادم)

برای status code می تونی بعد اون return سااده بنویسی 200 return "Would you like some tea?", 200

همیشه برای Return شدن میومدیم از render_template استفاده می کردیم ولی ایندفعه کار باحالی کرده و با make_response این کار را کرده. چه جالب اینجوری Set می کنه. دقت داشته باش که با این make response میشه خیلی کارا کرد از جمله Set کردن Status Code برای Response و mimetype و headers.

• myResponse = make_response('Response')
• myResponse.headers['customHeader'] = 'This is a custom header'
• myResponse.status_code = 403
• myResponse.mimetype = 'video/mp4'

Where to find example scripts for sui? I know about tests. Or interaction via ts scripts but not via move files.

pub

I wonder how much of this is accomplishable by automatically parameterizing code by the types that aren't used internally so they implementation can forget about the specifics. In addition some sort of meta-programming capability to automatically generate arbitrary instances or a richer form of trace types for user types would go a long way to simplifying the trace generation.

For a second there, I thought this website was promoting open source and was actually sharing the (source) code for their website...

But alas, it's actually for sharing a different kind (QR) of code instead...

Should we also provide instructions for running the built version of R?

We can probably delete this for the moment. That makes for simpler instructions

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

Learn more at Review Commons

Referee #1

Evidence, reproducibility and clarity

Uromodulin (Tamm-Horsfall protein) is the most abundant protein excreted in human urine.<br /> It plays role in protection against urinary tract infections and renal stones. Mutations in UMOD gene encoding uromodulin cause Autosomal Dominant Tubulointerstitial Disease (ADTKD) that slowly progresses to chronic kidney disease.

In this manuscript, Schiano et al. isolate 12 missense UMOD mutations, which they classify into two groups by age occurrence. They then proceed to study two of these mutations: one from the earlier-onset - Arg185Ser - and the second from the later-onset - Cys170Tyr.

The authors generate UmodC171Y and UmodR186S knock-in mice with distinct dynamic pathways impacting on ADTKD progression. These mutations are equivalent with UMOD mutations (C170Y and R185S) in patients. UmodC171Y and UmodR186S knock-in mice show impaired uromodulin biogenesis, with strong allelic and gene-dosage effects. The trafficking problem of ADTKD-UMOD mutants, involving ER retention, ER stress, and activation of the UPR is recapitulated in mIMCD-3 cells, where the R185S mutant reveals more aggregates that are triggering PERK and IRE1 pathways and ER stress responses.

The manuscript is well written, experiments are in general well described and performed, results offer important insights on cellular events eventually leading to organ damage in ADTKD resulting from missense mutation in the UMOD gene.<br /> The part of the work investigating the degradation mode of two different UMOD mutants, one relying on proteasomal and one relying on lysosomal clearance, is the most interesting for a general audience. Unfortunately, this last part of the work is too preliminary to be accepted as it is.

Comments/Suggestions:

• Selection of the UMOD variants, page 5: "R185S and C170Y are the most prevalent mutants in the clusters" please document/add reference.
• Fig. 1D: please show the position of the insets in the UMOD and BiP panels. Please separate the IF panels from the Picrosirius red panels (these are not the same samples that are shown),<br /> Formally, the BiP panels in Fig. 1D reveal that there is more BiP in cells expressing R185S. That this correlates with UPR induction (as confirmed in Fig. x) should be written at the end of page 5 to make this issue clear for non-experts.<br /> In Fig. 1D, the signal of BiP is not visible in WT and C170Y tissue/cells, which is odd because BiP is abundant protein. Moreover, the differences in BiP levels quantified in WB (semi-quantitative analyses) are not that dramatic in the mouse model (SFig. 3). Which panel in SFig. 3 (mouse) should be representative of the IF shown in Fig. 1D (patients)?<br /> Fig. 1D: Magnification of these images is not sufficient to conclude that R185S accumulates in the ER, and that WT and C170Y are at the apical cell's membrane as written (page 5). Authors should refer to Suppl Fig 1C, where individual cells are visible.<br /> Authors should briefly explain at the end of page 5 how the P. red staining in Fig. 1D informs on fibrosis.
• In the analyses of misfolded UMOD mutants (e.g., Fig. 2, 3, 4, ...) one would expect a test showing that BiP associates with R185S>C170Y>WT.
• Fig. 2F: in R186S there is a dramatic enlargement (at least 2x) of nuclei. Can the authors comment on that?
• Fig. 7E: Shouldn't one expects apical signal for C170Y?

• Fig. 7F: Why there is apical signal for R185S (and not for C170Y)?

• The part covering the degradation of the two UMOD variants would be of great interest for a wide audience of cell biologists. However, these data are too preliminary and, in this form, inconclusive.<br /> Few examples: MG132 is a non-specific inhibitor of the proteasome, which may enhance endogenous and trans-gene expression (check in Pubmed "mg132 promoter" for relevant literature). Thus, an increase in the intracellular level of C170Y on MG132 treatment does not necessarily indicate inhibition of the protein's proteasomal turnover. It could also, at least in part, be caused by an increased synthesis of UMOD. The authors should show that MG132 does not increase synthesis of mutant UMOD (or use the more selective proteasome inhibitor PS-341 in their experiments); similarly, the data on R185S do not prove that this protein is client of autophagy. They rather show that autophagy removes the protein when cells are under nutrient restriction (note that starvation activates bulk autophagy, the non-selective lysosomal clearance of cellular components). To show that misfolded R185S is removed from cells by misfolded protein-induced ER-phagy (i.e., ER-to-lysosome-associated degradation), the authors should monitor in WB the accumulation of R185S in the presence of BafA1 and/or in IF the accumulation of R185S within lysosomes in the presence of BafA1.

Minor comments

• Figure 1B: dotted lines should be defined in the legend.
• Figure 1C: "phenotypes are denoted as indicated". The color-code used for the phenotype is unclear to me. For example, what is the phenotype of the V.2 (grey square)?
• The meaning of "Unlike in UMOD R185S cells, higher SQSTM1 puncta colocalizing with uromodulin were initially present in C170Y mutant cells and further accumulated in MG132-treated cells (Supplementary Figures 10A, B). These data suggest that mutant cells respond differently to UPS inhibition, with C170Y mutant uromodulin being mainly targeted to this pathway." (page 14) and the interpretation of the results shown in 10A and 10B is unclear to me.
• Page 7: "The UmodC171Y mice showed a progressive increase in BUN at 4 months" please define BUN.
• Please, provide a complete list of primary antibodies used for immunoblotting, immunohistochemistry, and immunofluorescence staining.

Significance

The manuscript is well written, experiments are in general well described and performed, results offer important insights on cellular events eventually leading to organ damage in ADTKD resulting from missense mutation in the UMOD gene.<br /> The part of the work investigating the degradation mode of two different UMOD mutants, one relying on proteasomal and one relying on lysosomal clearance, is the most interesting for a general audience. Unfortunately, this last part of the work is too preliminary to be accepted as it is.

My expertise: protein quality control, ER-phagy

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

Learn more at Review Commons

Reply to the reviewers

1. General Statements [optional]

Reply to general assessment of referee #2:

1. General assessments: The current study adds some to these observations…some of these observations are incremental…biological significance is limited. While this reviewer does not suggest additional experimentation, this manuscript would be suitable as a resource paper.

Reply: It appears we were not clear enough in explaining the novel aspects of our study.

The starting points are two published studies from our lab demonstrating a global increase of ISGF3 association with ISG promoters in IFNγ-treated cells and a remarkable similarity of IFN-γ and type I IFN-induced early transcriptome changes. These findings challenge the notion in the field (as mentioned by the referee) that IFNγ specificity is produced by the predominant deployment of STAT1 homodimers. We thus tested the hypothesis that the specificity of the IFNγ-induced transcriptome is generated over time, rather than during the early response, and relies on secondary responses to transcription factors such as IRF1. In contrast, IRF1 plays no or only a small role in the type I IFN response that utilises ISGF3 and/or unknown secondary factors in the delayed response. We tested this hypothesis with PRO-seq technology to rule out confounding effects of mRNA processing over a 48h period. The data are clear in showing that many genes associated with the antibacterial or anti parasite profile of activated macrophages are indeed much more abundant in late-stage rather than briefly IFNγ-treated macrophages and these delayed changes are to a large extent dependent on IRF1. Our findings are based on the best available technologies, a combination of nascent transcript analysis with genetics and protein interaction studies. In addition, our findings rule out alternative models of sustained or secondary ISG transcription, such as the employment of alternative ISGF3 complexes (such as STAT2-IRF9) or of ISGF3 complexes formed with unphosphorylated STAT1 and STAT2. We provide evidence for higher order waves of transcription caused by unknow transcription factors that are produced by transcriptional activation of ISGF3 or IRF1 target genes and identify candidates among the AP1 and Ets transcription factor families. We agree that some of the data are confirmatory rather than novel (i.e. some of the genes we describe were known from previous literature to be IRF1 targets), but it is the systems approach of our study, and particularly the delineation of conditions under which the largely neglected delayed response diverts the IFNβ and IFNγ-induced transcriptomes, that generates a comprehensive and conclusive view of IFNγ acting predominantly as a macrophage activating factor, and IFNβ being an essential antiviral cytokine. We do think this main outcome is immunologically meaningful and not incremental. For this reason, we would prefer to publish the paper as a relevant contribution to innate immunology rather than a resource. Emphasizing our point, a paper appeared in ‘Cell’ while our study was under review, showing that human IRF1 mutations cause mendelian susceptibility to mycobacterial disease (MSMD), a term coined by JL Casanova and colleagues for immunological defects that reduce the ability of macrophages to cope with intracellular bacteria (new ref. 65). This important study emphasizes the main conclusions of our study about the relevance of IRF1 for macrophage activation. We discuss this paper on p. 14 lines 9-14.

Revision: We tried to better explain the scientific motivation for this study and the significance of the results (p. 4, lines, lines 12-25).

Revision plan: n. a.

2. Description of the planned revisions

Referee #3; major comment 1:

In Fig. 1d is difficult to interpret and misleading for many reasons. First, the cluster numbering is disconnected from the cluster order; why not numbering them based on the hierarchical clustering and writing the cluster number besides the cluster itself? Second, having a 2-color gradient is misleading; negative values shouldn't be in the same color tone than the positive values. Third, the authors did not provide adequate rationale behind using only the top 1,000 most expressed gene? Why not using all the differentially expressed genes in at least one of the condition to provide a comprehensive analysis? Could this potentially lead to bias in the data, and is there any information lost by not using the - lower - expressed genes fraction? Fourth, it is not clear what the color scale is representing and how the data was transformed. Was a mean centering of the expression values of the log2FC applied to the RNA-seq data to facilitate clustering? Mean centering and z-scoring is a common technique used to adjust expression data, but it can potentially exaggerate differences between samples. More information about the data and analysis should be provided, as it is difficult to determine whether this was a valid approach or not.

Reply:

• To create the heatmap, we used the pheatmap package from R and the cutree_rows option to separate 11 clusters with strikingly different patterns of gene expression based on visual exploration. The numbering was autogenerated by the program.
• The data is now shown in red-blue.
• We restricted our list to only 1000 genes from each comparison as we aimed to analyze the prominent patterns of gene expression across timepoints. Considering all differentially expressed genes based on a padj value would also include genes expressed at very low levels as evident from the low baseMean values obtained from DESeq2. Hence, we applied a selection of 1000 genes which effectively represented the major patterns of gene expression across timepoints.
• Variance stabilized transformation was applied on read counts obtained from PRO-seq using the DESeq2 package. The transformed reads were z-score normalized and used for performing hierarchical clustering by the “Ward.D2” method using the pheatmap package in R. A total of 3126 genes were used for this analysis. 11 distinct clusters were defined using cutree_rows option. The color scale represents z-score normalized counts. The genes represented in the heatmap were selected based on the following criteria: each timepoint of interferon treatment was compared to the homeostatic condition (untreated sample) in wildtype BMDMs. The differentially expressed genes from each comparison were selected based on the filtering criteria: absolute log2FoldChange >=1 and adjusted p value <0.01 by Wald test. Following the differential analysis, the first 1000 differentially expressed genes in each treatment condition (ordered based on adjusted p values) were selected for both IFN types and combined and selected for creating a list which consisted of 3126 unique genes. The scale in the heatmap represents z-scores of variance-stabilized reads, calculated across all genotype and treatment conditions, separately for each IFN type.

Revision plan: We will label the clusters with the cluster number next to it in addition to the color codes.

Referee #3; major comment 3:

The large standard deviation bars in the claim that ChIP data confirmed the binding of ISGF3 components to the promoter of Mx2 cast doubt on the validity of the results and conclusions. The authors should consider additional experiments or complementary analyses to validate their findings. Or alternative, to adjust their claims accordingly.

Reply: To demonstrate sufficient quality of the data the ratio of Stat1/ Stat2 was calculated for early (1.5hrs) and late (48h) separately. The unpaired two-tailed t test comparing this ratio between 1.5 hrs and 48hs, shows that they are not significantly different. This indicates that all ISGF3 components are associated with ISG during both early and delayed responses, i. e., that STAT2/IRF9 complexes are unlikely to contribute to delayed ISG control. However, we agree with the referee that the standard deviations of the kinetic ChIP experiment are high and that it would be good to generate additional data.

Revision plan: We will perform additional ChIP experiments to improve the statistical power of the results in fig. S2c.

Referee #3, major comment 6:

The authors interpret their ATAC-seq and ChIP-seq results based on a 2kb window to the TSS of genes, not considering relatively close enhancers or longer range cis-regulatory interactions in their interpretation. For example, they mention on p.7 "Contrasting the strong binding of IRF9 and IRF1 to the Mx2 (cluster 2) and Gbp2 (cluster 9) promoters, respectively, we saw no evidence for direct binding to Lrp11 (cluster 3) and Ptgs2 (cluster 10)", but on Fig 3d they show only the proximal regions. No scale bars are shown either. Moreover, exploring the same published IRF1 ChIP-seq dataset, there is a clear IRF1 binding site at the promoter of Ptgs2, while the authors report none.

Reply:

• According to the literature (e. g. refs. 11, 27), most IFN-induced accessibility changes occur in the vicinity of the TSS of ISG. This is further strengthened by the data shown in this manuscript. In addition, most functionally validated GAS and ISRE sequences are in the DNA interval chosen for our analysis. While distal ISG enhancers have been reported (e. g. DOI: 10.26508/lsa.202201823), an analysis beyond the placement of most control regions increases the risk of wrong assignments between ISG and their regulatory elements, hence the causality between transcription factor binding and accessibility changes.
• We extended the regions for the analysis of the Lrp11 and Ptgs2 regulatory regions and found no evidence for the binding of ISGF3 or IRF1. We find no evidence for a clear peak in the Ptgs2 promoter. There is a peak called by the Macs2 algorithm, but visual inspection of the track (bigwig file) shows it consists of a minor increase in reads above background that does not suggest a bona fide IRF1 binding site (see below). This view is supported by our inability to find an IRF binding site in the vicinity of the peak.

IRF1 binding indicated by bigWig browser tracks and corresponding peakfiles detected at the locus. We identified the peakfile from Langlais et al., 2016 and identified peaks using MACS2, however using mm10 genome as the analysis in the original paper was done with mm9 genome. The peak identified here appears to be an artefact of the MACS2 program as there is no evident enrichment at the gene promoter region upon inspection of the bigWig files.

Revision plan: Scales will be added to the browser tracks as requested.

Referee #3, major comment 7:

Lack of statistical analysis on chromatin accessibility claims: The authors claim that ATAC-seq data in BMDMs stimulated with IFNβ or IFNγ for a short (1.5 hours) or long (48 hours) period reveals a striking similarity between transcription and the general trends of chromatin accessibility at regions up to 1000 bp upstream of the TSS (Fig. 2a), suggesting continuous chromatin remodeling during the transcriptional response. However, I would like to know if this conclusion is well-supported by the correlation between the chromatin accessibility from ATAC-seq data from only one sample and the PRO-seq data.

Reply: See revision plan.

Revision plan: We will analyze single experiments whether they support the conclusions derived from the z-score of the triplicate samples.

Referee #3, major comment 8:

The need for additional experiments to verify claims such as the dependence of Ifi44 on IRF1 for gaining ATAC signal, as stated in the claim, "Expression required IRF1 for both, but accessibility of the Ifi44 regulatory region depended upon IRF1 whereas that of Gbp2 acquired an open structure independently of IRF1 (Fig. 5c).

Reply: We think the lack of clarity might be related to the size of figures 5a and 5b and the density of the dots in some areas of the plot. We agree it is very difficult to assign our gene labels unambiguously to a single dot.

Fig. 5a combines ATACseq data in wt and IRF1 knockout cells with the expression data from the Pro-seq experiment, Fig. 5b is the same set-up, but IRF9-deficient macrophages are analyzed.

Blue dots show ATACseq signals induced by IFN treatment. Violet dots represent genes that require IRF1 (Fig. 5a) or IRF9 (Fig. 5b) for transcriptional induction. Yellow dots mark genes such as IFI44 requiring IRF1 (Fig. 5a) or IRF9 (Fig. 5b) for both expression and the accessibility change in the promoter region. Fig. 5c visualizes representative examples of genes whose accessibility is coupled to the transcription factor dependence of the transcriptional induction (IFI44), or not (Gbp2). Thus Fig. 5c must be interpreted based on the dot color code in fig. 5a and we admit this has been difficult with the figure in its present form.

Revision plan: We will improve the clarity of figs 5a and 5b in several ways:

• We will label the panels to better indicate the intersected data sets.
• We will increase the size of the panels and figure legends and make sure that the correspondence between gene names and dots are unambiguous.
• We will include trend lines of the Ifi44 and Gbp2 genes to visualize their induction and IRF1 dependence.

Referee #3, major comment 13 (see also section 3):

The authors have not adequately addressed the methodological limitations in their discussion, which extends beyond the aforementioned comments. It is suggested they include a comprehensive discussion of the claims made pertaining to the necessity of IRF1 for accessibility and the potential biases in the interactomes, along with their associated consequences.

Reply: The contribution of IRF1 to the accessibility of ISG promoters emerges from the data in figures 5a, whose clarity will be improved (see reply to point 8). We do not interpret the impact of IRF1 beyond the data, in fact we state a relatively minor effect of IRF1 in the control of promoter accessibility (p. 10, lines 20-22) and we have added a reference in agreement with an impact of IRF1 on basal expression of antiviral genes (ref. 39, as suggested by the referee).

We have added discussion on potential limitations of the TurboID approach (p. 11, lines 22-24 and p. 15, lines 3-11).

Revision plan: Improvement of fig 5a (see ref. #3, point 8).

Referee #3, minor comment 2

Fig 1e. The color scales on the GO enrichment graphs are misleading since they use the same blue-to-red gradient for adj p-values ranging from 10-25 to 10-49 and 0.008 to 0.016, which could be considered non significant.

Reply: We agree that this is confusing. It results from automated assignments of the color gradients by the software.

Revision plan: We will investigate possibilities to change color codes for different ranges of p values.

Referee #3, minor comment 4

The incomplete schema in Figure 1a, which only focuses on PRO-seq and does not include the ATAC-seq element.

Reply: We will add a new figure to visualize the set-up of the ATAC seq experiments and their intersection with the Pro-seq data.

Revision plan: We will add a new figure in accordance with the referee’s request.

Referee #3, minor comment 6

The clearer labeling of Figure 5a and 5b.

Reply: Please refer to our reply to major point 8.

Referee #3, minor comment 10

Fig S1b, S3b. The PRO-seq was generated in triplicates, hence these graphs should include the Log2FC for the individual data points.

Reply: The Log2FC from DESeq2 were calculated from the triplicates, the software does not compute Log2FC from individual replicates.

Revision plan: We mention the p-values for the Log2FC to show the degree of consistency (figure legends). We will provide a table with log2FC and corresponding padj values of the genes represented at each timepoint (table_showing_padj_values_and_log2fc).

Referee #3, minor comment 12

In the genomic snapshot shown, only bars or fading triangles are shown in place of the gene body. The authors should provide an accurate gene structure; i.e., exons and introns.

Reply: We will try to include the exon-intron structure wherever the size of the figure allows this.

Revision: n. a.

Revision plan: If figure size permits, we will add the exon-intron structure of the genes in browser tracks as requested.

3. Description of the revisions that have already been incorporated in the transferred manuscript

Referee #1, major comment 1

Figure 2. Difficult to interpret data as it is presented. Consider quantifying figure 2C in order to make "changes in Pol II pausing were more pronounced during IFNb signaling" statement more apparent.

Reply: We presented the pausing data in two different graphic representations (figures 2c and S2) to make the understanding of the information content easier. In hindsight we may have generated more confusion than clarity.

Revision: We removed the original figure 2c and replaced it with original figure S2. This representation is quite intuitive as the graphs represent a direct quantitative logarithmic display whether and how much the relative amount of paused polymerase changes when comparing IFN-treated and untreated cells. The calculation of these ratios is now explained better in the legend to figure 2.

Referee #1, major comment 2

How are you distinguishing autocrine signaling in the BMDMs driven by IFN treatment from late transcripts (for example, at 48 hours are differential genes due to autocrine cytokine signaling or are they truly late transcripts)?

Reply: We do not exclude autocrine effects. In case of ISG, the most likely autocrine factor would be secreted interferon. According to our Proseq data, the differentially expressed genes do not include any interferon genes. That being said, it is possible that the transcription factors from the AP1 family we hypothesize as drivers of secondary or tertiary waves of transcription are activated by non-IFN cytokines secreted from IFN-treated cells (see also reply to comment 3).

Revision: We now mention that enhanced IFN production is not sustaining ISG responses (p.5 lines 18/20). We mention the possibility that secreted factors may drive secondary or tertiary waves of ISG transcription (p. 8, lines 21/23).

Referee #1, major comment 3

Figure 3D. Authors choose Gbp2 (as positive control for IFNg driven gene), but don't show that Gbp2 is a IFNb independent gene. Consider using IRF1 KO BMDMs in this data as well.

Reply: This is a misunderstanding. Gbp2 is not shown as an IFNγ-specific gene (it’s induction by both IFN types has been shown previously and emerges from our Pro-seq analysis, see also response to minor issue no. 2). It represents the cluster of genes that are sustained specifically after IFNγ treatment in an IRF1-dependent manner. The purpose of fig. 3D is to show that not all ISGF3/IRF9-dependent genes have promoter binding sites for ISGF3 and not all IRF1-dependent genes have binding sites for IRF1. This suggests indirect effects of both transcription factors in sustaining IFN-induced transcription (in line with the referee’s comment 1).

Previous figure S3e (now S2f) confirms binding of IRF1 to the GBP2 promoter by ChIP with kinetics correlating to its transcriptional effect. This experiment is normalized with an IgG control. IRF1 knockout cells did not produce a ChIP signal with IRF1 antibody, as expected (data not shown).

Revision: We better explain the rationale behind the experiments shown in figure 3D (text on p8, lines 12-16). In addition, we show the trend line of Gbp2 expression in WT vs IRF1KO as well as that of additional genes showing delayed/sustained responses in the new Figure S3.

Referee #1, minor comment 2

Define known IFNg and IFNb driven genes when they are introduced in figure 2 rather than in discussion.

Reply: Following the referee’s suggestion we provide the examples of IFNβ and IFNγ-controlled genes and the characteristics of their regulation in the context of our description of the results displayed by fig. 2 (p.6 lines 15-21). This includes Gbp2 (see major issue no. 3).

Revision: The text on p. 6 lines 15-21 has been modified in accordance with the request.

Referee #1, minor comment 4

Unclear whether IRF1 expression in figure 3A is from whole cell lysate or nuclear fraction.

Reply: We indicate in the figure legend that whole cell lysates were used.

Revision: We added a sentence with the relevant information in the legend of figure 3.

Referee #1, minor comment 5

Authors suggest IFNb treatment induces less IRF1 at later time points, however loading control also seems slightly lower than other considerations. Is it possible that IFNb treated cells are dying at later time points, given that type I IFN signaling can be pro-apoptotic.

Reply: The graph below the blot represents quantified IRF1 signals, normalized to the loading control. It shows that the differences are not generated by unequal loading of the blotted gel. We and others have shown that IFNβ may indeed enhance macrophage death, however only when the cells are simultaneously infected with an intracellular pathogen (e.g. new ref. 25). These studies also show that treatment with IFNβ alone over periods used in the present study does not affect macrophage viability.<br /> Revision: We added a sentence about the viability of IFN-treated macrophages (p. 4, lines 31-32).

Revision plan: n. a.

Referee #2, major comment 3

The sequencing and BioID data are not submitted to public databases.

Reply: An accession number has been added.

Revision: The accession number was added on p.29, line 25.

Referee #3, major comment 1 (see also revision plan, section 2):

Revision: The rationale for using the top 1.000 genes is explained (p.5, lines 7-9). The description of the pro-seq read count processing has been extended in accordance with our reply to the referee in the legend of figure 1d and in the methods section (p. 33, lines following line 10.)

Referee #3, major comment 2

Fig 2c. The authors claim that RNA Pol II pausing is a major factor in controlling the dynamics of ISG transcription. However, they did not provide sufficient explanation of the results, and in all fairness there is not much variation between the clusters to sustain the claim that this is a major factor in ISG transcriptional control.

Reply: We agree with the referee that we cannot posit RNA pol II pausing as a major factor for the differences of transcriptional control of ISG in individual clusters. We have made sure to remove any statements suggesting this possibility. We also try to better integrate our findings with RNA pol II pausing into the existing literature.

Revision: We added relevant literature on p. 6 lines 28-30 and p. 7, lines 4-6.

Referee #3, major comment 4

On p.5, the authors mention "Representative browser tracks from the Gbp2 and Slfn1 genes further validate this observation" but they are simply referring to genome browser snapshot, i.e., specific genomic examples, extracting from the same single dataset. Without using an independent dataset, this can not "further validate" the initial findings.

Reply: We agree the wording is incorrect.

Revision: We changed the paragraph describing this experiment (p. 6, lines 15-21).

Referee #3, major comment 5

IRF1 was successfully pulled down with STAT1 bait but not in the reciprocal experiment. The author should discuss this point as it is important for the conclusions. Could it potentially indicate issues with the technique used, and if this could introduce any bias into the results. The statement, "In contrast, interactors of the IRF1 bait did not include STAT1. This discrepancy could result from steric constraints of the tagged proteins due to the limitation of the 10nm distance reached by the biotin ligase," does not seem to be sufficient to explain this discrepancy.

Reply: STAT1 was present in the IRF1 pull-down and the interaction increased significantly after IFN treatment but after normalization to the NLS control it did not conform to our criterium of a 95% confidence interval for the FDR. To be consistent we did not include it in the list of IRF1 interactors. We have observed on several occasions that the significance of proximity is not reciprocal, even for well- documented physical interactions. A prime example for this is the interaction between STAT1 and IRF9 in IFN-treated cells which is recorded in the STAT1 pull-down, but not that with IRF9 (ref. 10). Apart from steric reasons the lack of reciprocity may result from different signal/noise ratios in pull downs with different baits.

Revision: We mention that IRF1 was a STAT1 interactor below the statistical cut-off (p. 11, lines 26-28) as well as the possibility of different signal/noise ratios in the IRF1 and STAT1 pull-downs on p.11, lines 22-24.

Referee #3, major comment 9

In the figure legends, there is missing information about the number of times experiments were replicated, suggesting that some were done a single time. Moreover, some graphs are missing statistical analysis, e.g., in Fig S3cS3e, S3f, the ChIP-qPCR experiments were done on biological triplicates, there is no mention of statistical test performed, it is not mentioned what the error bars represents (SD, SEM, etc.) and the variance is large, but the authors still interpret these results as significant enrichment of the transcription factors to the Mx2 promoter.

Reply: Where missing the relevant information has been added to figure legends. In brief, all experiments represent at least three biological replicates. The only exception is the western blot shown in figure S3a, (no S2a) which represents two independent replicates. Here, the clarity of the difference of IRF1 expression and the fact that the only purpose is to show that Raw264.7 macrophages behave like bone marrow-derived macrophages in fig. 3a justifies the omission of another replicate (please see also answer to point 3).

Revision: The relevant information has been added to figure legends where necessary (figs. 1, a, 3a, 6a-f, S1, S4, S5).

Referee #3, major comment 10

Another example are the RNA Pol II pausing index ratios, which show minor variations and not are supported by statistics to support a possible significance. Proper description, replication and statistical analyses of the results are critical.

Reply: We agree.

Revision: Statistics underlying the RNA Pol II pausing data are included in supplementary data 2.

Referee #3, major comment 11

The authors used CRISPR-Cas9 genome editing to generate knockout cell lines. However, they did not verify the knockouts at the protein level. Further experiments could confirm that the targeted proteins are not expressed in the knockout cell lines.

Reply: We included a western blot showing the lack of IRF1 and STAT1 expression in the respective cell lines.

Revision: New figure S6.

Referee #3, major comment 12

On p.9, it is mentioned "IRF1 affects chromatin structure ...". Here chromatin structure is related to minor changes in chromatin accessibility, this can not be qualified as changes in chromatin structure.

Reply: ‘structure’ has been changed in accordance with the request.

Revision: ‚structure‘ has been replaced with ‘accessibility’. (p. 10, lines 19 and 21).

Referee #3, major comment 13 (see also section 2, revision plan, major comment 8)

The authors have not adequately addressed the methodological limitations in their discussion, which extends beyond the aforementioned comments. It is suggested they include a comprehensive discussion of the claims made pertaining to the necessity of IRF1 for accessibility and the potential biases in the interactomes, along with their associated consequences.

Reply: The contribution of IRF1 to the accessibility of ISG promoters emerges from the data in figures 5a, whose clarity will be improved (see reply to point 8). We do not interpret the impact of IRF1 beyond the data, in fact we state a relatively minor effect of IRF1 in the control of promoter accessibility (p. 10, lines 20-22) and we have added a reference in agreement with an impact of IRF1 on basal expression of antiviral genes (ref. 39, as suggested by the referee).

We have added discussion on potential limitations of the TurboID approach (p. 11, lines 22-24 and p. 15, lines 3-11).

Revision: Change of the discussion section (p. 11, lines 22-24 and p. 15, lines 3-11).

Revision plan: Improvement of fig 5a (see ref. #3, point 8).

Referee #3, major comment 15

The work should be discussed in the context of the demonstrated physiopathological evidence of the IRF1 and IRF9 functions. IRF9 (Hernandez et al., JEM 2018) and more recently IRF1 (Rosain et al Cell, 2023) were identified as causing non overlapping phenotypes in human patients carrying loss-of-function mutations for these genes. The authors must interpret their results in this context.

Reply: We thank the referee for reminding us about the importance of these papers for our work.

Revision: The papers have been mentioned and discussed (p. 13 lines 19-28 and p.14, lines 9-14).

Referee #3, minor comment 3

The inconsistency in the title referring to IFNb as Type 1 but using IFNg instead of Type 2 nomenclature, perhaps consistency is best.

Reply: We agree about the importance of consistency but find ourselves in yet another quandary. While the use of ‘type I IFN’ is clearly indicated and widely used as a collective name for this group of cytokines, the use of ‘type II IFN’ for IFNγ is rare because it is the only member of this type. Hence, we decided for sticking with convention at the expense of a bit of consistency. We agree about the title, though, and have changed type I IFN to IFNβ.

Revision: We adapted the title in agreement with the referee’s comment.

Referee #3, minor comment 5

Figure 6d includes a color scale of -1 to +3, but it is unclear what these values represent and how they were calculated per interactor. The figure legend should be revised to clarify this information.

Reply: We agree. The relevant information has been added to the figure legend.

Revision: We added information (log2FC with regard to the NLS control) to the legend of fig. 6d.

Referee #3, minor comment 9

Fig 1e, S1c. Graphs having circles of varying sizes in function of a value are named "bubble plots" and not "dot plots".

Reply: Thank you for pointing this out, we corrected our mistake.

Revision: We changed dot plot to bubble plot in legend to figure S1c.

Referee #3, minor comment 11

Fig S3c legend. It is mentioned "Graph represents RT-qPCR of genomic Mx2". RT-qPCR usually stands for reverse transcription quantitative PCR, hence we suggest to change to "ChIP-qPCR" or qPCR. Confusingly, in the literature the term "RT-PCR" is used for real-time PCR and "qPCR" for quantitative PCR. Also, the authors should be specific about the "genomic" region targeted; the graphs mention "promoter", hence it would be appropriate to use the same designation in the legend.

Reply: We agree and thank the referee for correction of the terminology.

Revision: We changed RT-PCR to qPCR throughout the manuscript. Moreover, we specifically refer to ‘promoter region’ as the amplified DNA.

Referee #3, minor comment 12

Fig S3e. The y-axis names are missing.

Reply: Thanks for spotting this.

Revision: The y axis in the figure received its proper label.

Referee #3, minor comment 14

Raw cells are sometimes spelled as "Raw" and other times as "RAW". Please choose one for consistency.

Revision: This inconsistency has been corrected

Referee #3, minor comment 15

In p.10 l.20, the figure number is missing.

Revision: We corrected this mistake.

4. Description of analyses that authors prefer not to carry out

Referee #1, minor comment 1

Simplify figure 4B- consider focusing on most differentially expressed genes between clusters

Reply: The purpose of fig. 4B is to provide a visual overview of the kinetics of eRNA transcription in response to both IFN types and of the effects of IRF9 and IRF1 knockouts. This information needs to be given to demonstrate the similarities and differences between the control of eRNA and the corresponding ISG transcripts in the different regulatory clusters (as shown in figs. 1d and 2a).

Simplifying the figure would mean to separate it according to time point, IFN type treatment or knock-out effect. We think this would require to mentally reassemble the figure to understand the interrelationships between these parameters. To our opinion the visual display of the data interrelationship in fig. 4B facilitates the impropriation of the information content.

Revision: n. a. - we hope our reasoning has become sufficiently clear.

Revision plan: n. a.

Referee #1, minor comment 3

Clarify which cell types (IRF1 KO vs IRF9 KO) are used in figure 5 A/B.

Reply: The cell type (bone marrow-derived macrophages) is mentioned in the first sentence of the figure legend. Since all experiments except the Bio-ID experiment were performed with this cell type we decided not to label each figure.

Revision: n. a.

Revision plan: n. a.

Referee #2, major comment 2 and referee #3, major comment 14

Ref #2: Biological significance is limited as this study is largely descriptive and they do not test the hits obtained from BioID.

Ref #3: Although the TurboID experiments identify known STAT1 and IRF1 interactors, the proposed new interactors are numerous, and none are validated through independent co-IP experiments. Moreover, the results are very noisy, with little differences between untreated BMDMs (where IRF1 is barely expressed) and IFN-treated conditions.

Reply: The big advantage of BioID or TurboID is the ability to score proximity and very transient interactions. Validating BioID hits with technologies such as coIP is not particularly useful as the two technologies will obviously produce different interactomes. In fact, we show in this manuscript that IRF1 and STAT1 show proximity, but they do not form a stable complex under co-IP conditions. This leaves genetic approaches (LOF or GOF) as alternatives. However, apart from the workload (> 100 genes would have to be knocked out or their products overexpressed), most of our hits are expected to produce very broad effects in such experiments, hard to interpret regarding ISGF3 and IRF1 activities.

In view of this situation, we publish exclusively the high confidence nuclear interactors identified in our screen: biological replicates were performed in triplicate, a stringent internal control (TurboID-NLS) was used, and a stringent statistical cut-off for high-confidence interactors (95% FDR between groups) was applied. We further account for the experimental situation by limiting interpretation of the data to confirmed molecular events. For example, STAT1 dimers and the ISGF3 complex are required for histone acetylation in response to IFN, and ISGF3 is known to contribute to the exchange of the H2AZ histone variant (refs 11, 14, 71, 72). Our data show that IRF1 contributes to promoter accessibility changes and this is in line with its proximity to a remodelling complex. Thus, the BioID data indeed validate previous findings. However, in agreement with the referee’s comment, some of the data remain descriptive (such as the intriguing proximity of both STAT1 and IRF1 to nuclear products of ISG). To determine the importance of this molecular proximity is a major undertaking and beyond the scope of this study.

Revision: We added discussion to state the difficulty of validating TurboID-based interactions and the limitations of the TurboID experiments (p.15 lines 3-11).

Referee #3, minor comment 1

In most graphs the expression values or log2FC are shown separately for IFNb and IFNg, however in the heatmaps (Fig 1d, S1d) the IFNb and IFNg results are intercalated keeping them side-by-side for each time point, which makes them more difficult to interpret.

Reply: We are in a quandary about the design of the figure. On the one hand our goal is to visualize gene clusters with distinct behaviors for each IFN type. For this purpose, it would be advantageous to separate the IFN types. On the other hand, we aim at showing similarities and differences between genes induced by each IFN type, for this purpose it is better to maintain the current sample order. While understanding the referee’s point, we prefer to keep the figure as it is, because the suggested change will not increase its overall clarity.

Revision: n. a.

Revision plan: n. a.

Referee #3, minor comment 7

The statement that "IFN-I are the more important mediators of antiviral immunity" is not entirely accurate and may be an oversimplification, as there are certainly articles which suggest a larger role for type ll IFN elements than type l (ref: Yamane D et al., 2019 Nature microbiology). While yes, IFN-I plays a critical role in the innate immune response to viral infections, IFNγ also has antiviral activity and is involved in the adaptive immune response to viral infections, and in some instances to a larger extent than IFN l.

Reply: The Yamane et al study (now mentioned on p 10, lines 22-25 and referenced) agrees with our findings because it shows that IRF1 contributes to the basal expression of an ISRE-driven ISG subset. Our statement about the predominant role of type I IFN versus IFNγ refers to genetic data in both humans (mainly Casanova’s work including effects of autoantibodies against type I IFN, see also the paper about human STAT2 deficiency in the June 15th issue of the JCI, https://doi.org/10.1172/JCI168321) and mice (hundreds of papers) showing that disruption of type I IFN synthesis or response causes profound effects of antiviral immunity (i.e. resulting susceptibilities are first and foremost to viral pathogens) whereas susceptibilities as a consequence of disrupting the IFNγ pathway are first and foremost to intracellular nonviral pathogens such a mycobacteria. In fact, the term mendelian susceptibility to mycobacterial disease (MSMD) was coined by Casanova and colleagues to describe a variety of human mutations that include those of the IFNγ, but not the type I IFN pathway.

Maybe more importantly, the Rosain et al. paper mentioned by the referee which appeared in ‘Cell’ while our study was under review, shows that human IRF1 mutations also fall into the MSMD category (new ref. 65). In contrast, the authors did not observe diminished antiviral immunity. This emphasizes the main conclusions of our study about the relevance of IRF1 for macrophage activation. We discuss this paper on p 14. lines 9-14.

Obviously, this does not exclude a role of type I IFN in nonviral infection or of IFNγ in viral infection, in fact much of our own work has been dedicated to a role of type I IFN in infections with L. monocytogenes. Nevertheless, we think that in a generic statement about the difference between type I IFN and IFNγ it is correct to label the former as predominantly antiviral and the latter predominantly as a macrophage activating factor against nonviral, intracellular pathogens.

Revision: We added discussion of Rosain et al. (ref. 65) on p 14. lines 9-14.

Referee #3, minor comment 8

The authors claim that a significant portion of ISG promoters is associated with ISGF3 upon IFNγ receptor engagement and that the transcriptomes of macrophages treated briefly with IFNβ or IFNγ exhibit remarkable similarity and sensitivity to Irf9 deletion. However, I am uncertain about the extent of consensus on this claim.

Reply: The data were surprising but supported by ChIP-seq and RNA-seq in wt and IRF9 ko macrophages (ref 10). Data in a follow-up study (ref. 11) and in this manuscript support our original conclusion by demonstrating the impact of the IRF9 ko on IFNγ responses. Importantly, we don’t claim this is true in all cell types, it may well depend on STAT/IRF9 expression levels and tonic IFN signaling.

Revision: n. a.

Revision plan: n. a.

Ajouter l'éco index pour l'éco conception étant donné qu'il a été codé, et s'il n'est pas trop lourd avec les animations (= crédibilité pour éviter greenwashing/petwashing). Le bouton "bring me on top", avant le footer ?

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 3 Originality (7 points) : 7 Total (20 points) : 20

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 3 Originality (7 points) : 7 Total (20 points) : 20

same - did you run them? Cant see the code

Correct code / solutions (7 points) : 6 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 3 Originality (7 points) : 7 Total (20 points) : 19

Sometimes you think a bit too complicated! A lot of your code could be simplified. But good work

this is wrong - I dont really understand your code.

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 2 Originality (7 points) : 1 Total (20 points) : 13

Not clear what happened with the project.

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 2 Originality (7 points) : 6 Total (20 points) : 18

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 3 Originality (7 points) : 7 Total (20 points) : 20

Correct code / solutions (7 points) : 6 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 3 Originality (7 points) : 7 Total (20 points) : 19

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 2 Originality (7 points) : 7 Total (20 points) : 19

You did not execute your notebook before uploading as per the instructions so I had to exectute myself. Some of the code did not run for me.

Group 4:

Correct code / solutions (7 points) : 6 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 2 Originality (7 points) : 7 Total (20 points) : 18

How much code did you copy paste from your thesis code and how much is new based on what you learnt?

Correct code / solutions (7 points) : 2 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 1 Originality (7 points) : 6 Total (20 points) : 12

Many mistakes in the code for the guided exercises. Grade saved by the free coding project

Correct code / solutions (7 points) : 6 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 3 Originality (7 points) : 7 Total (20 points) : 19

Reviewer #1 (Public Review):

This work provides a new dataset of 71,688 images of different ape species across a variety of environmental and behavioral conditions, along with pose annotations per image. The authors demonstrate the value of their dataset by training pose estimation networks (HRNet-W48) on both their own dataset and other primate datasets (OpenMonkeyPose for monkeys, COCO for humans), ultimately showing that the model trained on their dataset had the best performance (performance measured by PCK and AUC). In addition to their ablation studies where they train pose estimation models with either specific species removed or a certain percentage of the images removed, they provide solid evidence that their large, specialized dataset is uniquely positioned to aid in the task of pose estimation for ape species.

The diversity and size of the dataset make it particularly useful, as it covers a wide range of ape species and poses, making it particularly suitable for training off-the-shelf pose estimation networks or for contributing to the training of a large foundational pose estimation model. In conjunction with new tools focused on extracting behavioral dynamics from pose, this dataset can be especially useful in understanding the basis of ape behaviors using pose.

Since the dataset provided is the first large, public dataset of its kind exclusively for ape species, more details should be provided on how the data were annotated, as well as summaries of the dataset statistics. In addition, the authors should provide the full list of hyperparameters for each model that was used for evaluation (e.g., mmpose config files, textual descriptions of augmentation/optimization parameters).

Overall this work is a terrific contribution to the field and is likely to have a significant impact on both computer vision and animal behavior.

Strengths:<br /> - Open source dataset with excellent annotations on the format, as well as example code provided for working with it.<br /> - Properties of the dataset are mostly well described.<br /> - Comparison to pose estimation models trained on humans vs monkeys, finding that models trained on human data generalized better to apes than the ones trained on monkeys, in accordance with phylogenetic similarity. This provides evidence for an important consideration in the field: how well can we expect pose estimation models to generalize to new species when using data from closely or distantly related ones?<br /> - Sample efficiency experiments reflect an important property of pose estimation systems, which indicates how much data would be necessary to generate similar datasets in other species, as well as how much data may be required for fine-tuning these types of models (also characterized via ablation experiments where some species are left out).<br /> - The sample efficiency experiments also reveal important insights about scaling properties of different model architectures, finding that HRNet saturates in performance improvements as a function of dataset size sooner than other architectures like CPMs (even though HRNets still perform better overall).

Weaknesses:<br /> - More details on training hyperparameters used (preferably full config if trained via mmpose).<br /> - Should include dataset datasheet, as described in Gebru et al 2021 (arXiv:1803.09010).<br /> - Should include crowdsourced annotation datasheet, as described in Diaz et al 2022 (arXiv:2206.08931). Alternatively, the specific instructions that were provided to Hive/annotators would be highly relevant to convey what annotation protocols were employed here.<br /> - Should include model cards, as described in Mitchell et al (arXiv:1810.03993).<br /> - It would be useful to include more information on the source of the data as they are collected from many different sites and from many different individuals, some of which may introduce structural biases such as lighting conditions due to geography and time of year.<br /> - Is there a reason not to use OKS? This incorporates several factors such as landmark visibility, scale, and landmark type-specific annotation variability as in Ronchi & Perona 2017 (arXiv:1707.05388). The latter (variability) could use the human pose values (for landmarks types that are shared), the least variable keypoint class in humans (eyes) as a conservative estimate of accuracy, or leverage a unique aspect of this work (crowdsourced annotations) which affords the ability to estimate these values empirically.<br /> - A reporting of the scales present in the dataset would be useful (e.g., histogram of unnormalized bounding boxes) and would align well with existing pose dataset papers such as MS-COCO (arXiv:1405.0312) which reports the distribution of instance sizes and instance density per image.

I was mistaken about [[OpenAI]]'s bug bounty program; it does not cover jailbreaks or hallucinations.

Limitations/Exceptions

Now reads:

(2) Limitation on application Paragraph (1) does not apply to-

(A) an entity exhibiting animals to the public under a Class C license from the Department of Agriculture, or a Federal facility registered with the Department of Agriculture that exhibits animals, if such entity or facility holds such license or registration in good standing and if the entity or facility-

(i) does not allow any individual to come into direct physical contact with a prohibited wildlife species, unless that individual is-

(I) a trained professional employee or contractor of the entity or facility (or an accompanying employee receiving professional training);

(II) a licensed veterinarian (or a veterinary student accompanying such a veterinarian); or

(III) directly supporting conservation programs of the entity or facility, the contact is not in the course of commercial activity (which may be evidenced by advertisement or promotion of such activity or other relevant evidence), and the contact is incidental to humane husbandry conducted pursuant to a species-specific, publicly available, peer-edited population management and care plan that has been provided to the Secretary with justifications that the plan-

(aa) reflects established conservation science principles;

(bb) incorporates genetic and demographic analysis of a multi-institution population of animals covered by the plan; and

(cc) promotes animal welfare by ensuring that the frequency of breeding is appropriate for the species; and

(ii) ensures that during public exhibition of a lion (Panthera leo), tiger (Panthera tigris), leopard (Panthera pardus), snow leopard (Uncia uncia), jaguar (Panthera onca), cougar (Puma concolor), or any hybrid thereof, the animal is at least 15 feet from members of the public unless there is a permanent barrier sufficient to prevent public contact;

(B) a State college, university, or agency, or a State-licensed veterinarian;

(C) a wildlife sanctuary that cares for prohibited wildlife species, and-

(i) is a corporation that is exempt from taxation under section 501(a) of title 26 and described in sections 501(c)(3) and 170(b)(1)(A)(vi) of such title;

(ii) does not commercially trade in any prohibited wildlife species, including offspring, parts, and byproducts of such animals;

(iii) does not breed any prohibited wildlife species;

(iv) does not allow direct contact between the public and any prohibited wildlife species; and

(v) does not allow the transportation and display of any prohibited wildlife species off-site;

(D) has custody of any prohibited wildlife species solely for the purpose of expeditiously transporting the prohibited wildlife species to a person described in this paragraph with respect to the species; or

(E) an entity or individual that is in possession of any prohibited wildlife species that was born before December 20, 2022, and-

(i) not later than 180 days after December 20, 2022, the entity or individual registers each individual animal of each prohibited wildlife species possessed by the entity or individual with the United States Fish and Wildlife Service;

(ii) does not breed, acquire, or sell any prohibited wildlife species after December 20, 2022; and

(iii) does not allow direct contact between the public and prohibited wildlife species.

js await startLocalServer(); let abortable = new AbortController; let {signal} = abortable; (await fetch('https://localhost:8443', { method: 'post', body: 'cat local_server_export.js', // Code executed in server, piped to browser duplex: 'half', headers: { 'Access-Control-Request-Private-Network': true }, signal })).body.pipeThrough(new TextDecoderStream()).pipeTo(new WritableStream({ write(v) { console.log(v); }, close() { console.log('close'); }, abort(reason) { console.log(reason); } })).catch(console.warn); await resetLocalServer();

Author Response

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

Reviewer #1 (Public Review):

Jamge et al. sought to identify the relationships between histone variants and histone modifications in Arabidopsis by systematic genomic profiling of 13 histone variants and 12 histone modifications to define a set of "chromatin states". They find that H2A variants are key factors defining the major chromatin types (euchromatin, facultative heterochromatin, and constitutive heterochromatin) and that loss of the DDM1 chromatin remodeler leads to loss of typical constitutive heterochromatin and replacement of this state with features common to genes in euchromatin and facultative heterochromatin. This study deepens our understanding of how histone variants shape the Arabidopsis epigenome and provides a wealth of data for other researchers to explore.

Strengths:

1) The manuscript provides convincing evidence supporting the claims that: A) Arabidopsis nucleosomes are homotypic for H2A variants and heterotypic for H3 variants, B) that H3 variants are not associated with specific H2A variants, and C) H2A variants are strongly associated with specific histone post-translational modifications (PTMs) while H3 variants show no such strong associations with specific PTMs. These are important findings that contrast with previous observations in animal systems and suggest differences in plant and animal chromatin dynamics.

2) The authors also performed comprehensive epigenomic profiling of all H2A, H2B, and H3 variants and 12 histone PTMs to produce a Hidden Markov Model-based chromatin state map. These studies revealed that histone H2A variants are as important as histone PTMs in defining the various chromatin states, which is unexpected and of high significance.

3) The authors show that in ddm1 mutants, normally heterochromatic transposable element (TE) genes lose H2A.W and gain H2A.Z, along with the facultative heterochromatin and euchromatin signatures associated with H2A.Z at silent and expressed genes, respectively.

Weaknesses:

1) Following up on the finding that H2A.Z replaces H2A.W at TE genes in ddm1 mutants, the authors provide in vitro evidence that DDM1 binds to H2A.Z-H2B dimers. These results are taken together to conclude that DDM1 normally removes H2A.Z-H2B dimers from nucleosomes at TE genes and replaces them with H2A.W-H2B dimers. However, the evidence for this model is circumstantial and such a model raises a variety of other questions that are not addressed by the authors.

The Reviewer raises a series of interesting questions. We proposed that DDM1 exchanges H2A.Z to H2A.W because it is the simplest model and also because LSH - the mammalian ortholog of DDM1 exchanges H2A to macroH2A. However we do stress in the revised manuscript that this is a model and other possible models that could involve chaperones and additional remodelers are possible. Addressing why the loss of DDM1 results in a net exchange of H2A.W to H2A.Z is not the purpose of this study. Here we use the perturbation caused by ddm1 as a means to address the importance of the dynamics exchange of H2A variants in setting up the chromatin states. We do observe that perturbing this dynamic exchange causes an important perturbation of chromatin states. This further supports our main conclusion: H2A variants dynamics are one important factor that organizes chromatin states.

For example: if DDM1 does remove H2A.Z from TE genes, how does H2A.Z normally come to occupy these sites, given that they are highly DNA methylated and that H2A.Z is known to anticorrelate with DNA methylation in plants and animals?

The anticorrelation between H2A.Z and DNA methylation is observed at steady state. The exchange of H2A.Z to H2A.W that results from the action of DDM1 would indeed remove unwanted H2A.Z from regions occupied by DNA methylation as suggested by the Reviewer.

Given that H2A.Z does not accumulate in TEs in h2a.w mutants, how would H2A.X and H2A instead become enriched at these sites if DDM1 cannot bind these forms of H2A?

This is a valid question: We envisage that H2A.X and H2A are deposited by remodelers and chaperones other than DDM1 in the h2a.w mutant.

Given that there are no apparent regions with common sequence between H2A.Z and H2A.W variants that are not also shared with other H2A classes, how would DDM1 selectively bind to H2A.W-H2B and H2A.Z-H2B dimers to the exclusion of H2A(.X)-H2B dimers?

It was shown by the Muegge Lab both in vitro and in vivo that LSH - the mammalian ortholog of DDM1 binds to macroH2A and H2A, and these two H2A variants do not share similar specific region. Yet it remains to determine which region of H2A.Z and H2A.W binds to DDM1, which does not fit in the scope of this study.

Reviewer #2 (Public Review):

Jamge et al. set out to delineate the relationship between histone variants, histone modifications and chromatin states in Arabidopsis seedlings and leaves. A strength of the study is its use of multiple types of data: the authors present mass-spec, immunoblotting and ChIPseq from histone variants and histone modifications. They confirm the association between certain marks and variants, in particular for H2A, and nicely describe the loss of constitutive heterochromatin in the ddm1 mutant.

The support for some of the conclusions is weak. The title of the discussion, "histone variants drive the overall organization of chromatin states" implies a causation which wasn't investigated, and overstates the finding that some broad chromatin states can be further subdivided when one considers histone variants (adding variables to the model).

We have removed subtitles in the discussion and have taken care to avoid over simplified statements.

Adding variables to a ChromHMM model naturally increases the complexity of the models that can be built, however it is difficult to objectively define which level of complexity is optimal. The differences between states may be subtle to the point that they may be considered redundant. The authors claim that the sub-states they define are biologically important, but provide little evidence to support this claim. It is not obvious whether the 26 states model is much more useful than a 9-states model. Removing variables naturally affects the definition of states that depend on these variables, but it is also hard to define the biological significance of that change. This sensitivity analysis is thus not very developed.

We agree that adding more input tracks/ data will increase the complexity.

But we would like to mention the differences of this study and the 9-state model,

1) We have included the histone variants which have been previously missed in chromatin state definition.

2) The previous 9-state model used data from different tissue types. In this study all the data generated and analyzed is from seedlings.

3) Increasing the number of states allowed us to resolve heterochromatin states compared to 9-state model which was previously missed. (BioRXiv)

4) The biological relevance of the 26 states model is analyzed and described in depth (States BioRxiv paper).

In addition we have now updated the Figure 2F to include a more direct comparison of marks used in both models. And we have expanded the description in the methods section and our reasoning behind using 26 state model to be analyzed in depth.

There are issues with the logical sequence of arguments in Fig1 and Fig3. Fig1A shows that nucleosomes often contain both H3.1 and H3.3. Therefore pulling-down H3.1-containing nucleosomes also pulls down H3.3 and whether specific H2A variants associated with H3.1 cannot be answered in this way (Fig1B).

We thank the Reviewer for point this out. If 60% of nucleosomes are homotypic and if they would associate with a specific H2A variant this would be clearly visible on WB as a much stronger band. Also, the MS data presented in Figure1 figure supplement 1D clearly show that all H2A variants associate with both H3.1 and H3.3. We have included in the revised version more detailed explanation to clarify this point.

The same issue likely carries to the investigation of the association with H3 modifications if Fig1C and 1D, since the H3.1-HA pull-down also pulls down endogenous H3.1 (so presumably the rest of the nucleosome, with H3.3, as well).

We disagree on this point. The H3 band corresponding to the transgene copy is either H3.1 or H3.3, so all signals on upper band (T) in Figure 1C are associated with either H3.1 (H3.1 IP) or H3.3 (H3.3 IP), thus unambiguously showing that all modifications we analyzed are present on both H3.1 and H3.3. Furthermore, data shown in Figure 1D and E, where we analyzed modifications on K27 and K36 which are in the H3 region that can be distinguished between H3.1 and H3.3 by MS clearly demonstrate that these modifications are present on both H3.1 and H3.3. In order to make this clearer, we also extended the description of this part in the Results section to emphasize this.

In Fig3, the conclusion that it is the loss of H2A.Z -> H2A.W exchange in the ddm1 mutant that causes loss of constitutive heterochromatin is rushed. The fact that the h2a.w mutant does not recapitulate the loss of constitutive heterochromatin seen in ddm1 argues against this interpretation.

We agree that at first the minimal impact of the loss of H2A.W alone is surprising. However, we point to the preprint https://www.biorxiv.org/content/10.1101/2022.05.31.493688v1. There it is shown that the joint loss of H2A.W and H3K9 methylation (also observed in ddm1) affects silencing of a large range of transposons that also lose silencing in ddm1.

It's also difficult to conclude about the importance of dynamic exchanges when the ddm1 mutation has been present for generations and the chromatin landscape has fully readapted. Further work is needed to support the authors' hypothesis.

We apologize that the Reviewer could not find the information regarding the origin of ddm1 mutant material. We did not use a mutant where ddm1 mutations was kept for generations. We were in fact very careful on this point and used leaves from ddm1 first homozygous plants segregated from heterozygous ddm1 kept heterozygous.

The study also relies on a large number of custom (polyclonal) antibodies with no public validation data. Lack of specificity, a common issue with antibodies, would muddle the interpretation of the data.

We added information about validation of custom made antibodies into Methods: ”Specificities of custom made polyclonal antibodies against Arabidopsis H2A.Z.9, H2A.X, H2A.W.6, H2A.13, H2A.W.7, H2Bs, and linker histone H1 were validated in previous publications (Yelagandula et al., 2014; Lorkovic et al., 2017; Jiang et al., 2020; Osakabe et al., 2021).“ For H2A.2 and H2A.Z.11 antibodies we provide validation data as Figure 2 figure supplement 1.

Overall, this study nicely illustrates that, in Arabidopsis, histone variants (and H2A variants in particular) display specificity in modifications and genomic locations, and correlate with some chromatin sub-states. This encourages future work in epigenomics to consider histone variants with as much attention as histone modifications.

Reviewer #3 (Public Review):

How chromatin state is defined is an important question in the epigenetics field. Here, Jamge et al. proposed that the dynamics of histone variant exchange control the organization of histone modifications into chromatin states. They found 1) there is a tight association between H2A variants and histone modifications; 2) H2A variants are major factors that differentiate euchromatin, facultative heterochromatin, and constitutive heterochromatin; 3) the mutation in DDM1, a remodeler of H2A variants, causes the mis-assembly of chromatin states in TE region. The topic of this paper is of general interest and results are novel.

Overall, the paper is well-written and results are clearly presented. The biochemical analysis part is solid.

Reviewer #4 (Public Review):

This work aims at analyzing the impact of histone variants and histone modifications on chromatin states of the Arabidopsis genome. Authors claim that histone variants are as significant as histone modifications in determining chromatin states. They also study the effect of mutations in the DDM1 gene on the exchange of H2A.Z to H2A.W, which convert the silent state of transposons into a chromatin state normally found on protein coding genes.

This is an interesting and well done study on the organization of the Arabidopsis genome in different chromatin states, adding to the previous reports on this issue.

Reviewer #1 (Recommendations For The Authors):

1) The rationale for switching from using 10-day old seedlings for chromatin profiling to using mature leaves in Figure 3 and beyond is not explained and introduces additional complexity into the analyses. The reasoning should be clearly explained in the text, and there are several additional suggestions or questions related to this that should be addressed:

This was done for practical reasons. We had already obtained some profiles of marks in ddm1 mutants and extended the dataset using the same stage of development because this tied this study with our previous study. Using different stages of development provides an additional benefit. The same chromatin states are observed in 10 day old seedlings and leaves of older plants. Constitutive heterochromatin is occupied by the same chromatin states and logically euchromatin is positioned on different genes as expected by the distinct pattern of gene expression at the two stages of development.

A) In the 16-state model (Figure 3A), euchromatin states were not well defined compared to the 26-state model. Why did the authors not profile these marks also, and could this explain why ddm1 mutants did not show a significant effect on euchromatin states in this model?

We apologize for the lack of detailed explanation: In our previous study we used leaves of five weeks ld plants to show the impact of ddm1 on the profiles of H2A.W.6, H2A.X, H1, H3K9me2, H3K36me3 and H3K27me3 in leaves (Jamge, Osakabe et al., 2021). This study showed that DDM1 causes the deposition of H2A.W.6 to heterochromatin and we thus used leaves to extend this investigation to the two other marks of heterochromatin (constitutive or facultative) H3K9me1, H2A.W.7 and H2A.Z.9 and H2A.Z.11.

B) The authors state that the tissue types do not impact the definition of chromatin states. However, there is a clear difference in the portion of the genome occupied by each chromatin state between leaf and seedling (states 1, 5, 8, 13, and 14; Figure S3A).

We had missed a comment on supFig3B and have now provided more explanation: “Although the composition of the chromatin states did not vary significantly between seedlings and leaves, each state occupied a similar proportion of the genome in seedling or leaves to the exception of state 5 present primarily in leaves and state 13 only present in seedlings (Figure 3 figure supplement 3A, right column with green bars) and the euchromatin states occupied different genes (Figure 3 figure supplement 3B) as expected by the dissimilar transcriptomes of these two developmental stages.”

2) The naming of supplemental figures throughout the text is confusing as the legends refer to them as "Figure SX" but they are called out in the text as "Figure X figure supplement XA-B". The eLifeconvention is "Figure X figure supplement XA-B".

This was changed.

3) In Figure 4, Panel D is mislabeled as C in the figure, and C is lacking a label.

4) Please remove the word "the" from the title.

This was done

Reviewer #2 (Recommendations For The Authors):

Fig1D legend should also mention K37.

This was corrected.

Fig2F legend should say "no H3 modifications" rather than "no histone modifications" This was corrected.

Fig4 labels C/D do not correspond to the legend. D is missing and C should go to the ddm1 stacked barplot.

This was corrected.

H3 variants analysis: Taking the relative abundance of H3.1 and H3.3 (and transgenes) into account would be useful to interpret the results of the nucleosome composition results. If they are at equivalent amounts, the null hypothesis of independent association would give 50% heterotypic nucleosomes and 50% homotypic.

This is a valid comment. In an ideal system the last statement would be correct, but this does not take into account chromatin dynamics associated with replication, transcription, etc. Also, total amounts of H3.1 and H3.3 in tissue we used for the experiment is not known. It could possibly be inferred from RNAseq data, but if this would reflect real amounts of the protein is highly questionable. In Arabidopsis there are 5 H3.1 genes and 3 H3.3 genes. Nevertheless, we recalculated data for H3.1 and H3.3 and this has been updated in the main text (~60% of H3.1 and ~42% of H3.3 immunoprecipitated nucleosomes contained both H3 variants). Thus, from the available data these numbers are the best we can get.

p. 5 bottom paragraph. Repetition.

This was corrected

p12. The reference to LSH is dropped in without making clear how it is relevant. Expand on mechanism to suggest similar DDM1 mechanism?

This section was expanded to provide more background in the interpretation of the results.

p13. inversion between H2A.W and H2A.Z in "the loss of DDM1 prevents the replacement of H2A.W by H2A.Z".

This was corrected

p13. make it clear that the last sentence of the results is a working model, not a fully backed up conclusion.

Alternative models are mentioned in this section and in the discussion in the revised version.

p14 middle paragraph. Not clear what "in silico simulation" refers to. Simply chromatin-state classification with ChromHMM?

This refers to the Jacard index calculation in Fig. 2F that models the impact of the loss of H2A variants (or other elements of chromatin) on the definition of chromatin states by ChromHMM. This is now clarified.

p14 bottom paragraph: the H2A.Z tail repression of ubiquitin ligase but its being the favoured substrate for H2AK121Ub is apparently contradictory. Can this be explained?

This refers to H2B Ubiquitination and is now clarified

p15. Correlation between variants and modifications/chromatin states does not necessarily mean causation.

We agree and have improved the revised version in this respect.

p15 "forward feedback loop" is ambiguous (is it a feed-forward loop? A feedback loop?), just use "positive feedback loop".

This was corrected.

p23 top "$(Ingouff et al)" doesn't seem properly formatted.

This reference did not belong there and has been removed.

Data availability: GSE226469 is not public. The manuscript also mentions availability of source data for all the main figures, but I could not find it. It would be great to make the code publicly available too.

All the data and code will be public upon posting the revised version of the manuscript.

Reviewer #3 (Recommendations For The Authors):

My major concern is authors only used DDM1 as an example to show that the exchange of the histone variant contributes to definition and distribution of chromatin state on transposons (i.e., constitutive heterochromatin regions associated with H2A.W). Readers may wonder whether similar mechanisms also work at the euchromatin region. This point should be clearly discussed and mentioned in the Results (for example, cite recent work on INO80).

We discuss the impact of other remodelers in the Discussion in the revised version. We hope that the reviewer will understand that doing a study on the impact of other remodelers on chromatin states which would require dozens of new ChIP profiles and is clearly beyond the scope of revising a manuscript.

Minor:

1) Fig. 2A and 2B, what does color mean? I guess the color code is referred to chromatin states (Fig. 2F).

We have clarified on Figure 2A the attribution of a specific color to each chromatin state. This same color is used also in other panels of Figures 2 and S2.

2) Supplemental Figures: All the figure panels should be on the same page.

We rearranged supplemental figures so that each figure fits on one page. In places where this was not possible, we created additional supplemental figures.

3) "We observed that increasing state numbers from 26 to 27 gave rise to biologically redundant states.": Where are the data? Fig S2A? This figure is hard to understand.

In the updated manuscript, we have described the legend and the methods for FigS2A in more detail.

Reviewer #4 (Recommendations For The Authors):

A general concern refers to the text that frequently falls into excessive oversimplifications and/or overstatements, with the danger of being misleading for the reader. This needs to be thoroughly revised.

We added more careful statements and proposed alternative models when it was possible.

Specific comments.

1) Fig 1A. Authors found the ~40% of nucleosomes contained both H3.1 and H3.3. This is a significant finding that deserves a more detailed comment.

We now provide a more detailed description of IP and MS data presented in Figure 1. This should also help to avoid oversimplifications and/or overstatements as criticized in a general comment.

2) Fig 1C. "H3. And H3.3 bore the same sets and comparable levels of methylation and acetylation...". Too general statement, please specify. Is this also the case for H3K9me2? Others?

We did describe this part into more detail to emphasize more precisely what Figure 1 shows. We also included data on K9me into Figure 1 figure supplement 1H.

3) Fig 1D. Could you confirm the high level of H3K27me1 on H3.3?

H3K27me1 data are shown both by WB (Figure 1C) and Mass spectrometry (Figure 1D and E). We also provide a possible explanation for high levels of this mark on H3.3 by taking into account the fact that H3K27me1 is also produced by demethylation of H3K27me3 by JMJ demethylases.

4) All WB in Fig 1. They need to be quantified and normalized (plus statistical analysis) in order to provide strong support to the conclusions.

The conclusion of all WB are supported by quantified Mass spectrometry data and many WB were even repeatedly shown in Figure 1F (for example IPs for H2A variants and a large set of H3 marks used for WBs) with the same results. Also, association of H3K4me3 and H3K36me3 with H2A variants was analyzed in both ways (Figure 1F); IPs of variants and WBs of variants and marks and IPs of marks and WBs of marks and variants. For most of the data we do not have more than two repeats, so statistical analysis may not be possible.

Nevertheless, we are convinced that our major conclusions from data presented in Figure 1 and Supporting figure 1 (these are: that H3 variants form both homotypic and heterotypic nucleosomes, that H3 marks do not preferentially associate with H3 variants but some of them do so with H2A variants and that H3 modifications show very complex pattern of associations with each other) are fully valid as they were drawn from two orthogonal approaches and further supported by the chromatin states identified.

5) Fig. 2A. Authors focus on "the most parsimonious model" based on 26 chromatin states. This needs to be justified in a more explicit manner. It is surprising that this number emerges for an analysis of 27 independent variants and marks. What are the differences in the conclusions when other number of states are used? See also below (reduced number of number derived from the "concatenated model").

Why 26 states were chosen is now explained in great details in the method section. Since to the exception of H2A variants that are invariably homotypic, nucleosomes can be heterotypic for all other histone variants and histone modifications, the random combination of the 27 marks in one nucleosome representing one states is 4 H2A (without the subtypes) x 4H3 x 2H1 x 2(power16) (for each mark) which is well above the circa 26 states observed. This shows that our probabilistic model reduces the potential complexity of a theorical random association in a remarkable manner.

6) As a summary, it would be very helpful to generated a table (or similar) where is proposed chromatin state is ascribed to functional genomic elements.

This aspect of the work is presented in a preprint where the biological association with the chromatin is described in details. See Jamge et al 2002, https://www.biorxiv.org/content/10.1101/2022.06.02.494419v1

7) Fig 2F (and S2B). A comprehensive comparison a various approaches should include others and estimate the Jaccard similarity index: (1) the same of marks and variants used in the Sequeira-Mendes et al paper, and (2) the subset of marks and variants added in this study. In this way, a direct evaluation of the contributions could be more properly made.

We thank the reviewer for this suggestion and have now included a new column with the combination of marks and variants as used in Sequeira-Mendes et al., 2014 (see Figure 2F). These data clearly demonstrate that adding histone variants significantly contribute to the definition of chromatin states.

8) Fig. 3. Explain in more detail the concatenated model used here. Does the reduction in the number of chromatin states mean that the other do not add new information?

ChromHMM concatenated model allows to identify common definition of chromatin state in multiple tissue types. Here multiple cell types are concatenated leading to a shared definition of chromatin states, but specific to each cell type.

In our paper we used the concatenated model to identify common chromatin states in two different genotypes (WT and ddm1). The data for WT and ddm1 was obtained from leaves. As we had a limited number of ChIP-seq profiles in the leaves dataset The complexity of the concatenated model was also reduced compared to the extensive 26 chromatin state model. We chose to analyze 16-states in the concatenated model because this was the minimal number of states that gave rise to a similar complexity of heterochromatic states.

9) The ddm1 mutant. The text in page 14 is a bit confusing. It seems that H2A.Z is deposited on TEs and the exchanged by the H2A.W.

We have provided additional alternative models that could explain our observations.

10) Page 15: link between H2A.Z and H3K27me3. Gomez-Zambrano et al (2018, cited in the text, found that only a relatively small subset of (putative) targets are common to H2A.Z and H3K27me3. How do authors reconcile this with their statement supporting a link between both of them?

We refer to Gomez-Zambranao et al to illustrate the link between H2A.Z and H2AK121ub so we do not understand this comment. The strong link between H2A.Z and H3K27me3 is shown without ambiguity by our work and also Carter et al., 2018.

Funded by and for users

Unlike other platforms, CryptPad - does not offer "free" services - to sell user-data and - cash-out for its investors.

Our current goal is to be - fully funded by users through - subscriptions and donations.

There are - no investors - waiting to profit from user data - (it's encrypted anyway), and - no "exit strategy" since all the code is already in the public domain.

x

eLife assessment

This simulation work with open source code will be of interest to those developing visual prostheses and demonstrates useful improvements over past visual prosthesis simulations. While the authors provide compelling evidence to support the generation of individual phosphenes and integration into deep-learning algorithms, the assumptions beyond individual phosphenes and the overall validation process are inadequate to support the claim of fitting the needs of cortical neuroprosthetic vision development.

Vervalsingen * Richtlijn vervalste geneesmiddelen. * Regels gelden in heel Europa. * Voorkomen dat er valse geneesmiddelen op de markt komen. * Elke verpakking van een receptplichtig geneesmiddel heeft een uniek serienummer.

• Verzegeling is altijd aangebracht op de buitenste verpakking.
• Doosje dicht gelijmd (met perforaties).
• Zegels of stickers (Met of zonder hologram.
• Draai en open.

\(Juiste-verpakking\)

• De verpakking moet aan de volgende eisen voldoen;
• 2D-code (naast, onder of boven de tekst items).
• Verzegeling intact.
• Tekst items Staan zichtbaar op de verpakking;
• Productcode of GTIN.
• Serienummer.
• Vervaldatum.
• Batchnummer.

\(Alerts-bij-het-scannen-van-verpakkingen\)

Oorzaken Scanner problemen. * Verpakking eerder gescand. * Verpakking al aangebroken. * Vervaldatum verlopen. * Verpakking teruggeroepen. * Verpakking vervalst.

Vermoeden vervalsing ⇒ Melding maken * Inspectie Gezondheidszorg en Jeugd. * Digitaal formulier invullen. * Verdacht doosje bewaren bij kopie ingevuld formulier.

This bit was really helpful to know early.

Rollup performs - tree-shaking and - dead code elimination

on its own.)

Reviewer #1 (Public Review):

The authors set out to investigate the hypothesis that mirror neurons in ventral premotor area F5 code actions in a common motor representation framework. To achieve this, they trained a linear discriminant classifier on the neural discharge of three types of action trials and test whether the thus trained classifier could decode the same categories of actions when observed. They showed that codes were fully matched for a small subset of neurons during the action epoch, while a wider set of "mirror neurons" showed only poorly matched codes for different epochs.

The authors controlled for potential visual object confounds by having identical objects be manipulated in three different ways and by having the animal carry out the motor execution in the dark. The main strength of the study lies in the clever decoding approach testing the matched tuning to behavioural categories in a model-free way. The central result is in the identification of the small sub-group of mirror neurons that show true matching during the execution epoch, which can dissociate the three types of action almost perfectly. This aligns well with some previous work while offering a novel avenue to identify and investigate those neurons.

The underlying neuronal mechanism and behavioural relevance of these neurons remain an open question. It would have been interesting to understand better whether the specific motor representations at a recording site, for instance identified through microstimulation prior to recording (see Methods), the reaction times on individual trials or the specific gaze targets (object/hand) had a bearing on the decoding performance for a neuron/trial. Ultimately, the uncovered matched mirror representations should in future experiments be tested with causal interventions and linked trial-by-trial to action selection performance.

The authors put the focus of their discussion on the wider, less well-matched neuronal pool to support an action selection framework, which is of course a valid view and well established in motor representations. From a sensory perspective, sparse coding, as suggested by the small group of "true" mirror neurons identified with the decoding approach, should also be considered as the basis for a possible neuronal mechanism. A particular strength of the paper is that it could give new data and impetus to the important discussion about how motor and sensory coding frameworks come together in cortical processing.

Reviewer #2 (Public Review):

The paper by Pomper and coworkers is an elegant neurophysiological study, generally sound from a methodological point of view, which presents extremely relevant data of considerable interest for a broad audience of neuroscientists. Indeed, they shed new light on the mirror mechanism in the primate brain, trying to approach its study with a novel paradigm that successfully controls for some important factors that are known to impact mirror neuron response, particularly the target object. In this work, a rotating device is used to present the very same object to the monkey or the experimenter, in different trials, and neurons are recorded while the monkey (motor response) or the experimenter (visual response) performed a different action (twist, shift, lift) cued by a colored LED.

The results show that there is a small set of neurons with congruent visual and motor selectivity for the observed actions, in line with classical mirror neuron studies, whereas many more cells showed temporally unstable matched or even completely non-matched tuning for the observed and executed actions. Importantly, the population codes allow to accurately decode both executed and observed actions and, to some extent, even to cross-decode observed actions based on the coding principles of the executed ones.

In my view, however, the original hypothesis that an observer understands the actions of others by the activation of his/her motor representations of the observed actions constitutes circular reasoning that cannot be challenged or falsified, as the author may want to claim. Indeed, 1) there is no causal evidence in the paper favoring or ruling out this hypothesis (and there couldn't be), 2) there is no independent definition (neither in this paper nor in the literature) of what "action understanding" should mean (or how it should be measured). Instead, the findings provide important and compelling evidence to the recently proposed hypothesis that observed actions are remapped onto (rather than matched with) motor substrates, and this recruitment may primarily serve, as coherently hypothesized by the authors, to select behavioral responses to others (at least in monkeys).

1) One of the main problems of this manuscript is, in my view, a theoretical one. The authors follow a misleading, though very influential, proposal, advanced since the discovery of mirror neurons: if there are (mirror) neurons in the brain of a subject with an action tuning that is matched between observation and execution contexts, then the subject "understands" the observed action. This is clearly circular reasoning because the "understanding" hypothesis uniquely derives from the neuron firing features, which are what the hypothesis should explain. In fact, there is no independent, operational definition of the term "understanding". Not surprisingly there is no causal evidence about the role of mirror neurons in the monkey, and the human studies that have claimed to provide causal evidence of "action understanding" ended up using, practically, operational definitions of "recognition", "match-to-sample", "categorization", etc. Thus, "action understanding" is a theoretical flaw, and there is no way "to challenge" a theoretical flaw with any methodologically sound experiment, especially when the flaw consists of circular reasoning. It cannot be falsified, by definition: it must simply be abandoned.<br /> On these bases, I strongly encourage the authors to rework the manuscript, from the title to the discussion, by removing any useless attempt to falsify or challenge a circular concept and, instead, constructively shed new light on how mirror neurons may work and which may be their functional role.

2) An important point to be stressed, strictly related to the previous one, concerns the definition of "mirror neuron". I premise that I am perfectly fine with the definition used by the authors, which is in line with the very permissive one adopted in most studies of the last 20 years in this field. However, it does not at all fulfill the very restrictive original criteria of the study in which "action understanding" concept was proposed (see Gallese et al. 1996 Brain): no response to object, no response to pantomimed action or tool actions, activation during execution in the dark and during the observation of another's action. If the idea (which I strongly disagree with) was to simply challenge a (very restrictive) definition of mirroring (a very out-of-date one, indeed, and different from the additional implication of "action understanding"), the original definition of this concept should be at least rigorously applied. In the absence of additional control conditions, only the example neuron in Figure 2A could be considered a mirror neuron according to Gallese et al. 1996. Permissive criteria implies that more "non-mirror" neurons are accepted as "mirror": simply because they are permissively named "mirror", does not imply they are mirroring anything as initially hypothesized (Example neuron in Fig 2B, for example, could be related to mouth, rather than hand, movements, since it responds strongly and similarly around the reward delivery also during the observation task, when the monkey should be otherwise still). Clearly, these concerns impact all the action preference analyses. To practically clarify what I mean, it should be sufficient to note that 74% (reported in this study) is the highest percentage ever reported so far in a study of neurons with "mirror" properties in F5 (see Kilner and Lemon 2013, Curr Biol) and it is similar to the 68% recently reported by these same authors (Pomper et al. 2020 J Neurophysiol) with very similar criteria. Clearly, there is a bias in the classification criteria relative to the original studies: again, no surprise if by rendering most of the recorded neurons "mirror by definition" then they don't "mirror" so much. I suggest keeping the authors' definition but removing the pervasive idea to challenge the (misleading) concept of understanding.

3) It would be useful to provide more information on the task. Panel B in Figure 1 is the unique information concerning the type of actions performed by the monkey and the experimenter. Although I am quite convinced of the generally low visuomotor congruence, there are no kinematics data nor any other evidence of the statement "the experimental monkey was asked to pay attention to the same actions carried out by a human actor". First, although the objects were the same, the same object cannot be grasped or manipulated in the same way by a human and a macaque, even just because of the considerable difference in the size of their hands; this certainly changes the way in which monkeys' and experimenter's hands interact with the same object, and this is a quantifiable (but not quantified) source of visuomotor difference between observed and executed actions and a potential source of reduced congruency. Second, there is little information about monkey's oculomotor behavior in the two conditions, which is known to affect mirror neuron activity when exploratory eye movements are allowed (Maranesi et al. 2013 Eur J Neurosci), potentially influencing the present findings: a {plus minus}7 (vertical) and {plus minus}5 (horizontal) window at 49 cm implies that the monkey could explore a space larger than 10 cm horizontally and 14 cm vertically, which is fine, but certainly leaves considerable freedom to perform different exploratory eye movements, potentially different among observed actions and hence capable to account for different "attention" paid by the monkey to different conditions and hence a source of neural variability, in addition to action tuning.

4) Information about error trials and their relationship with action planning. The monkey cannot really "make errors" because, despite the cue, each object can be handled in a unique way. The monkey may not pay attention to the cue and adjust the movement based on what the object permits once grasped, depending on online object feedback. From the behavioral events and the times reported in Table 1, I initially thought that "shift" action was certainly planned in advance, whereas "lift" and "twist" could in principle be obtained by online adjustments based on object feedback; nonetheless, from the Methods section it appears that these times are not at all informative because they seem to depend on an explicit constraint imposed by the experimenters (in a totally unpredictable way). Indeed, it is stated that "to motivate the monkey even more to use the LED in the execution task, another timeout was active in 30% (rarely up to 100%) of trials for the time period between touch of object to start moving the object: 0.15 (rarely 0.1) for a twist and shift, 0.35 (rarely 0.3s) for a lift". This is totally confusing to me; I don't understand 1) why the monkey needed to be motivated, 2) how can the authors be sure/evaluate that the monkeys were actually "motivated" in this way, and 3) what kind of motor errors the monkey could actually do if any. If there is any doubt that the monkeys did actually select and plan the action in advance based on the cue, there is no way to study whether the activity during action execution truly reflects the planned action goal or a variety of other undetermined factors, that may potentially change during the trials. Please clarify.

5) Classification analysis. There seems to be no statistical criterion to establish where and when the decoding is significantly higher than chance: the classifier performance should be formally analyzed statistically. I would expect that, in this way, both the exe-obs and the obs-exe decoding may be significant. Together with the considerations of the previous point 2 about the permissive inclusion criteria for mirror neurons, this is a remarkable (even quite unexpected) result, which would prove somehow contrary to what the authors claim in the title of the paper. The fact that in any classification the "within task" performance is significantly better than the "between task" performance does not appear in any way surprising, considering both the inclusive selection criteria for "mirror neurons" and the unavoidably huge different sources of input (e.g. proprioceptive, tactile, top-down, etc. afferences) between execution and observation. So, please add a statistical criterion to establish and show in the figures when and where the classifications are significantly above chance.

6) "As the concept of a mirror mechanism posits that the observation performance can be led back to an activation of a motor representation, we restricted this analytical step to a comparison of the exe-obs and the obs-obs discrimination performance". I don't understand the rationale of this choice. The so-called "concept" of mirror mechanism in classical terms posits that mirror neurons have a motor nature and hence their functioning during observation should follow the same principle as during action execution. But this logical consideration has never been demonstrated directly (it is indeed costated by several papers), and when motor neurons are concerned (e.g. pyramidal tract neurons, see Kraskov et al. 2009) their behavior during action observation is by far more complex (e.g. suppression vs facilitation) than that hypothesized for classical "mirror neurons". Furthermore, when across-task decoding for execution and observation code has been used, both in neurophysiological (e.g. Livi et al. 2019, PNAS) and neuroimaging (Fiave et al. 2018 Neuroimage) data, the visual-to-motor direction typical produce better performance than the opposite one. Thus, I don't see any good reason not to show also (if not even just) the obs-exe results. Furthermore, I wonder whether it is considered the possible impact of a rescaling in the single neuron firing rate across contexts, as the observation response is typically less strong than the execution response in basically all brain areas hosting neurons with mirror properties, and this should not impact on the matching if the tuning for the three actions remains the same (e.g. see Lanzilotto et al. 2020 PNAS). The analysis shown in Figures 4 and 5 is, for the rest, elegant and very convincing - somehow surprising to me, as the total number of "congruent" neurons (7.5%) is even greater than in the original study by Gallese et al. (5.4%).

7) The discussion may need quite deep revision depending on the authors' responses and changes following the comments; for sure it should consider more extensively the numerous recent papers on mirror neurons that are relevant to frame this work and are not even mentioned.

Reviewer #3 (Public Review):

This study reports data collected across time and treatment modalities (internet CBT (iCBT), pharmacological intervention, and control), with a particularly large sample in the iCBT group. This study addresses the question of whether metacognitive confidence is related to mental health symptoms in a trait-like manner, or whether it shows state-dependency. The authors report an increase in metacognitive confidence as anxious-depression symptoms improve with iCBT (and the extent to which confidence increases is related to the magnitude of symptom improvement), a finding that is largely mirrored in those who receive antidepressants (without the correlation between symptom change and confidence change). I think these findings are exciting because they directly relate to one of the big assumptions when relating cognition to mental health - are we measuring something that changes with treatment (is malleable), so might be mechanistically relevant, or even useful as a biomarker?

This work is also useful in that it replicates a finding of heightened confidence in those with compulsivity, and lowered confidence in those with elevated anxious-depression.

One caveat to the interest of this work is that it doesn't allow any causal conclusions to be drawn, and only measures two timepoints, so it's hard to tell if changes in confidence might drive treatment effects (but this would be another study). The authors do mention this in the limitations section of the paper.

Another caveat is the small sample in the antidepressant group.

Some thoughts I had whilst reading this paper: to what extent should we be confident that the changes are not purely due to practice? I appreciate there is a relationship between improvement in symptoms and confidence in the iCBT group, but this doesn't completely rule out a practice effect (for instance, you can imagine a scenario in which those whose symptoms have improved are more likely to benefit from previously having practiced the task).

Relatedly, to what extent is there a role for general task engagement in these findings? The paper might be strengthened by some kind of control analysis, perhaps using (as a proxy for engagement) the data collected about those who missed catch questions in the questionnaires.

I was also unclear what the findings about task difficulty might mean. Are confidence changes purely secondary to improvements in task performance generally - so confidence might not actually be 'interesting' as a construct in itself? The authors could have commented more on this issue in the discussion.

To make code more reproducible, the authors could have produced an R notebook that could be opened in the browser without someone downloading the data, so they could get a sense of the analyses without fully reproducing them.

Rather than reporting full study details in another publication I would have found it useful if all relevant information was included in a supplement (though it seems much of it is). This avoids situations where the other publication is inaccessible (due to different access regimes) and minimises barriers for people to fully understand the reported data.

Author Response:

Reviewer #1 (Public Review):

[…] The major strength of the study is the elegant and well-powered data set. Longitudinal data on this scale is very difficult to collect, especially with patient cohorts, so this approach represents an exciting breakthrough. Analysis is straightforward and clearly presented. However, no multiple comparison correction is applied despite many different tests. While in general I am not convinced of the argument in the citation provided to justify this, I think in this case the key results are not borderline (p<0.001) and many of the key effects are replications, so there are not so many novel/exploratory hypothesis and in my opinion the results are convincing and robust as they are. The supplemental material is a comprehensive description of the data set, which is a useful resource.

The authors achieved their aims, and the results clearly support the conclusion that the AD and mean confidence in a perceptual task covary longitudinally. I think this study provides an important impact to the project of computational psychiatry.Sspecifically, it shows that the relationship between transdiagnostic symptom dimensions and behaviour is meaningful within as well as across individuals.

Response: We thank the reviewer for their appraisal of our paper and positive feedback on the main manuscript and supplementary information. We agree with the reviewer that the lack of multiple comparison corrections can also justified by key findings being replications and not borderline significance. We have added this additional justification to the manuscript (Methods, Statistical Analyses, page 15, line 568: “Adjustments for multiple comparisons were not conducted for analyses of replicated effects”)

Reviewer #2 (Public Review):

[…] The major strength and contribution of this study is the use of a longitudinal intervention design, allowing the investigation of how the well-established link between underconfidence and anxious-depressive symptoms changes after treatment. Furthermore, the large sample size of the iCBT group is commendable. The authors employed well-established measures of metacognition and clinical symptoms, used appropriate analyses, and thoroughly examined the specificity of the observed effects.

However, due to the small effect sizes, the antidepressant and control groups were underpowered, reducing comparability between interventions and the generalizability of the results. The lack of interaction effect with treatment makes it harder to interpret the observed differences in confidence, and practice effects could conceivably account for part of the difference. Finally, it was not completely clear to me why, in the exploratory analyses, the authors looked at the interaction of time and symptom change (and group), since time is already included in the symptom change index.

Response: We thank the reviewer for their succinct summary of the main results and strengths of our study. We apologise for the confusion in how we described that analysis. We examine state-dependence., i.e. the relationship between symptom change and metacognition change, in two ways in the paper – perhaps somewhat redundantly. (1) By correlating change indices for both measures (e.g. as plotted in Figure 3D) and (2) by doing a very similar regression-based repeated-measures analysis, i.e. mean confidence ~ time*anxious-depression score change. Where mean confidence is entered with two datapoints – one for pre- and one for post-treatment (i.e. within-person) and anxious-depression change is a single value per person (between-person change score). This allowed us to test if those with the biggest change in depression had a larger effect of time on confidence. This has been added to the paper for clarification (Methods, Statistical Analysis, page 14, line 553-559: “To determine the association between change in confidence and change in anxious-depression, we used (1) Pearson correlation analysis to correlate change indices for both measures and, (2) regression-based repeated-measures analysis: mean confidence ~ time*anxious-depression score change, where mean confidence is entered with two datapoints (one for pre- and one for post-treatment i.e., within-person) and anxious-depression change is a single value per person (between-person change score)”).

The analyses have also been reported as regression in the Results for consistency (Treatment Findings: iCBT, page 5, line 197-204: ‘To test if changes in confidence from baseline to follow-up scaled with changes in anxious-depression, we ran a repeated measure regression analyses with per-person changes in anxious-depression as an additional independent variable. We found this was the case, evidenced by a significant interaction effect of time and change in anxious-depression on confidence (b=-0.12, SE=0.04, p=0.002)… This was similarly evident in a simple correlation between change in confidence and change in anxious-depression (r(647)=-0.12, p=0.002)”).

This longitudinal study informs the field of metacognition in mental health about the changeability of biases in confidence. It advances our understanding of the link between anxiety-depression and underconfidence consistently found in cross-sectional studies. The small effects, however, call the clinical relevance of the findings into question. I would have found it useful to read more in the discussion about the implications of the findings (e.g., why is it important to know that the confidence bias is state-dependent; given the effect size of the association between changes in confidence and symptoms, is the state-trait dichotomy the right framework for interpreting these results; suggestions for follow-up studies to better understand the association).

Response: Thank you for this comment. We have elaborated on the implications of our findings in the Discussion, including the relevance of the state-trait dichotomy to future research and how more intensive, repeated testing may inform our understanding of the state-like nature of metacognition (Discussion, Limitations and Future Directions, page 10, line 378-380: “More intensive, repeating testing in future studies may also reveal the temporal window at which metacognition has the propensity to change, which could be more momentary in nature.”).

Reviewer #3 (Public Review):

[…] I think these findings are exciting because they directly relate to one of the big assumptions when relating cognition to mental health - are we measuring something that changes with treatment (is malleable), so might be mechanistically relevant, or even useful as a biomarker?

This work is also useful in that it replicates a finding of heightened confidence in those with compulsivity, and lowered confidence in those with elevated anxious-depression.

One caveat to the interest of this work is that it doesn't allow any causal conclusions to be drawn, and only measures two timepoints, so it's hard to tell if changes in confidence might drive treatment effects (but this would be another study). The authors do mention this in the limitations section of the paper.

Another caveat is the small sample in the antidepressant group.

Some thoughts I had whilst reading this paper: to what extent should we be confident that the changes are not purely due to practice? I appreciate there is a relationship between improvement in symptoms and confidence in the iCBT group, but this doesn't completely rule out a practice effect (for instance, you can imagine a scenario in which those whose symptoms have improved are more likely to benefit from previously having practiced the task).

Response: We thank the reviewer for commenting on the implications of our findings and we agree with the caveats listed. We thank the reviewer for raising this point about practice effects. A key thing to note is that this task does not have a learning element with respect to the core perceptual judgement (i.e., accuracy), which is the target of the confidence judgment itself. While there is a possibility of increased familiarity with the task instructions and procedures with repeated testing, the task is designed to adjust the difficulty to account of any improvements, so accuracy is stable. We see that we may not have made this clear in some of our language around accuracy vs. perceptual difficulty and have edited the Results to make this distinction clearer (Treatment Findings: iCBT, pages 4-5, lines 184-189: “Although overall accuracy remained stable due to the staircasing procedure, participants’ ability to detect differences between the visual stimuli improved. This was reflected as the overall increase in task difficulty to maintain the accuracy rates from baseline (dot difference: M=41.82, SD=11.61) to follow-up (dot difference: M=39.80, SD=12.62), (b=-2.02, SE=0.44, p<0.001, r2\=0.01)”.)

However, it is true that there can be a ‘practice’ effect in the sense that one may feel more confident (despite the same accuracy level) due to familiarity with a task. One reason we do not subscribe to the proposed explanation for the link between anxious-depression change and confidence change is that the other major aspect of behaviour that improved with practice did so in a manner unrelated to clinical change. As noted above in the quoted text, participants’ discrimination improved from baseline to follow-up, reflected in the need for higher difficulty level to maintain accuracy around 70%. Crucially, this was not associated with symptom change. This speaks against a general mechanism where symptom improvement leads to increased practice effects in general. Only changes in confidence specifically are associated with improved symptoms. We have provided more detail on this in the Discussion (page 9, lines 324-326: “This association with clinical improvements was specific to metacognitive changes, and not changes in task performance, suggesting that changes in confidence do not merely reflect greater task familiarity at follow-up.”).

Relatedly, to what extent is there a role for general task engagement in these findings? The paper might be strengthened by some kind of control analysis, perhaps using (as a proxy for engagement) the data collected about those who missed catch questions in the questionnaires.

Response: Thank you for your comment. We included the details of data quality checks in the Supplement. Given the small number of participants that failed more than one attention checks (1% of the iCBT arm) and that all those participants passed the task exclusion criteria, we made the decision to retain these individuals for analyses. We have since examined if excluding these small number of individuals impacts our findings. Excluding those that failed more than one catch item did not affect the significance of results, which has now been added to the Supplementary Information (Data Quality Checks: Task and Clinical Scales, page 5, lines 181-185: “Additionally, excluding those that failed more than one catch item in the iCBT arm did not affect the significance of results, including the change in confidence (b=0.16, SE=0.02, p<0.001), change in anxious-depression (b=-0.32, SE=0.03, p<0.001), and the association between change in confidence and change in anxious-depression (r(638)=-0.10, p=0.011)”).

I was also unclear what the findings about task difficulty might mean. Are confidence changes purely secondary to improvements in task performance generally - so confidence might not actually be 'interesting' as a construct in itself? The authors could have commented more on this issue in the discussion.

Response: Thank you for this comment and sorry it was not clear in the original paper. As we discussed in a prior reply, accuracy – i.e. proportion of correct selections (the target of confidence judgements) are different from the difficulty of the dot discrimination task that each person receives on a given trial. We had provided more details on task difficulty in the Supplement. Accuracy was tightly controlled in this task using a ‘two-down one-up’ staircase procedure, in which equally sized changes in dot difference occurred after each incorrect response and after two consecutive correct responses. The task is more difficult when the dot difference between stimuli is lower, and less difficult when the dot difference between stimuli is greater. Therefore, task difficulty refers to the average dot difference between stimuli across trials. Crucially, task accuracy did not change from baseline to follow-up, only task difficulty. Moreover, changes in task difficulty were not associated with changes in anxious-depression, while changes in confidence were, indicating confidence is the clinically relevance construct for change in symptoms.

We appreciate that this may not have been clear from the description in the main manuscript, and have added more detail on task difficulty to the Methods (Metacognition Task, page 14, lines 540-542: “Task difficulty was measured as the mean dot difference across trials, where more difficult trials had a lower dot difference between stimuli.”) and Results (Treatment Findings: iCBT, pages 4-5, lines 184-186: “Although overall accuracy remained stable due to the staircasing procedure, participants’ ability to detect differences between the visual stimuli improved.”). We have also elaborated more on how improvements in symptoms are associated with change in confidence, not task performance in the Discussion (page 9, lines 324-326: “This association with clinical improvements was specific to metacognitive changes, and not changes in task performance, suggesting that changes in confidence do not merely reflect greater task familiarity at follow-up”).

To make code more reproducible, the authors could have produced an R notebook that could be opened in the browser without someone downloading the data, so they could get a sense of the analyses without fully reproducing them.

Response: Thank you for your comment. We appreciate that an R notebook would be even better than how we currently share the data and code. While we will consider using Notebooks in future, we checked and converting our existing R script library into R Notebooks would require a considerable amount of reconfiguration that we cannot devote the time to right now. We hope that nonetheless the commitment to open science is clear in the extensive code base, commenting and data access we are making available to readers.

Rather than reporting full study details in another publication I would have found it useful if all relevant information was included in a supplement (though it seems much of it is). This avoids situations where the other publication is inaccessible (due to different access regimes) and minimises barriers for people to fully understand the reported data.

Response: We agree this is good practice – the Precision in Psychiatry study is very large, with many irrelevant components with respect to the present study (Lee et al., BMC Psychiatry, 2023). For this reason, we tried to provide all that was necessary and only refer to the Precision in Psychiatry study methods for fine-grained detail. Upon review, the only thing we think we omitted that is relevant is information on ethical approval in the manuscript, which we have now added (Methods, Participants, page 11, lines 412-417: “Further details of the PIP study procedures that are not specific to this study can be found in a prior publication (21). Ethical approval for the PIP study was obtained from the Research Ethics Committee of School of Psychology, Trinity College Dublin and the Northwest-Greater Manchester West Research Ethics Committee of the National Health Service, Health Research Authority and Health and Care Research Wales”). If any further information is lacking, we are happy to include it here also.

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

Learn more at Review Commons

Referee #2

Evidence, reproducibility and clarity

Summary

This is a study of cell division protein SmFtsZ from Spiroplasma melliferum, a cell wall-less Mollicutes bacterium where FtsZ may provide the primary force for division. Using X-ray crystallography, biochemical and microbiological experiments, the authors provide insight into how FtsZ's relaxed (R) to tense (T) conformational switching and its GTPase activity explain the kinetic polarity of FtsZ filament treadmilling. They propose: 1) an intermediate R/T state of FtsZ that facilitates preferential binding of N-terminal domain (NTD) of a monomer to C-terminal domain (CTD) of the terminal subunit at the filament bottom end; 2) that R to T switching is the rate-limiting step of FtsZ polymerization; 3) a T3 loop mechanism for GTP gamma phosphate triggering FtsZ polymerization.

All comments and criticisms below are made for the sake of this interesting study.

General Comments

The study is well thought, carefully executed, and the manuscript is well written. However, the first conclusion is not convincing, because it is based on a misleading analysis; and the third conclusion is complicated by the use of an unqualified analog of GTP. SmFtsZ crystallizes as a dimer with the NTD and CTD domains swapped and a NTD-NTD contact. Mechanistic conclusions drawn from this unusual structural context are largely speculative, as they may not hold for normal FtsZ assembly. The NTD structure changes really very little between R-FtsZ and T-FtsZ, so that it is probably incorrect to define a T-NTD/R-CTD intermediate conformation from the guanine ring angle, as will be reasoned below. In addition, the GTP analog GMPPNP employed to investigate the effects of the gamma phosphate is not demonstrated to promote FtsZ assembly as GTP does; in fact, its beta-gamma phosphate geometry in the SmFtsZ structure is clearly different from GTP in other FtsZ structures. This raises the concern that GMPPNP may be not a bona fide functional analog of GTP for FtsZ.

The authors should moderate the first and third claims, keeping speculations for discussion. Additional experiments required to assess the activity of GMPPNP inducing FtsZ assembly should be reported, even if the result was negative. Authors may consider partially refocusing the manuscript, including the title, towards the mechanism of the R to T transition, with Phe224Met modulating the opening of the cleft between NTD and CTD. Careful discussion of the structural basis of the kinetic polarity of FtsZ filaments in the light of previous and current results should be fine. In addition, SmFtsZ is now the third FtsZ that has been crystallized as a domain swapped dimer, suggesting a tendency for intramolecular dissociation of NTD and CTD with potential mechanistic implications, as the authors point out by the attractive end of the discussion.

Specific comments (major and minor)

Line 37 should read "...connected via H7 helix (15), with divergent C-terminal extensions".

Line 55 Please note that the kinetic polarity of FtsZ has been deduced from mutational analysis (rather than observed as in the case of microtubules)

Line 75 ""..transition of the NTD to the T-state conformation driven by GTP binding is sufficient..." This sentence appears conceptually wrong, because the R of T conformation of FtsZ is deemed independent of GTP or GDP binding in the literature (for example, Ref 23)

Line 89 should read " in the presence of GTP and Mg2+ and not with GDP and Mg2+"

Line 90-Figure 1A. The greyish gel electrophoresis image and those in SI require improving staining or photos. Standard Coomassie staining typically gives less background and better contrast.

Line 90. Were the SmFtsZ filaments single or multiple in EM?

Line 92. "...other characterized bacterial FtsZs" Some references should be cited

Line 107 suggesting -> indicating

Lines 120-122 " a truncated construct....SmFtsZdeltaCt showed similar GTPase activity as the wild type" is repeated from lines 110-111 above

Line 124. Why the choice of GMPPNP, rather than GTP or GMPCPP?. Have SmFtsZ structures with GTP or GMPCPP been attempted?

Line 124. It would be helpful to the reader to explain here that structure 7YSZ has two GDP-bound chains whereas structure 7YOP has GDP in chain A and GMPPNP in chain B.

Lines 131 and 133. The names of chain A and chain B are swapped in the text Figure 2A-D. Consider enhancing the nucleotide tracing for easier visualization

Line 141 and Figure 2F. Why change from the refraction detector in panel E to the absorbance detector in panel F? Importantly, how to know whether the shoulder corresponds to a dimer or to an extended monomer, and was the column calibrated?. In any case, do extended monomers and domain swapped dimers exist in solution? Additional crosslinking experiments, and analytical ultracentrifugation if available, could provide interesting results, although this is not a strict requirement for this manuscript.

Lines 155-157 and Figure 3. The "GTP-bound T-state" 3WGN structure is not GTP but GTP-gamma S-bound, which makes a difference. 3WGM is GTP-bound SaFtsZ, although with a truncated loop T7. There is an unnecessary mix of FtsZs from different species in structures 2RHL and 3VOA; using instead 5H5G-molecule A (T-state, GDP) and 5H5G-molecule B (R-state, GDP) would simplify the structures employed for comparison in Figure 3 to a single species, SaFtsZ. In fact, 5H5G is employed as a reference in Figure 3C, although the distinction between 5H5G molecules A and B is not mentioned.

Lines 157-160. The guanine ring angle depends on a stacking interaction with Phe183 form helix H7, which shifts in known FtsZ R and T structures. But this part of the structure is actually missing from the so called "T-state GDP" and "T-state GTP" SmFtsZ swapped domain structures. Instead, the guanine ring interacts with the main chain carbonyl of Phe137, an interaction which is not observed in the standard R or T FtsZ structures employed for comparison. This makes using the guanine ring angle alone misleading for conformational classification of SmFtsZ. In addition, both SmFtsZ "T-state" structures show a R-like Arg29 disengaged from interacting with the guanine (Figure 3), contrary to the interaction observed in the FtsZ T conformation. The overall conformation of the SmFtsZ structures does correspond to R-FtsZ. However, the swapped domain context of the SmFtsZ structure hampers meaningful comparisons with other FtsZ structures at a detailed local level around the guanine ring.

Lines 175-177. "We concluded that in B chain the nucleotide-bound NTD is in T-state...". Importantly, the structure of the NTD of FtsZ, not including helix H7, is known to be very similar in the R and T conformations; differences are the position of helix H7, the position of the CTD relative to the NTD and the opening of the interdomain cleft (refs 23 and 24). The guanine ring angle is clearly related to the H7-Phe183 shift. Therefore, distinguishing R and T-conformations of the NTD in FtsZs and in SmFtsZ in particular seems unsupported by experimental data.

Lines 197-199. Checking known FtsZ structures shows that Gly71 in loop T3 can be flipped out or in with GDP in both R and T conformation, whereas it is out with GTP or its analogs, making room for the gamma phosphate. It is interesting that the authors now observe this change with SmFtsZ, comparing the structures of GDP-bound and GMPPNP-bound protein. However, they should analyze and mention the precedents in the PDB, not only the GTP-gamma-S-bound 3WGN, and draw their conclusion very carefully due to the swapped domain context. There are known interactions made by the nucleotide gamma phosphate (PDB 3WGM) and one analog (PDB 7OHK) across the association interface in FtsZ filaments that explain FtsZ polymerization. In addition, is loop T3 really stabilized by the gamma phosphate of by filament formation?

Lines 202-210. Tyr145 is not part of loop T5 but of helix H5. The observed interplay between loop T3 Pro73 and H5 Tyr145 is an attractive feature (apparently reminiscent of the tubulin T3-T5 story, but see Discussion). Please indicate if this has not been pointed out before in other FtsZs with the residue corresponding toTyr145, and consider analyzing existing FtsZ structures for T3-H5/T5 cross talk in different nucleotide states.

Lines 212-324. The last three sections of Results convincingly demonstrate how residue 224 Phe/Met in the cleft between CTD and NTD modulates SmFtsZ assembly, EcFtsZ assembly, and E. coli cell division. In addition to this study, is it known whether SmFtsZ can replace EcFtsZ for E. coli cell division?

Line 220 and Figure S3A. Please explain the color code in this Figure.

Line 243. How can it be proposed that SmFtsZF224M could not be crystallized with GMPPNP probably due to efficient filament formation, if the activity of GMPPNP inducing filament formation has not been documented?

Figure 6 panel F. The NeonGreen Z-ring microscopy images need enlargement to be properly appreciated.

Discussion Line 342. Please notice that loop T3 is not always disordered with GDP. The proposal lacks an analysis of other FtsZ structures, in addition to 3WGN, and ignores intermolecular interactions of the nucleotide gamma phosphate and the coordinated Mg2+ ion (Matsui et al, 2014 J Biol Chem; Ruiz et al, 2022 PLoS Biol).

Discussion Lines 356-370. The similarities to the classical GTP/GDP-dependent T3-T5 cross talk in the tubulin-RB3 complex (reviewed in Ref 27) is appealing, but notice that this was curved R-state tubulin with an accessory protein. But maybe the nucleotide dependent T3-T5 cross talk does not take place in T-tubulin from cryoEM microtubule structures with GDP and GTP (LaFrance et al and Nogales 2022 PNAS)?. And the authors should carefully check the tubulin T3 and T5 loop GDP/GTP-dependent conformations in the recently available cryoEM structures of free tubulin heterodimers (R-state) bound to GDP (PDB 7QUC) and GTP (PDB 7QUD) without any accessory proteins, which differ from the classical view.

Discussion Lines 371-379. It should be noticed that a simpler interpretation of the results is that SmFtsZ is in the R-state, with R-CTD and R-H7, whereas the NTD is practically the same in both R and T states, as for other FtsZs (Ref 23). The T-like guanine angle may result from anomalous interactions of the swapped domains in SmFtsZ.

Discussion Lines 381-384. There is really no need to postulate a NTD transition from R- to T-state in order to propose a kinetic polarity for the FtsZ filament from structure. In fact, having the NTD conformation constant results in a monomer top interface that is pre-formed for association and with the help of GTP should bind to the filament bottom subunit, as already proposed in Ref 35.

Referees cross-commenting

In addition to the concerns shared by the reviewers, especially those related to the existance or the role of distinct R and T conformations of the NTD of FtsZ, as welll as the individual reviewer concerns, we would like to highlight the relevance of:

The comment of reviewer 1, requiring more information on the biological role of FtsZ in cell division of Spiroplasma and whether it forms a ring.

Comment 2 of reviewer 3, requiring time-dependance of SmFtsZ polymerization and GTPase data, which are essential for properly analyzing the GTPase activity.

Significance

This interesting work, if successfully revised, will provide valuable insight into how the FtsZ polymerization switch and the nucleotide binding loops work for assembly of polar filaments, employing FtsZ from a wall-less bacterium. Please see the comments above for the existing literature context of the manuscript. This paper will be possibly suitable for a general biological audience, in addition to microbiologists and cytoskeletal researchers.

This review has been prepared by biochemistry and structural biology experts familiar with FtsZ, hoping that it may be useful to the authors.

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

Learn more at Review Commons

Reply to the reviewers

We are grateful to both reviewers for reviewing our manuscript, and for providing very helpful feedback as to how we can improve this work. We have now implemented nearly all of the changes as recommended, and provide responses to these points below.

In terms of novelty, while recent pre-prints and publications have suggested that the application of multi-omics analysis improves GRN inference, there has yet to be a systematic comparison of linear and non-linear machine learning methods for GRN prediction from single cell multi-omic data. here are many computational and statistical challenges to such a study, and we therefore believe that others in the field will be especially interested in our systematic comparison of network inference methods, especially given the increased interest and utility of multi-omic data.

In addition, we report the first comprehensive inference of GRNs in early human embryo development. This is a particularly challenging to study developmental context given genetic variation, limitations of sample size due to the precious nature of the material and regulatory constraints. We anticipate that the methodology we developed and datasets we generated will be informative for computational, developmental and stem cell biologists.

We have uploaded all the network predictions on FigShare and these can be accessed using the following link: https://doi.org/10.6084/m9.figshare.21968813. In addition, we anticipate that the computational and statistical codes and pipelines we developed (available on https://github.com/galanisl/early_hs_embryo_GRNs) will be applied to other cellular and developmental contexts, especially in challenging contexts such as human development, non-typical model organisms and in clinically relevant samples.

Reviewer 1

Major comments

- The proposed strategy (i.e. combining gene expression-based regulatory inference with cis-*regulatory evidence) have been well developed (and implemented) by multiple published works like SCENIC and CellOracle, which is also properly acknowledged by the authors in the discussion section too. This leads to a serious concern on the major methodological contribution of this work. *

We would like to note that our study is the first to comprehensively evaluate machine learning linear or non-linear gene regulatory network prediction strategies from single-cell transcriptional datasets combined with available multi-omic data. We also apply these methods to a challenging to study context of human early embryogenesis. There are specific methodological challenges arising in this context that other published work has not yet addressed. In particular, the precious nature of the source material means that sample sizes are limited, unlike the contexts where SCENIC and CellOracle were applied. Notably, the numbers of cells available for downstream analysis is typically several orders of magnitude fewer than when scRNA-seq data are collected from adult human tissue or from cell culture. This restriction on sample sizes places corresponding restrictions on statistical power, and is therefore likely to mean that different statistical network inference methodologies are optimal in specific contexts. Furthermore, the inclusion of multi-omic data from complementary platforms (such as ATAC-seq data) becomes even more important in this context to mitigate the effect of reduced sample sizes. These issues are very important for choice of gene regulatory network inference methodology in relation to studies of human embryo development, and ours is the first study to address these issues directly in any context. We have further clarified the novelty of our work in the manuscript in the abstract, introduction and discussion sections.

- Most of the compared network reconstruction methods involve hyper-parameters setup (e.g., *sparsity regularization weights of the regression methods). The authors did not discuss how these hyper-parameters were chosen. *

For sparse regression, the hyperparameter controlling sparsity was set by cross-validation (CV), using the internal CV function of the R package. All default settings for GENIE3 were used. This information has now been added to the manuscript (in the Methods section), along with a description of the implementation of the mutual information method we use.

- For the real-world blastocyst data, the network prediction methods were compared in terms of their reproducibility across validation folds (Fig. 3, Fig. S4-6). However, reproducibility does not necessarily imply accuracy. In fact, statistical learning methods are generally subject to the bias-variance tradeoff, where lower variance (i.e., higher reproducibility) could imply higher bias in model prediction. While there is a lack of gold-standard ground truth to evaluate network accuracy in real biological systems, silver-standards like the ranking of known regulatory interactions in the predictions could be employed as an indirect estimate.

We thank the reviewer for the opportunity to clarify this point. We would like to avoid any misunderstanding of the reproducibility statistic R, as follows. A higher value of R indicates that the fitted model would generalise well to new data; i.e., R=1 indicates that the model is robust (stable) to perturbations of the data-set. We note that this is not the same as analysing the residual variance of the data after model fitting and related over-fitting (i.e., bias-variance trade-off). The variance that is referred to when discussing bias-variance trade-off is the mean-squared error (of data compared to model), which is not the same as what is assessed by reproducibility statistic R . Specifically, R is a Bayesian estimate of the posterior probability of observing a gene regulation given the data. R is calculated by taking a random sample of the data, doing the network inference again, checking if each gene regulation still appears in the GRN, and then recording (as the R statistic) the average fraction of inclusions over many repetitions. So when we have R close to 1, this indicates that our model predictions generalise well to new data, which is the opposite of what is suggested in this comment. In summary, the accuracy quantified by the reproducibility statistic R relates to the stability of the model predictions to perturbation of the data. We thank the reviewer for the helpful comment to draw our attention to this point, and have now clarified this point in the manuscript on page 6 line 252.

- The gene set enrichment results were reported only on EPI and TE cell types (Fig. 4C and Fig. *S12), due to the reason that CA data is only available for TE and ICM. However, many of the other results presented in Fig. 3-6 did include the PE cell type albeit using the same CA data. It is not particularly convincing why the cell type inclusion standard for gene set enrichment is different from the other results. *

We thank the reviewer for noting this and would like to clarify that we restricted the analysis to the EPI and TE, because similar lists of gene-sets were not available for primitive endoderm, where it is currently unclear which pathways are most relevant to this cell type. This has now been clarified in the manuscript on page 8, line 337.

- The authors cited TF binding in cis-regulatory regions as supporting evidence of several MICA-inferred regulatory interactions (e.g., NANOG -> ZNF343). However, the same cis-regulatory *evidence has already been used in the CA filtering step. All interactions passing CA filtering should in principle have TF-binding support. It would be more convincing if the authors provided other types of evidence as independent support, such as genetic associations like eQTL, experimental perturbations like gene knockdown/knockout, etc. *

We appreciate the reviewer’s point. We address this by describing published ChIP-seq validation in human pluripotent stem cells which is widely used as a proxy for the study of the epiblast. We feel that the ChIP-seq validation in this context is an appropriate independent validation to support the MICA-inferred cis regulatory interactions predicted from the human embryo datasets we analysed. Our inferences from ATAC-seq data cannot identify TF-DNA binding directly. ChIP-seq data is a widely accepted independent methods to support the inferred interactions from ATAC-seq data.

We agree that knockdown/knockout would provide further evidence suggesting gene regulation, and indeed these are experiments we would like to conduct systematically in the future, but such perturbations are difficult to achieve at genome-wide scale, especially with very restricted quantities of human embryo material. Notably, these studies would not be evidence of direct regulation and the gold-standard in our opinion is to perturb the cis regulatory region to demonstrate its functional importance in gene regulation. These are important experiments to conduct systematically in the future. We also note that assessing quantitative trait loci in the context of human pre-implantation embryos is extremely challenging due to the restricted sample sizes and genetic variance in the samples collected.

*- Many of the MICA-inferred regulatory interactions do not exhibit Spearman correlation (Fig. 5, Fig. S17), which could probably be explained by the ability of mutual information to capture complex non-monotonic dependencies. It would be interesting to provide further investigation on these "uncorrelated" edges, which may help demonstrate the superiority of mutual information over Spearman correlation. *

This has been added as a new Fig.S18.

- The authors conducted immunostaining experiments to validate the MICA-inferred regulatory *interaction between TFAP2C and JUND. While the identified protein co-localization is a step further than RNA co-expression, it is still correlation rather than causality. Additional evidence like the effect of knockout/knockdown perturbations would be more convincing. *

We agree with Reviewer 1 that experimental perturbations of TFAP2C and JUND to determine what consequence this has for interactions between these proteins would be informative. However due to the complexity of such an investigation in human embryos, we feel that this is beyond the scope of the current study. One option is to conduct the perturbations in human pluripotent stem cells, however it is unclear if the GRN in this context reflects the same interactions as human embryos and is a distinct question to address in the future. Moreover, while knockdown/knockout studies would be suggestive of up-stream regulation, it will not address the question of whether this is a direct or indirect effect without systematic further analysis including transcription factor-DNA binding (such as CUT&RUN, CUT&Tag or ChIP-seq) analysis as well as perturbations of the putative cis regulatory regions. These are all exciting future experiments and our study provides us and others with hypotheses to functionally test in the future. These are future directions and we have clarified this in the discussion section on page 16, line 576.

__Minor comments __

• *The γ symbols in AP-2γ are not correctly rendered. *

We note that this applies only to the way AP-2γ appears on the Review Commons website, and we are trying to fix this issue. We hope this transformation after the manuscript upload will not apply to a subsequent transfer to a journal.

• The UMAP figures (Fig. 4A, Fig. S7) are of low resolution compared to other figures.

We thank the reviewer for noting this. These figures have now been added as vector graphics files to overcome this issue.

• As the authors are focused on studying the blastocyst regulatory network, the inferred regulatory interactions should be provided as supplementary data.

We have included all of the inferred gene regulatory interactions as a supplementary folder for the MICA predictions using FigShare: doi.org/10.6084/m9.figshare.21968813. We have included code to reproduce the inferred gene regulatory interactions for the other methods which we compared to MICA. Because this includes 100,000 regulatory interactions per method, we feel that it would be impractical to include the alternative inferred interaction as supplementary data.

Reviewer 2

Minor comments

*- In the abstract, it would be adequate to already mention which normalisation method works the best. *

This has now been added to the abstract and we appreciate this suggestion.

*- In Fig. 1: *

* Describe what are squares and circles

This information has been included in the figure 1 legend.

** In the GRNs refined by keeping CA-predicted regulations only, mention that this are Cis interactions *

We have modified the figure 1 legend and the text on page 5, line 224 to clarify that these are putative cis-regulatory interactions.

* The ATAC seq shows KRT8, GATA3, RELB motifs, while the rest of the figure is very general. Maybe make the ATAC-seq peaks panel also as a sketch and relate it to the square/circles graphs on the right hand side to showcase how the filtering of the network is performed.

We appreciate this suggestion and modified figure 1 accordingly.

** The caption says Five GRN inference approaches, while abstract and text say 4. If is clear after reading that the 5th is a random approach. However, it was a surprise at first. *

We have modified the figure 1 legend to clarify that we also compared random prediction in addition to the 4 GRN inference approaches.

*- How the Simulation study was performed is not understandable for non experts as it is described in the Methods section. This is an important approach in general, and I think the audience would benefit if the authors add a full section about it in their supplementary data. *

Further details have now been added to the subsection ‘simulation study’ in the Methods section.

*- Fig. 2: *

** As it is, it is hard to tell the difference between GRN inference methods for a given sample size and number of regulators. Could the authors add a comparative panel for this (maybe some scatter plots would be enough)? MI by itself looks worse here? *

We thank the reviewer for this helpful suggestion. This comparative plot has now been included in figure 2 and indicates that MI is on par with the other GRN inference methods using simulation RNA-seq data.

*- When mentioning "samples" (e.g. last paragraph of section 1 in results), do the authors refer to "cells"? *

We appreciate the reviewer pointing this out and have amended the text throughout to state that these are cells.

*- What about normalisation effects in the simulated data? *

With regards to the simulated data, normalisation effects are not relevant as we are generating data that are idealised and therefore not subject to unwanted sources of variation such as read depth. However, in future work, this could be investigated with an expanded simulation study and we appreciate the reviewer’s suggestion.

*- Figure S7 should be cited in the first paragraph of section 2 in results. *

This has now been cited.

*Could the authors add a panel to indicate whether the data is SMART-seq2 or 10X. *

We thank the reviewer for the suggestion to clarify this, which we think is an important point. We have included a statement that all data used was generated using the SMART-seq2 sequencing technique in the figure legend. The choice of sequencing method/depth of sequencing will likely impact on the choice of GRN inference method and we have also clarified this in the discussion section on page 13, line 516.

*- In the association of inferred GRNs to human blastocyst cell lineages, the authors find the GRN edges predicted that overlap between the 4 inference methods in each cell type. Do they, therefore, recommend to always use more than one GRN inference method? *

Identifying overlapping inferences by comparing more than one GRN inference method may be a strategy to identify network edges with more confidence due to the agreement between several inference methodologies. However, this strategy may also miss some edges which can only be detected by one method and not another. We have included a statement in the discussion section to clarify this point on page 15, line 571.

- If the CA data used was only generated for the TE and ICM only, how do the authors use it to perform MICA on PE?

We appreciate that this is confusing and have since revised the manuscript on page 5, line 223 to state that the inner cell mass (ICM), comprises EPI (epiblast) and PE (primitive endoderm) cells. It may be that we miss putative cis-regulatory interactions if the ICM CA data does not reflect developmentally progressed PE and EPI cells and we have noted this caveat in the discussion section on page 15, line 561.

Deckt das die introspektiven Methoden, die in 2 thematisiert wurden, noch mit ab?

JAVA formatted code snippets for Logseq https://docs.logseq.com/#/page/advanced%20commands

In other words, the "problem" is the opposite of "fragmented runtimes"—it's that JS runtimes are so strictly committed to compatibility at the language level (to a degree not seen in any other widely deployed tech that I can think of). There are clear downsides (such as the sentiment expressed here), but there are also massive upsides. On net, the alignment is a huge win.

vernacular code

from : Ward

The "continue" part could be long and create difficult to understand nesting, or it could be a function which creates another layer of nesting. Code-flow-wise I would turn it around. This way you have abort criteria at the beginning and the rest of the code below.

if my_assumption is not TRUE: do something else: continue

or even better, skip the else for less nesting:

``` if my _assumption is not TRUE: do something abort

continue ```

What do you mean exactly? Compared clocks, or at least reading of them. What's a good example of this? If it's 3:55, we would say 3:55, or "5 to 4:00", but wouldn't probably say that it's "3".

1. As a partner, you create an application in the partner Dashboard and download your client credentials.
2. The merchant visits a section in your platform with the embedded Razorpay Payments setup.
3. They click Connect with Razorpay and visit the Razorpay authorization URL you initiated with the client credentials downloaded in Step 1.
4. They complete the guided onboarding process and are redirected to an authorization window.
5. The merchant gives authorization, which allows Razorpay to connect their merchant account to your Partner account.
6. On successful authorization, Razorpay redirects the user back to a URL configured by you in your application settings. While redirecting, Razorpay shares an authentication code. You need to hit our token API with this Auth code to generate Auth token.

This completes the connection setup. You should use this token to start accepting payments on behalf of the merchant.

If we are explicitly saying above that the example will be of 25RM in the code, then we should change the code here to item[amount]=2500 and item[currency]=MYR

Else we should change the line above to keep it as 500 INR

Author Response

Reviewer #2 (Public Review):

Here, a simple model of cerebellar computation is used to study the dependence of task performance on input type: it is demonstrated that task performance and optimal representations are highly dependent on task and stimulus type. This challenges many standard models which use simple random stimuli and concludes that the granular layer is required to provide a sparse representation. This is a useful contribution to our understanding of cerebellar circuits, though, in common with many models of this type, the neural dynamics and circuit architecture are not very specific to the cerebellum, the model includes the feedforward structure and the high dimension of the granule layer, but little else. This paper has the virtue of including tasks that are more realistic, but by the paper’s own admission, the same model can be applied to the electrosensory lateral line lobe and it could, though it is not mentioned in the paper, be applied to the dentate gyrus and large pyramidal cells of CA3. The discussion does not include specific elements related to, for example, the dynamics of the Purkinje cells or the role of Golgi cells, and, in a way, the demonstration that the model can encompass different tasks and stimuli types is an indication of how abstract the model is. Nonetheless, it is useful and interesting to see a generalization of what has become a standard paradigm for discussing cerebellar function.

We appreciate the Reviewer’s positive comments. Regarding the simplifications of our model, we agree that we have taken a modeling approach that abstracts away certain details to permit comparisons across systems. We now include an in-depth discussion of our simplifying assumptions (Assumptions & Extensions section in the Discussion) and have further noted the possibility that other biophysical mechanisms we have not accounted for may also underlie differences across systems.

Our results predict that qualitative differences in the coding levels of cerebellum-like systems, across brain regions or across species, reflect an optimization to distinct tasks (Figure 7). However, it is also possible that differences in coding level arise from other physiological differences between systems.

Reviewer #3 (Public Review):

1) The paper by Xie et al is a modelling study of the mossy fiber-to-granule cell-to-Purkinje cell network, reporting that the optimal type of representations in the cerebellar granule cell layer depends on the type task. The paper stresses that the findings indicate a higher overall bias towards dense representations than stated in the literature, but it appears the authors have missed parts of the literature that already reported on this. While the modelling and analysis appear mathematically solid, the model is lacking many known constraints of the cerebellar circuitry, which makes the applicability of the findings to the biological counterpart somewhat limited.

We thank the Reviewer for suggesting additional references to include in our manuscript, and for encouraging us to extend our model toward greater biological plausibility and more critically discuss simplifying assumptions we have made. We respond to both the comment about previous literature and about applicability to cerebellar circuitry in detail below.

2) I have some concerns with the novelty of the main conclusion, here from the abstract: ’Here, we generalize theories of cerebellar learning to determine the optimal granule cell representation for tasks beyond random stimulus discrimination, including continuous input-output transformations as required for smooth motor control. We show that for such tasks, the optimal granule cell representation is substantially denser than predicted by classic theories.’ Stated like this, this has in principle already been shown, i.e. for example: Spanne and Jo¨rntell (2013) Processing of multi-dimensional sensorimotor information in the spinal and cerebellar neuronal circuitry: a new hypothesis. PLoS Comput Biol. 9(3):e1002979. Indeed, even the 2 DoF arm movement control that is used in the present paper as an application, was used in this previous paper, with similar conclusions with respect to the advantage of continuous input-output transformations and dense coding. Thus, already from the beginning of this paper, the novelty aspect of this paper is questionable. Even the conclusion in the last paragraph of the Introduction: ‘We show that, when learning input-output mappings for motor control tasks, the optimal granule cell representation is much denser than predicted by previous analyses.’ was in principle already shown by this previous paper.

We thank the Reviewer for drawing our attention to Spanne and Jo¨rntell (2013). Our study shares certain similarities with this work, including the consideration of tasks with smooth input-output mappings, such as learning the dynamics of a two-joint arm. However, our study differs substantially, most notably the fact that we focus our study on parametrically varying the degree of sparsity in the granule cell layer to determine the circumstances under which dense versus sparse coding is optimal. To the best of our ability, we can find no result in Spanne and J¨orntell (2013) that indicates the performance of a network as a function of average coding level. Instead, Spanne and Jo¨rntell (2013) propose that inhibition from Golgi cells produces heterogeneity in coding level which can improve performance, which is an interesting but complementary finding to ours. We therefore do not believe that the quantitative computations of optimal coding level that we present are redundant with the results of this previous study. We also note that a key contribution of our study is mathemetical analysis of the inductive bias of networks with different coding levels which supports our conclusions.

We have included a discussion of Spanne and Jo¨rntell (2013) and (2015) in the revised version of our manuscript:

"Other studies have considered tasks with smooth input-output mappings and low-dimensional inputs, finding that heterogeneous Golgi cell inhibition can improve performance by diversifying individual granule cell thresholds (Spanne and J¨orntell, 2013). Extending our model to include heterogeneous thresholds is an interesting direction for future work. Another proposal states that dense coding may improve generalization (Spanne and Jo¨rntell, 2015). Our theory reveals that whether or not dense coding is beneficial depends on the task."

3) However, the present paper does add several more specific investigations/characterizations that were not previously explored. Many of the main figures report interesting new model results. However, the model is implemented in a highly generic fashion. Consequently, the model relates better to general neural network theory than to specific interpretations of the function of the cerebellar neuronal circuitry. One good example is the findings reported in Figure 2. These represent an interesting extension to the main conclusion, but they are also partly based on arbitrariness as the type of mossy fiber input described in the random categorization task has not been observed in the mammalian cerebellum under behavior in vivo, whereas in contrast, the type of input for the motor control task does resemble mossy fiber input recorded under behavior (van Kan et al 1993).

We agree that the tasks we consider in Figure 2 are simplified compared to those that we consider elsewhere in the paper. The choice of random mossy fiber input was made to provide a comparison to previous modeling studies that also use random input as a benchmark (Marr 1969, Albus 1971, Brunel 2004, Babadi and Sompolinsky 2014, Billings 2014, LitwinKumar et al., 2017). This baseline permits us to specifically evaluate the effects of lowdimensional inputs (Figure 2) and richer input-output mappings (Figure 2, Figure 7). We agree with the Reviewer that the random and uncorrelated mossy fiber activity that has been extensively used in previous studies is almost certainly an unrealistic idealization of in vivo neural activity—this is a motivating factor for our study, which relaxes this assumption and examines the consequences. To provide additional context, we have updated the following paragraph in the main text Results section:

"A typical assumption in computational theories of the cerebellar cortex is that inputs are randomly distributed in a high-dimensional space (Marr, 1969; Albus, 1971; Brunel et al., 2004; Babadi and Sompolinsky, 2014; Billings et al., 2014; Litwin-Kumar et al., 2017). While this may be a reasonable simplification in some cases, many tasks, including cerebellumdependent tasks, are likely best-described as being encoded by a low-dimensional set of variables. For example, the cerebellum is often hypothesized to learn a forward model for motor control (Wolpert et al., 1998), which uses sensory input and motor efference to predict an effector’s future state. Mossy fiber activity recorded in monkeys correlates with position and velocity during natural movement (van Kan et al., 1993). Sources of motor efference copies include motor cortex, whose population activity lies on a lowdimensional manifold (Wagner et al., 2019; Huang et al., 2013; Churchland et al., 2010; Yu et al., 2009). We begin by modeling the low dimensionality of inputs and later consider more specific tasks."

4) The overall conclusion states: ‘Our results....suggest that optimal cerebellar representations are task-dependent.’ This is not a particularly strong or specific conclusion. One could interpret this statement as simply saying: ‘if I construct an arbitrary neural network, with arbitrary intrinsic properties in neurons and synapses, I can get outputs that depend on the intensity of the input that I provide to that network.’ Further, the last sentence of the Introduction states: ‘More broadly, we show that the sparsity of a neural code has a task-dependent influence on learning...’ This is very general and unspecific, and would likely not come as a surprise to anyone interested in the analysis of neural networks. It doesn’t pinpoint any specific biological problem but just says that if I change the density of the input to a [generic] network, then the learning will be impacted in one way or another.

We agree with the Reviewer that our conclusions are quite general, and we have removed the final sentence as we agree it was unspecific. However, we disagree with the Reviewer’s paraphrasing of our results.

First, we do not select arbitrary intrinsic properties of neurons and synapses. Rather, we construct a simplified model with a key quantity, the neuronal threshold, that we vary parametrically in order to assess the effect of the resulting changes in the representation on performance. Second, we do not vary the intensity/density of inputs provided to the network – this is fixed throughout our study for all key comparisons we perform. Instead, we vary the density (coding level) of the expansion layer representation and quantify its effect on inductive bias and generalization. Finally, our study’s key contribution is an explanation of the heterogeneity in average coding level observed across behaviors and cerebellum-like systems. We go beyond the empirical statement that there is a dependence of performance on the parameter that we vary by developing an analytical theory. Our theory describes the performance of the class of networks that we study and the properties of learning tasks that determine the optimal expansion layer representation.

To clarify our main contributions, we have updated the final paragraph of the Introduction. We have also removed the sentence that the Reviewer objects to, as it was less specific than the other points we make here.

"We propose that these differences can be explained by the capacity of representations with different levels of sparsity to support learning of different tasks. We show that the optimal level of sparsity depends on the structure of the input-output relationship of a task. When learning input-output mappings for motor control tasks, the optimal granule cell representation is much denser than predicted by previous analyses. To explain this result, we develop an analytic theory that predicts the performance of cerebellum-like circuits for arbitrary learning tasks. The theory describes how properties of cerebellar architecture and activity control these networks’ inductive bias: the tendency of a network toward learning particular types of input-output mappings (Sollich, 1998; Jacot et al., 2018; Bordelon et al., 2020; Canatar et al., 2021; Simon et al., 2021). The theory shows that inductive bias, rather than the dimension of the representation alone, is necessary to explain learning performance across tasks. It also suggests that cerebellar regions specialized for different functions may adjust the sparsity of their granule cell representations depending on the task."

5) The interpretation of the distribution of the mossy fiber inputs to the granule cells, which would have a crucial impact on the results of a study like this, is likely incorrect. First, unlike the papers that the authors cite, there are many studies indicating that there is a topographic organization in the mossy fiber termination, such that mossy fibers from the same inputs, representing similar types of information, are regionally co-localized in the granule cell layer. Hence, there is no support for the model assumption that there is a predominantly random termination of mossy fibers of different origins. This risks invalidating the comparisons that the authors are making, i.e. such as in Figure 3. This is a list of example papers, there are more: van Kan, Gibson and Houk (1993) Movement-related inputs to intermediate cerebellum of the monkey. Journal of Neurophysiology. Garwicz et al (1998) Cutaneous receptive fields and topography of mossy fibres and climbing fibres projecting to cat cerebellar C3 zone. The Journal of Physiology. Brown and Bower (2001) Congruence of mossy fiber and climbing fiber tactile projections in the lateral hemispheres of the rat cerebellum. The Journal of Comparative Neurology. Na, Sugihara, Shinoda (2019) The entire trajectories of single pontocerebellar axons and their lobular and longitudinal terminal distribution patterns in multiple aldolase C-positive compartments of the rat cerebellar cortex. The Journal of Comparative Neurology.

6) The nature of the mossy fiber-granule cell recording is also reviewed here: Gilbert and Miall (2022) How and Why the Cerebellum Recodes Input Signals: An Alternative to Machine Learning. The Neuroscientist. Further, considering the re-coding idea, the following paper shows that detailed information, as it is provided by mossy fibers, is transmitted through the granule cells without any evidence of re-coding: Jo¨rntell and Ekerot (2006) Journal of Neuroscience; and this paper shows that these granule inputs are powerfully transmitted to the molecular layer even in a decerebrated animal (i.e. where only the ascending sensory pathways remains) Jo¨rntell and Ekerot 2002, Neuron.

We agree that there is strong evidence for a topographic organization in mossy fiber to granule cell connectivity at the microzonal level. We thank the Reviewer for pointing us to specific examples. We acknowledge that our simplified model does not capture the structure of connectivity observed in these studies.

However, the focus of our model is on cerebellar neurons presynaptic to a single Purkinje cell. Random or disordered distribution of inputs at this local scale is compatible with topographic organization at the microzonal scale. Furthermore, while there is evidence of structured connections at the local scale, models with random connectivity are able to reproduce the dimensionality of granule cell activity within a small margin of error (Nguyen et al., 2022). Finally, our finding that dense codes are optimal for learning slowly varying tasks is consistent with evidence for the lack of re-coding – for such tasks, re-coding may absent because it is not required.

We have dedicated a section on this issue in the Assumptions and Extensions portion of our Discussion:

"Another key assumption concerning the granule cells is that they sample mossy fiber inputs randomly, as is typically assumed in Marr-Albus models (Marr, 1969; Albus, 1971; LitwinKumar et al., 2017; Cayco-Gajic et al., 2017). Other studies instead argue that granule cells sample from mossy fibers with highly similar receptive fields (Garwicz et al., 1998; Brown and Bower, 2001; J¨orntell and Ekerot, 2006) defined by the tuning of mossy fiber and climbing fiber inputs to cerebellar microzones (Apps et al., 2018). This has led to an alternative hypothesis that granule cells serve to relay similarly tuned mossy fiber inputs and enhance their signal-to-noise ratio (Jo¨rntell and Ekerot, 2006; Gilbert and Chris Miall, 2022) rather than to re-encode inputs. Another hypothesis is that granule cells enable Purkinje cells to learn piece-wise linear approximations of nonlinear functions (Spanne and J¨orntell, 2013). However, several recent studies support the existence of heterogeneous connectivity and selectivity of granule cells to multiple distinct inputs at the local scale (Huang et al., 2013; Ishikawa et al., 2015). Furthermore, the deviation of the predicted dimension in models constrained by electron-microscopy data as compared to randomly wired models is modest (Nguyen et al., 2022). Thus, topographically organized connectivity at the macroscopic scale may coexist with disordered connectivity at the local scale, allowing granule cells presynaptic to an individual Purkinje cell to sample heterogeneous combinations of the subset of sensorimotor signals relevant to the tasks that Purkinje cell participates in. Finally, we note that the optimality of dense codes for learning slowly varying tasks in our theory suggests that observations of a lack of mixing (J¨orntell and Ekerot, 2002) for such tasks are compatible with Marr-Albus models, as in this case nonlinear mixing is not required."

7) I could not find any description of the neuron model used in this paper, so I assume that the neurons are just modelled as linear summators with a threshold (in fact, Figure 5 mentions inhibition, but this appears to be just one big lump inhibition, which basically is an incorrect implementation). In reality, granule cells of course do have specific properties that can impact the input-output transformation, PARTICULARLY with respect to the comparison of sparse versus dense coding, because the low-pass filtering of input that occurs in granule cells (and other neurons) as well as their spike firing stochasticity (Saarinen et al (2008). Stochastic differential equation model for cerebellar granule cell excitability. PLoS Comput. Biol. 4:e1000004) will profoundly complicate these comparisons and make them less straight forward than what is portrayed in this paper. There are also several other factors that would be present in the biological setting but are lacking here, which makes it doubtful how much information in relation to the biological performance that this modelling study provides: What are the types of activity patterns of the inputs? What are the learning rules? What is the topography? What is the impact of Purkinje cell outputs downstream, as the Purkinje cell output does not have any direct action, it acts on the deep cerebellar nuclear neurons, which in turn act on a complex sensorimotor circuitry to exert their effect, hence predictive coding could only become interpretable after the PC output has been added to the activity in those circuits. Where is the differentiated Golgi cell inhibition?

Thank you for these critiques. We have made numerous edits to improve the presentation of the details of our model in the main text of the manuscript. Indeed, granule cells in the main text are modeled as linear sums of mossy fiber inputs with a threshold-linear activation function. A more detailed description of the model for granule cells can now be found in Equation 1 in the Results section:

"The activity of neurons in the expansion layer is given by: h = φ(Jeffx − θ), (1) where φ is a rectified linear activation function φ(u) = max(u,0) applied element-wise. Our results also hold for other threshold-polynomial activation functions. The scalar threshold θ is shared across neurons and controls the coding level, which we denote by f, defined as the average fraction of neurons in the expansion layer that are active."

Most of our analyses use the firing rate model we describe above, but several Supplemental Figures show extensions to this model. As we mention in the Discussion, our results do not depend on the specific choice of nonlinearity (Figure 2-figure supplement 2). We have also considered the possibility that the stochastic nature of granule cell spikes could impact our measures of coding level. In Figure 7-figure supplement 1 we test the robustness of our main conclusion using a spiking model where we model granule cell spikes with Poisson statistics. When measuring coding level in a population of spiking neurons, a key question is at what time window the Purkinje cell integrates spikes. For several choices of integration time windows, we show that dense coding remains optimal for learning smooth tasks. However, we agree with the Reviewer that there are other biological details our model does not address. For example, our spiking model does not capture some of the properties the Saarinen et al. (2008) model captures, including random sub-threshold oscillations and clusters of spikes. Modeling biophysical phenomena at this scale is beyond the scope of our study. We have added this reference to the relevant section of the Discussion:

"We also note that coding level is most easily defined when neurons are modeled as rate, rather than spiking units. To investigate the consistency of our results under a spiking code, we implemented a model in which granule cell spiking exhibits Poisson variability and quantify coding level as the fraction of neurons that have nonzero spike counts (Figure 7-figure supplement 1; Figure 7C). In general, increased spike count leads to improved performance as noise associated with spiking variability is reduced. Granule cells have been shown to exhibit reliable burst responses to mossy fiber stimulation (Chadderton et al., 2004), motivating models using deterministic responses or sub-Poisson spiking variability. However, further work is needed to quantitatively compare variability in model and experiment and to account for more complex biophysical properties of granule cells (Saarinen et al., 2008)."

A second concern the Reviewer raises is our implementation of Golgi cell inhibition as a homogeneous rather than heterogeneous input onto granule cells. In simplified models, adding heterogeneous inhibition does not dramatically change the qualitative properties of the expansion layer representation, in particular the dimensionality of the representation (Billings et al., 2014, Cayco-Gajic et al., 2017, Litwin-Kumar et al., 2017). We have added a section about inhibition to our Discussion:

"We also have not explicitly modeled inhibitory input provided by Golgi cells, instead assuming such input can be modeled as a change in effective threshold, as in previous studies (Billings et al., 2014; Cayco-Gajic et al., 2017; Litwin-Kumar et al., 2017). This is appropriate when considering the dimension of the granule cell representation (Litwin-Kumar et al., 2017), but more work is needed to extend our model to the case of heterogeneous inhibition."

Regarding the mossy fiber inputs, as we state in response to paragraph 3, we agree with the Reviewer that the random and uncorrelated mossy fiber activity that has been used in previous studies is an unrealistic idealization of in vivo neural activity. One of the motivations for our model was to relax this assumption and examine the consequences: we introduce correlations in the mossy fiber activity by projecting low-dimensional patterns into the mossy fiber layer (Figure 1B):

"A typical assumption in computational theories of the cerebellar cortex is that inputs are randomly distributed in a high-dimensional space (Marr, 1969; Albus, 1971; Brunel et al., 2004; Babadi and Sompolinsky, 2014; Billings et al., 2014; Litwin-Kumar et al., 2017). While this may be a reasonable simplification in some cases, many tasks, including cerebellumdependent tasks, are likely best-described as being encoded by a low-dimensional set of variables. For example, the cerebellum is often hypothesized to learn a forward model for motor control (Wolpert et al., 1998), which uses sensory input and motor efference to predict an effector’s future state. Mossy fiber activity recorded in monkeys correlates with position and velocity during natural movement (van Kan et al., 1993). Sources of motor efference copies include motor cortex, whose population activity lies on a low-dimensional manifold (Wagner et al., 2019; Huang et al., 2013; Churchland et al., 2010; Yu et al., 2009). We begin by modeling the low dimensionality of inputs and later consider more specific tasks.

We therefore assume that the inputs to our model lie on a D-dimensional subspace embedded in the N-dimensional input space, where D is typically much smaller than N (Figure 1B). We refer to this subspace as the “task subspace” (Figure 1C)."

The Reviewer also mentions the learning rule at granule cell to Purkinje cell synapses. We agree that considering online, climbing-fiber-dependent learning is an important generalization. We therefore added a new supplemental figure investigating whether we would still see a difference in optimal coding levels across tasks if online learning were used instead of the least squares solution (Figure 7-figure supplement 2). Indeed, we observed a similar task dependence as we saw in Figure 2F. We have added a new paragraph in the Discussion under Assumptions and Extensions describing our rationale and approach in detail:

"For the Purkinje cells, our model assumes that their responses to granule cell input can be modeled as an optimal linear readout. Our model therefore provides an upper bound to linear readout performance, a standard benchmark for the quality of a neural representation that does not require assumptions on the nature of climbing fiber-mediated plasticity, which is still debated. Electrophysiological studies have argued in favor of a linear approximation (Brunel et al., 2004). To improve the biological applicability of our model, we implemented an online climbing fiber-mediated learning rule and found that optimal coding levels are still task-dependent (Figure 7-figure supplement 2). We also note that although we model several timing-dependent tasks (Figure 7), our learning rule does not exploit temporal information, and we assume that temporal dynamics of granule cell responses are largely inherited from mossy fibers. Integrating temporal information into our model is an interesting direction for future investigation."

Finally, regarding the function of the Purkinje cell, our model defines a learning task as a mapping from inputs to target activity in the Purkinje cell and is thus agnostic to the cell’s downstream effects. We clarify this point when introducing the definition of a learning task:

"In our model, a learning task is defined by a mapping from task variables x to an output f(x), representing a target change in activity of a readout neuron, for example a Purkinje cell. The limited scope of this definition implies our results should not strongly depend on the influence of the readout neuron on downstream circuits."

8) The problem of these, in my impression, generic, arbitrary settings of the neurons and the network in the model becomes obvious here: ‘In contrast to the dense activity in cerebellar granule cells, odor responses in Kenyon cells, the analogs of granule cells in the Drosophila mushroom body, are sparse...’ How can this system be interpreted as an analogy to granule cells in the mammalian cerebellum when the model does not address the specifics lined up above? I.e. the ‘inductive bias’ that the authors speak of, defined as ‘the tendency of a network toward learning particular types of input-output mappings’, would be highly dependent on the specifics of the network model.

We agree with the Reviewer that our model makes several simplifying assumptions for mathematical tractability. However, we note that our study is not the first to draw analogies between cerebellum-like systems, including the mushroom body (Bell et al., 2008; Farris, 2011). All the systems we study feature a sparsely connected, expanded granule-like layer that sends parallel fiber axons onto densely connected downstream neurons known to exhibit powerful synaptic plasticity, thus motivating the key architectural assumptions of our model. We have constrained anatomical parameters of the model using data as available (Table 1). However, we agree with the Reviewer that when making comparisons across species there is always a possibility that differences are due to physiological mechanisms we have not fully understood or captured with a model. As such, we can only present a hypothesis for these differences. We have modified our Discussion section on this topic to clearly state this.

"Our results predict that qualitative differences in the coding levels of cerebellum-like systems, across brain regions or across species, reflect an optimization to distinct tasks (Figure 7). However, it is also possible that differences in coding level arise from other physiological differences between systems."

9) More detailed comments: Abstract: ‘In these models [Marr-Albus], granule cells form a sparse, combinatorial encoding of diverse sensorimotor inputs. Such sparse representations are optimal for learning to discriminate random stimuli.’ Yes, I would agree with the first part, but I contest the second part of this statement. I think what is true for sparse coding is that the learning of random stimuli will be faster, as in a perceptron, but not necessarily better. As the sparsification essentially removes information, it could be argued that the quality of the learning is poorer. So from that perspective, it is not optimal. The authors need to specify from what perspective they consider sparse representations optimal for learning.

This is an important point that we would like to clarify. It is not the case that sparse coding simply speeds up learning. In our study and many related works (Barak et al. 2013; Babadi and Sompolinsky 2014; Litwin-Kumar et al. 2017), learning performance is measured based on the generalization ability of the network – the ability to predict correct labels for previously unseen inputs. As our study and previous studies show, sparse codes are optimal in the sense that they minimize generalization error, independent of any effect on learning speed. To communicate this more effectively, we have added the following sentence to the first paragraph of the Introduction:

"Sparsity affects both learning speed (Cayco-Gajic et al., 2017), and generalization, the ability to predict correct labels for previously unseen inputs (Barak et al., 2013; Babadi and Sompolinsky, 2014; Litwin-Kumar et al., 2017)."

10) Introduction: ‘Indeed, several recent studies have reported dense activity in cerebellar granule cells in response to sensory stimulation or during motor control tasks (Knogler et al., 2017; Wagner et al., 2017; Giovannucci et al., 2017; Badura and De Zeeuw, 2017; Wagner et al., 2019), at odds with classic theories (Marr, 1969; Albus, 1971).’ In fact, this was precisely the issue that was addressed already by Jo¨rntell and Ekerot (2006) Journal of Neuroscience. The conclusion was that these actual recordings of granule cells in vivo provided essentially no support for the assumptions in the Marr-Albus theories.

In our reading, the main finding of J¨orntell and Ekerot (2006) is that individual granule cells are activated by mossy fibers with overlapping receptive fields driven by a single type of somatosensory input. However, there is also evidence of nonlinear mixed selectivity in granule cells in support of the re-coding hypothesis (Huang et al., 2013; Ishikawa et al., 2015). Jo¨rntell and Ekerot (2006) also suggest that the granule cell layer shares similar topographic organization as mossy fibers, organized into microzones. The existence of topographic organization does not invalidate Marr-Albus theories. As we have suggested earlier, a local combinatorial expansion can coexist with a global topographic organization.

We have described these considerations in the Assumptions and Extensions portion of the Discussion:

"Another key assumption concerning the granule cells is that they sample mossy fiber inputs randomly, as is typically assumed in Marr-Albus models (Marr, 1969; Albus, 1971; LitwinKumar et al., 2017; Cayco-Gajic et al., 2017). Other studies instead argue that granule cells sample from mossy fibers with highly similar receptive fields (Garwicz et al., 1998; Brown and Bower, 2001; J¨orntell and Ekerot, 2006) defined by the tuning of mossy fiber and climbing fiber inputs to cerebellar microzones (Apps et al., 2018). This has led to an alternative hypothesis that granule cells serve to relay similarly tuned mossy fiber inputs and enhance their signal-to-noise ratio (Jo¨rntell and Ekerot, 2006; Gilbert and Chris Miall, 2022) rather than to re-encode inputs. Another hypothesis is that granule cells enable Purkinje cells to learn piece-wise linear approximations of nonlinear functions (Spanne and J¨orntell, 2013). However, several recent studies support the existence of heterogeneous connectivity and selectivity of granule cells to multiple distinct inputs at the local scale (Huang et al., 2013; Ishikawa et al., 2015). Furthermore, the deviation of the predicted dimension in models constrained by electron-microscopy data as compared to randomly wired models is modest (Nguyen et al., 2022). Thus, topographically organized connectivity at the macroscopic scale may coexist with disordered connectivity at the local scale, allowing granule cells presynaptic to an individual Purkinje cell to sample heterogeneous combinations of the subset of sensorimotor signals relevant to the tasks that Purkinje cell participates in. Finally, we note that the optimality of dense codes for learning slowly varying tasks in our theory suggests that observations of a lack of mixing (J¨orntell and Ekerot, 2002) for such tasks are compatible with Marr-Albus models, as in this case nonlinear mixing is not required."

We have also included the Jo¨rntell and Ekerot (2006) study as a citation in the Introduction:

"Indeed, several recent studies have reported dense activity in cerebellar granule cells in response to sensory stimulation or during motor control tasks (Jo¨rntell and Ekerot, 2006; Knogler et al., 2017; Wagner et al., 2017; Giovannucci et al., 2017; Badura and De Zeeuw, 2017; Wagner et al., 2019), at odds with classic theories (Marr, 1969; Albus, 1971)."

11) Results: 1st para: There is no information about how the granule cells are modelled.

We agree that this should information should have been more readily available. We now more completely describe the model in the main text. Our model for granule cells can be found in Equation 1 in the Results section and also the Methods (Network Model):

"The activity of neurons in the expansion layer is given by: h = φ(Jeffx − θ), (2)

where φ is a rectified linear activation function φ(u) = max(u,0) applied element-wise. Our results also hold for other threshold-polynomial activation functions. The scalar threshold θ is shared across neurons and controls the coding level, which we denote by f, defined as the average fraction of neurons in the expansion layer that are active."

12) 2nd para: ‘A typical assumption in computational theories of the cerebellar cortex is that inputs are randomly distributed in a high-dimensional space.’ Yes, I agree, and this is in fact in conflict with the known topographical organization in the cerebellar cortex (see broader comment above). Mossy fiber inputs coding for closely related inputs are co-localized in the cerebellar cortex. I think for this model to be of interest from the point of view of the mammalian cerebellar cortex, it would need to pay more attention to this organizational feature.

As we discuss in our response to paragraphs 5 and 6, we see the random distribution assumption at the local scale (inputs presynaptic to a single Purkinje cell) as being compatible with topographic organization occurring at the microzone scale. Furthermore, as discussed earlier, we specifically model low-dimensional input as opposed to the random and high-dimensional inputs typically studied in prior models.

"A typical assumption in computational theories of the cerebellar cortex is that inputs are randomly distributed in a high-dimensional space (Marr, 1969; Albus, 1971; Brunel et al., 2004; Babadi and Sompolinsky, 2014; Billings et al., 2014; Litwin-Kumar et al., 2017). While this may be a reasonable simplification in some cases, many tasks, including cerebellumdependent tasks, are likely best-described as being encoded by a low-dimensional set of variables. For example, the cerebellum is often hypothesized to learn a forward model for motor control (Wolpert et al., 1998), which uses sensory input and motor efference to predict an effector’s future state. Mossy fiber activity recorded in monkeys correlates with position and velocity during natural movement (van Kan et al., 1993). Sources of motor efference copies include motor cortex, whose population activity lies on a low-dimensional manifold (Wagner et al., 2019; Huang et al., 2013; Churchland et al., 2010; Yu et al., 2009). We begin by modeling the low dimensionality of inputs and later consider more specific tasks. We therefore assume that the inputs to our model lie on a D-dimensional subspace embedded in the N-dimensional input space, where D is typically much smaller than N (Figure 1B). We refer to this subspace as the “task subspace” (Figure 1C)."

References

Albus, J.S. (1971). A theory of cerebellar function. Mathematical Biosciences 10, 25–61.

Apps, R., et al. (2018). Cerebellar Modules and Their Role as Operational Cerebellar Processing Units. Cerebellum 17, 654–682.

Babadi, B. and Sompolinsky, H. (2014). Sparseness and expansion in sensory representations. Neuron 83, 1213–1226.

Badura, A. and De Zeeuw, C.I. (2017). Cerebellar granule cells: dense, rich and evolving representations. Current Biology 27, R415–R418.

Barak, O., Rigotti, M., and Fusi, S. (2013). The sparseness of mixed selectivity neurons controls the generalization–discrimination trade-off. Journal of Neuroscience 33, 3844– 3856.

Bell, C.C., Han, V., and Sawtell, N.B. (2008). Cerebellum-like structures and their implications for cerebellar function. Annual Review of Neuroscience 31, 1–24.

Billings, G., Piasini, E., Lo˝rincz, A., Nusser, Z., and Silver, R.A. (2014). Network structure within the cerebellar input layer enables lossless sparse encoding. Neuron 83, 960–974.

Bordelon, B., Canatar, A., and Pehlevan, C. (2020). Spectrum dependent learning curves in kernel regression and wide neural networks. International Conference on Machine Learning 1024–1034.

Brown, I.E. and Bower, J.M. (2001). Congruence of mossy fiber and climbing fiber tactile projections in the lateral hemispheres of the rat cerebellum. Journal of Comparative Neurology 429, 59–70.

Brunel, N., Hakim, V., Isope, P., Nadal, J.P., and Barbour, B. (2004). Optimal information storage and the distribution of synaptic weights: perceptron versus Purkinje cell. Neuron 43, 745–757.

Canatar, A., Bordelon, B., and Pehlevan, C. (2021). Spectral bias and task-model alignment explain generalization in kernel regression and infinitely wide neural networks. Nature Communications 12, 1–12.

Cayco-Gajic, N.A., Clopath, C., and Silver, R.A. (2017). Sparse synaptic connectivity is required for decorrelation and pattern separation in feedforward networks. Nature Communications 8, 1–11.

Chadderton, P., Margrie, T.W., and Ha¨usser, M. (2004). Integration of quanta in cerebellar granule cells during sensory processing. Nature 428, 856–860.

Churchland, M.M., et al. (2010). Stimulus onset quenches neural variability: a widespread cortical phenomenon. Nature Neuroscience 13, 369–378.

Farris, S.M. (2011). Are mushroom bodies cerebellum-like structures? Arthropod structure & development 40, 368–379.

Garwicz, M., Jorntell, H., and Ekerot, C.F. (1998). Cutaneous receptive fields and topography of mossy fibres and climbing fibres projecting to cat cerebellar C3 zone. The Journal of Physiology 512 ( Pt 1), 277–293.

Gilbert, M. and Chris Miall, R. (2022). How and Why the Cerebellum Recodes Input Signals: An Alternative to Machine Learning. The Neuroscientist 28, 206–221.

Giovannucci, A., et al. (2017). Cerebellar granule cells acquire a widespread predictive feedback signal during motor learning. Nature Neuroscience 20, 727–734.

Huang, C.C., et al. (2013). Convergence of pontine and proprioceptive streams onto multimodal cerebellar granule cells. eLife 2, e00400.

Ishikawa, T., Shimuta, M., and Ha¨usser, M. (2015). Multimodal sensory integration in single cerebellar granule cells in vivo. eLife 4, e12916.

Jacot, A., Gabriel, F., and Hongler, C. (2018). Neural tangent kernel: Convergence and generalization in neural networks. Advances in Neural Information Processing Systems 31.

Jo¨rntell, H. and Ekerot, C.F. (2002). Reciprocal Bidirectional Plasticity of Parallel Fiber Receptive Fields in Cerebellar Purkinje Cells and Their Afferent Interneurons. Neuron 34, 797–806.

Jorntell, H. and Ekerot, C.F. (2006). Properties of Somatosensory Synaptic Integration in Cerebellar Granule Cells In Vivo. Journal of Neuroscience 26, 11786–11797.

Knogler, L.D., Markov, D.A., Dragomir, E.I., Stih, V., and Portugues, R. (2017). Senso-ˇ rimotor representations in cerebellar granule cells in larval zebrafish are dense, spatially organized, and non-temporally patterned. Current Biology 27, 1288–1302.

Litwin-Kumar, A., Harris, K.D., Axel, R., Sompolinsky, H., and Abbott, L.F. (2017). Optimal degrees of synaptic connectivity. Neuron 93, 1153–1164. Marr, D. (1969). A theory of cerebellar cortex. Journal of Physiology 202, 437–470.

Nguyen, T.M., et al. (2022). Structured cerebellar connectivity supports resilient pattern separation. Nature 1–7.

Saarinen, A., Linne, M.L., and Yli-Harja, O. (2008). Stochastic Differential Equation Model for Cerebellar Granule Cell Excitability. PLOS Computational Biology 4, e1000004.

Simon, J.B., Dickens, M., and DeWeese, M.R. (2021). A theory of the inductive bias and generalization of kernel regression and wide neural networks. arXiv: 2110.03922.

Sollich, P. (1998). Learning curves for Gaussian processes. Advances in Neural Information Processing Systems 11.

Spanne, A. and Jo¨rntell, H. (2013). Processing of Multi-dimensional Sensorimotor Information in the Spinal and Cerebellar Neuronal Circuitry: A New Hypothesis. PLOS Computational Biology 9, e1002979.

Spanne, A. and Jo¨rntell, H. (2015). Questioning the role of sparse coding in the brain. Trends in Neurosciences 38, 417–427.

van Kan, P.L., Gibson, A.R., and Houk, J.C. (1993). Movement-related inputs to intermediate cerebellum of the monkey. Journal of Neurophysiology 69, 74–94.

Wagner, M.J., Kim, T.H., Savall, J., Schnitzer, M.J., and Luo, L. (2017). Cerebellar granule cells encode the expectation of reward. Nature 544, 96–100.

Wagner, M.J., et al. (2019). Shared cortex-cerebellum dynamics in the execution and learning of a motor task. Cell 177, 669–682.e24.

Wolpert, D.M., Miall, R.C., and Kawato, M. (1998). Internal models in the cerebellum. Trends in Cognitive Sciences 2, 338–347.

Yu, B.M., et al. (2009). Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity. Journal of Neurophysiology 102, 614–635.

Reviewer #3 (Public Review):

The paper by Xie et al is a modelling study of the mossy fiber-to-granule cell-to-Purkinje cell network, reporting that the optimal type of representations in the cerebellar granule cell layer depends on the type task. The paper stresses that the findings indicate a higher overall bias towards dense representations than stated in the literature, but it appears the authors have missed parts of the literature that already reported on this. While the modelling and analysis appear mathematically solid, the model is lacking many known constraints of the cerebellar circuitry, which makes the applicability of the findings to the biological counterpart somewhat limited.

I have some concerns with the novelty of the main conclusion, here from the abstract:<br /> 'Here, we generalize theories of cerebellar learning to determine the optimal granule cell representation for tasks beyond random stimulus discrimination, including continuous input-output transformations as required for smooth motor control. We show that for such tasks, the optimal granule cell representation is substantially denser than predicted by classic theories.'<br /> Stated like this, this has in principle already been shown, i.e. for example:<br /> Spanne and Jorntell (2013) Processing of multi-dimensional sensorimotor information in the spinal and cerebellar neuronal circuitry: a new hypothesis. PLoS Comput Biol. 9(3):e1002979.<br /> Indeed, even the 2 DoF arm movement control that is used in the present paper as an application, was used in this previous paper, with similar conclusions with respect to the advantage of continuous input-output transformations and dense coding. Thus, already from the beginning of this paper, the novelty aspect of this paper is questionable. Even the conclusion in the last paragraph of the Introduction: 'We show that, when learning input-output mappings for motor control tasks, the optimal granule cell representation is much denser than predicted by previous analyses.' was in principle already shown by this previous paper.

However, the present paper does add several more specific investigations/characterizations that were not previously explored. Many of the main figures report interesting new model results. However, the model is implemented in a highly generic fashion. Consequently, the model relates better to general neural network theory than to specific interpretations of the function of the cerebellar neuronal circuitry. One good example is the findings reported in Figure 2. These represent an interesting extension to the main conclusion, but they are also partly based on arbitrariness as the type of mossy fiber input described in the random categorization task has not been observed in the mammalian cerebellum under behavior in vivo, whereas in contrast, the type of input for the motor control task does resemble mossy fiber input recorded under behavior (van Kan et al 1993).

The overall conclusion states:<br /> 'Our results....suggest that optimal cerebellar representations are task-dependent.'<br /> This is not a particularly strong or specific conclusion. One could interpret this statement as simply saying: ' if I construct an arbitrary neural network, with arbitrary intrinsic properties in neurons and synapses, I can get outputs that depend on the intensity of the input that I provide to that network.'<br /> Further, the last sentence of the Introduction states: 'More broadly, we show that the sparsity of a neural code has a task-dependent influence on learning...' This is very general and unspecific, and would likely not come as a surprise to anyone interested in the analysis of neural networks. It doesn't pinpoint any specific biological problem but just says that if I change the density of the input to a [generic] network, then the learning will be impacted in one way or another.

The interpretation of the distribution of the mossy fiber inputs to the granule cells, which would have a crucial impact on the results of a study like this, is likely incorrect. First, unlike the papers that the authors cite, there are many studies indicating that there is a topographic organization in the mossy fiber termination, such that mossy fibers from the same inputs, representing similar types of information, are regionally co-localized in the granule cell layer. Hence, there is no support for the model assumption that there is a predominantly random termination of mossy fibers of different origins. This risks invalidating the comparisons that the authors are making, i.e. such as in Figure 3. This is a list of example papers, there are more:<br /> van Kan, Gibson and Houk (1993) Movement-related inputs to intermediate cerebellum of the monkey. Journal of Neurophysiology.<br /> Garwicz et al (1998) Cutaneous receptive fields and topography of mossy fibres and climbing fibres projecting to cat cerebellar C3 zone. The Journal of Physiology.<br /> Brown and Bower (2001) Congruence of mossy fiber and climbing fiber tactile projections in the lateral hemispheres of the rat cerebellum. The Journal of Comparative Neurology.<br /> Na, Sugihara, Shinoda (2019) The entire trajectories of single pontocerebellar axons and their lobular and longitudinal terminal distribution patterns in multiple aldolase C-positive compartments of the rat cerebellar cortex. The Journal of Comparative Neurology.

The nature of the mossy fiber-granule cell recording is also reviewed here:<br /> Gilbert and Miall (2022) How and Why the Cerebellum Recodes Input Signals: An Alternative to Machine Learning. The Neuroscientist<br /> Further, considering the recoding idea, the following paper shows that detailed information, as it is provided by mossy fibers, is transmitted through the granule cells without any evidence of recoding: Jorntell and Ekerot (2006) Journal of Neuroscience; and this paper shows that these granule inputs are powerfully transmitted to the molecular layer even in a decerebrated animal (i.e. where only the ascending sensory pathways remains) Jorntell and Ekerot 2002, Neuron.

I could not find any description of the neuron model used in this paper, so I assume that the neurons are just modelled as linear summators with a threshold (in fact, Figure 5 mentions inhibition, but this appears to be just one big lump inhibition, which basically is an incorrect implementation). In reality, granule cells of course do have specific properties that can impact the input-output transformation, PARTICULARLY with respect to the comparison of sparse versus dense coding, because the low-pass filtering of input that occurs in granule cells (and other neurons) as well as their spike firing stochasticity (Saarinen et al (2008). Stochastic differential equation model for cerebellar granule cell excitability. PLoS Comput. Biol. 4:e1000004) will profoundly complicate these comparisons and make them less straight forward than what is portrayed in this paper. There are also several other factors that would be present in the biological setting but are lacking here, which makes it doubtful how much information in relation to the biological performance that this modelling study provides:<br /> What are the types of activity patterns of the inputs? What are the learning rules? What is the topography? What is the impact of Purkinje cell outputs downstream, as the Purkinje cell output does not have any direct action, it acts on the deep cerebellar nuclear neurons, which in turn act on a complex sensorimotor circuitry to exert their effect, hence predictive coding could only become interpretable after the PC output has been added to the activity in those circuits. Where is the differentiated Golgi cell inhibition?

The problem of these, in my impression, generic, arbitrary settings of the neurons and the network in the model becomes obvious here: 'In contrast to the dense activity in cerebellar granule cells, odor responses in Kenyon cells, the analogs of granule cells in the Drosophila mushroom body, are sparse...' How can this system be interpreted as an analogy to granule cells in the mammalian cerebellum when the model does not address the specifics lined up above? I.e. the 'inductive bias' that the authors speak of, defined as 'the tendency of a network toward learning particular types of input-output mappings', would be highly dependent on the specifics of the network model.

More detailed comments:<br /> Abstract:<br /> 'In these models [Marr-Albus], granule cells form a sparse, combinatorial encoding of diverse sensorimotor inputs. Such sparse representations are optimal for learning to discriminate random stimuli.' Yes, I would agree with the first part, but I contest the second part of this statement. I think what is true for sparse coding is that the learning of random stimuli will be faster, as in a perceptron, but not necessarily better. As the sparsification essentially removes information, it could be argued that the quality of the learning is poorer. So from that perspective, it is not optimal. The authors need to specify from what perspective they consider sparse representations optimal for learning.

Introduction:<br /> 'Indeed, several recent studies have reported dense activity in cerebellar granule cells in response to sensory stimulation or during motor control tasks (Knogler et al., 2017; Wagner et al., 2017; Giovannucci et al., 2017; Badura and De Zeeuw, 2017; Wagner et al., 2019), at odds with classic theories (Marr, 1969; Albus, 1971).' In fact, this was precisely the issue that was addressed already by Jorntell and Ekerot (2006) Journal of Neuroscience. The conclusion was that these actual recordings of granule cells in vivo provided essentially no support for the assumptions in the Marr-Albus theories.

Results:<br /> 1st para: There is no information about how the granule cells are modelled.

2nd para: 'A typical assumption in computational theories of the cerebellar cortex is that inputs are randomly distributed in a high-dimensional space.' Yes, I agree, and this is in fact in conflict with the known topographical organization in the cerebellar cortex (see broader comment above). Mossy fiber inputs coding for closely related inputs are co-localized in the cerebellar cortex. I think for this model to be of interest from the point of view of the mammalian cerebellar cortex, it would need to pay more attention to this organizational feature.

I updated the req doc with code snippets. Please check and update here

Reviewer #1 (Public Review):

Summary:<br /> Codol et al. present a toolbox that allows simulating biomechanically realistic effectors and training Artificial Neural Networks (ANNs) to control them. The paper provides a detailed explanation of how the toolbox is structured and several examples that demonstrate its usefulness.

Main comments:<br /> 1. The paper is well written and easy to follow. The schematics help in understanding how the toolbox works and the examples provide an idea of the results that the user can obtain.

2. As I understand it, the main purpose of the paper should be to facilitate the usage of the toolbox. For this reason, I have missed a more explicit link to the actual code. As I see it, researchers will read this paper to figure out whether they can use MotorNet to simulate their experiments, and how they should proceed if they decide to use it. I'd say the paper provides an answer to the first question and assures that the toolbox is very easy to install and use. Maybe the authors could support this claim by adding "snippets" of code that show the key steps in building an actual example.

3. The results provided in Figures 1, 4, 5 and 6 are useful, because they provide examples of the type of things one can do with the toolbox. I have a few comments that might help improving them:<br /> a. The examples in Figures 1 and 5 seem a bit redundant (same effector, similar task). Maybe the authors could show an example with a different effector or task? (see point 4).<br /> b. I missed a discussion on the relevance of the results shown in Figure 4. The moment arms are barely mentioned outside section 2.3. Are these results new? How can they help with motor control research?<br /> c. The results in Figure 6 are important, since one key asset of ANNs is that they provide access to the activity of the whole population of units that produces a given behavior. For this reason, I think it would be interesting to show the actual "empirical observations" that the results shown in Fig. 6 are replicating, hence allowing a direct comparison between the results obtained for biological and simulated neurons.

4. All examples in the paper use the arm26 plant as effector. Although the authors say that "users can easily declare their own custom-made effector and task objects if desired by subclassing the base Plant and Task class, respectively", this does not sound straightforward. Table 1 does not really clarify how to do it. Maybe an example that shows the actual code (see point 2) that creates a new plant (e.g. the 3-joint arm in Figure 7) would be useful.

5. One potential limitation of the toolbox is that it is based on Tensorflow, when the field of Computational Neuroscience seems to be, or at least that's my impression, transitioning to pyTorch. How easy would it be to translate MotorNet to pyTorch? Maybe the authors could comment on this in the discussion.

6. Supervised learning (SL) is widely used in Systems Neuroscience, especially because it is faster than reinforcement learning (RL). Thus providing the possibility of training the ANNs with SL is an important asset of the toolbox. However, SL is not always ideal, especially when the optimal strategy is not known or when there are different alternative strategies and we want to know which is the one preferred by the subject. For instance, would it be possible to implement a setup in which the ANN has to choose between 2 different paths to reach a target? (e.g. Kaufman et al. 2015 eLife). In such a scenario, RL seems to be a more natural option Would it be easy to extend MotorNet so it allows training with RL? Maybe the authors could comment on this in the discussion.

Impact:<br /> MotorNet aims at simplifying the process of simulating complex experimental setups to rapidly test hypotheses about how the brain produces a specific movement. By providing an end-to-end pipeline to train ANNs on the simulated setup, it can greatly help guide experimenters to decide where to focus their experimental efforts.

Additional context:<br /> Being the main result a toolbox, the paper is complemented by a GitHub repository and a documentation webpage. Both the repository and the webpage are well organized and easy to navigate. The webpage walks the user through the installation of the toolbox and the building of the effectors and the ANNs.

Reviewer #2 (Public Review):

MotorNet aims to provide a unified interface where the trained RNN controller exists within the same TensorFlow environment as the end effectors being controlled. This architecture provides a much simpler interface for the researcher to develop and iterate through computational hypotheses. In addition, the authors have built a set of biomechanically realistic end effectors (e.g., an 2 joint arm model with realistic muscles) within TensorFlow that are fully differentiable.

MotorNet will prove a highly useful starting point for researchers interested in exploring the challenges of controlling movement with realistic muscle and joint dynamics. The architecture features a conveniently modular design and the inclusion of simpler arm models provides an approachable learning curve. Other state-of-the-art simulation engines offer realistic models of muscles and multi-joint arms and afford more complex object manipulation and contact dynamics than MotorNet. However, MotorNet's approach allows for direct optimization of the controller network via gradient descent rather than reinforcement learning, which is a compromise currently required when other simulation engines (as these engines' code cannot be differentiated through).

The paper could be reorganized to provide clearer signposts as to what role each section plays (e.g., that the explanation of the moment arms of different joint models serves to illustrate the complexity of realistic biomechanics, rather than a novel discovery/exposition of this manuscript). Also, if possible, it would be valuable if the authors could provide more insight into whether gradient descent finds qualitatively different solutions to RL or other non gradient-based methods. This would strengthen the argument that a fully differentiable plant is useful beyond improving training time / computational power required (although this is a sufficiently important rationale per se).

Reviewer #3 (Public Review):

Artificial neural networks have developed into a new research tool across various disciplines of neuroscience. However, specifically for studying neural control of movement it was extremely difficult to train those models, as they require not only simulating the neural network, but also the body parts one is interested in studying. The authors provide a solution to this problem which is built upon one of the main software packages used for deep learning (Tensorflow). This allows them to make use of state-of-the-art tools for training neural networks.

They show that their toolbox is able to (re-)produce several commonly studied experiments e.g., planar reaching with and without loads. The toolbox is described in sufficient detail to get an overview of the functionality and the current state of what can be done with it. Although the authors state that only a few lines of code can reproduce such an experiment, they unfortunately don't provide any source code to reproduce their results (nor is it given in the respective repository).

The modularity of the presented toolbox makes it easy to exchange or modify single parts of an experiment e.g., the task or the neural network used as a controller. Together with the open-source nature of the toolbox, this will facilitate sharing and reproducibility across research labs.

I can see how this paper can enable a whole set of new studies on neural control of movement and accelerate the turnover time for new ideas or hypotheses, as stated in the first paragraph of the Discussion section. Having such a low effort to run computational experiments will be definitely beneficial for the field of neural control of movement.

It's useful to have a deep learning method for taxonomic classification of species. However, for the method to be broadly useful across the scientific community the specific methods need to be accessible. Do you have a Github repository where you share the code for this work so that other scientists can try this method on their own biological samples?

Mistaken human assumptions that AI will generate what a human would given the same prompt are reinforced by claims by those selling AI tools that such tools "understand human language." We don't actually know that AI understands, just that it provides a result that we can interpret as understanding (with the help of our cognitive biases).

This claim to understanding is especially misleading for neural network-based AI. We don't know how neural networks think. With older Lisp based AI we could at least trace through the code to see how the AI thinks.

This is how I was taught to write C++ code

In other words, make the member variable itself private, but can be abstracted (and access provided) via public methods/properties

Doesn't this violate their own non-goal #6, "Provide additional runtime functionality", since it emits a this.x = x run-time side effect in the body that isn't explicitly written out in the source code?

Read-Eval-Print Loop and is a software programming environment that allows developers to interactively execute their code

Hello from Berkeley. Thank you for sharing this wonderful lecture. But I cannot run the code in Deepnote which keeps saying pydrake.all not installed. I duplicated it and ran it.

Reviewer #2 (Public Review):

This study explores the variability of cerebellar anatomy in the mammal. By capturing a set of anatomical measures in the cerebellum and including previously reported cerebral and cerebellar metrics in a set of 58 different mammalian species, this study depicts both consistency and heterogeneity in the co-occurrence of different brain features, with a focus on cerebellar structures such as folial wavelength or median depth of the molecular layer. This is very informative as the cerebellum is currently under-explored and the phylogenetic aspect of this work gives insights into evolutionary processes linked to the morphology of the cerebellum.

Strengths:

- The methods used to capture the different brain features are relevant, and include the reuse of previously reported metrics, which makes sense and valorises the previous work of other teams.<br /> - One interesting novel method to detect the depth of the molecular layer is implemented.<br /> - A generous amount of results are reported (including correlations, phylogenetic principal component analyses, ancestor character state estimation, and allometries), with visually effective figures to support them.<br /> - A remarkable effort has been made to make data and code available, which will be of great use to the community.

Weaknesses:

- The methods section does not address all the numerical methods used to make sense of the different brain metrics. In the results section, it sometimes makes it difficult for the reader to understand the reason for a sub-analysis and the interpretation of the numerical findings.<br /> - The originality of the article is not sufficiently brought forward:<br /> a) the novel method to detect the depth of the molecular layer is not contextualized in order to understand the shortcomings of previously-established methods. This prevents the reader from understanding its added value and hinders its potential re-use in further studies.<br /> b) The numerous results reported are not sufficiently addressed in the discussion for the reader to get a full grasp of their implications, hindering the clarity of the overall conclusion of the article.

Correct code / solutions (7 points) : 7 Correct plots (3 points) : 3 Clarity of code / document (3 points) : 1 Originality (7 points) : 2 Total (20 points) : 13

no. This is not from you and does not explain what the code does.

While referring principally to the hero of the Homeric epic, ‘Ulisse’ also represents the kind of moniker that could serve Italian partisans as a nom de guerre. In his 1981 poem Partigia, Primo Levi enquires into the fate of his companions in the Resistance: ‘Dove siete, partigia di tutte le valli, | Tarzan, Riccio, Sparviero, Saetta, Ulisse?’ Historian Sergio Luzzatto reports that Levi’s 1946 application for recognition as a partisan listed his own code name as 'Ferrero' (Luzzatto 2016, 165-66).

PB

Authors' Response (2 June 2023)

GENERAL ASSESSMENT

The objectives of the study: This paper aims to characterize the dynamics that drive allostery of the adenosine A1 receptor (A1R) via computational analysis of its activation free energy landscape and measurements of the appropriate geometrical parameters. This is done by focusing on the allosteric signaling pathways in different activation states, from inactive to active states via intermediate and pre-active ones, as well as the characterization of putative drug-binding pockets. The long-term objectives are to eventually be able to aid drug discovery efforts for this therapeutically important GPCR.

Key findings and major conclusions: Conventional MD does not enable the sampling of the complete conformational landscape of receptor activation. Instead, enhanced sampling MD simulations are required to achieve this. Using metadynamics, the authors decipher the activation pathway of A1R, decode the allosteric networks and identify transient pockets. The protein energy networks computed throughout the inactive, intermediate active, pre-active and active conformational states unravel the extra and intracellular allosteric centers and the communication pathways that couple them, whereby the pathways are reinforced in the activated state. These conformations primarily differ in the dynamics of the ionic lock motif that couples TM3 to TM6 in the inactive conformation and reveal that G-proteins are required to fully stabilize the active conformation. Support for these findings comes from prior mutagenesis work on the A1R that identified key allosteric residues that in many cases map to identified communication nodes. Finally, the authors identified allosteric pockets throughout the A1R in four different conformational states that support prior experimental and MD studies on the mechanism of the positive allosteric modulator MIPS521 and which could be targeted for the design of modulators. Overall, these findings provide complementary support to a structure-based mechanism of activation and allosteric modulation of A1R, and extend the findings to incorporate dynamics across the full activation pathway.

The perceived strengths and weaknesses: This preprint employs a combination of computational techniques to successfully reconstruct and analyze the conformational ensemble of the A1R activation. The metadynamics simulations supported the aim of the study, the results are clearly presented and the work is very well written. The authors could improve the discussion of how the protein energy network analysis could further advance rational design of specific modulators with a desired mode of action. The computational approach needs to be refined to be robust, with a focus on characterizing the convergence of the free energy landscapes. Overall, A1R is a good choice as the target for this study as there is existing structural and pharmacological data to support preliminary findings. Moreover, the framework presented herein could be adapted and scaled to other GPCRs with structural templates, which might enable comparison of allosteric pathways across families and classes.

We thank the reviewers for contextualizing the findings of this study and for highlighting that our work provides complementary support to the mechanism of activation and allosteric modulation of A1R. We also thank the reviewers for their comments and suggestions, which had a great contribution to improve the quality and significance of the revised version of the preprint.

RECOMMENDATIONS

Revisions essential for endorsement:

1. The paper could better demonstrate how the insights gained herein will or could lead to progress in the rational design of specific modulators with a desired effect. The authors should outline and discuss how they envision the modeling pipeline they have designed will be used towards this goal and tone-down or explain why "this information is essential to ease the design of allosteric modulators for A1R.". A recent study on FFAR1, where the authors targeted a specific dynamic pocket could be helpful in this respect (https://www.pnas.org/doi/full/10.1073/pnas.1811066116). Specifically, this might entail: How does specificity for a receptor correlate with pockets forming in a specific state? From this, how does one design an agonist vs. an antagonist vs. an inverse agonist? Does breaking a specific network select a function of the drug? How would another group follow up on this work, for example in a virtual screening campaign?

We agree with the reviewers that a more comprehensive discussion of the points they mention is highly relevant to the study. Firstly, we have rephrased the last sentence of the abstract as "This information can be useful to ease the design of allosteric modulators for A1R" to ensure the significance of our results is not overstated. However, to address the reviewer's feedback more thoroughly, we conducted additional simulations with a positive allosteric modulator (MIPS51) and added an additional sub-section in the results, which includes three new figures (Figure 7-8 and S12):

"ADO and MIPS51 PAM have a significant impact on the energy networks. In order to establish a connection between the energy networks and the mode of action of allosteric modulators, we focus on exploring the effect of MIPS521 positive allosteric modulator (PAM) and ADO agonist as a proof of concept. Experimental assays and Gaussian accelerated MD determined that MIPS521 PAM increases the binding affinity of ADO in the orthosteric site.[19] Thus, PB and PD must be allosterically coupled. Among MIPS521 PAM pocket residues, only L2426.43, L2456.46, S2466.47 and G2797.44 were experimentally found to affect the PAM cooperativity (Figure S11). Interestingly, the PEN obtained in presence of ADO captures these key residues along activation, including TM6 (L2426.43 and L2456.46) in the intermediate, L2426.43 and S2466.47 in the pre-active and L2426.43 and TM7(G2797.44) allosteric residues in the fully-active ensemble (Figure 7). Indeed, G2797.44 becomes a key node in the PEN of the fully-active ensemble. This evidence suggests that although both PD and PB are open in all conformational states, their energy coupling is particularly stronger during the receptor activation.

This prompted us to investigate whether the binding of ADO and the MIPS521 PAM can affect the allosteric communication between PB and PD sites. To that end, we performed cMD of the heterotrimeric Gi2 protein ADO-A1R-Gi2 complex in presence of the PAM (PAM-ADO-A1R-Gi2 complex) and in absence of adenosine (A1R-Gi2 complex) in order to compute their conformational landscape and energy networks following the same protocol for the ADO-A1R-Gi2 complex (Figure 8 and S12). The analysis of the PEN of A1R-Gi2 complex reveals that in the absence ADO, the receptor displays a reduced allosteric communication between PB and functional regions of the receptor, such as the extracellular allosteric center, TM6 and PD allosteric site. As expected, the presence of ADO restores the allosteric coupling between PB and TM6, which could explain the increase of receptor activity associated with agonist binding. Additionally, our analysis of the PAM-ADO-A1R-Gi2 complex shows that the PAM reinforces the TM7-ECL3-ECL2 allosteric pathway that couple PD with PB, and ECL2 now communicates to the intracellular region through TM5 (Figure 8). Notably, a recently published study reported that the orhosteric pocket contracts after ADO binding, as demonstrated by shortened distances of the so-called vestibular lid (defined as the sum of length of the triangle perimeters formed by E17045.51-Y2717.36-E17245.53 interacting residues) and the E17245.53-K26567 salt bridge.[48] Remarkably, the TM7-ECL3-ECL2 enhanced pathway by PAM effect contains the vestibular lid and the E17245.53-K26567 salt bridge residues (Figure 8). This suggest that PAM promotes the contraction of PB, leading to the stabilization of the ADO-bound state. Thus, the enhanced energy coupling between PB and PD may be responsible for the increase in the binding affinity of ADO in presence of the PAM, as observed experimentally.[19] This data indicates that allosteric modulators are able to enhance and redistribute the energy networks, which is likely attributed for their effects on the receptor activity."

The new insights gained from our additional simulations have significantly enriched the discussion on how the protein energy network analysis can contribute to the rational design of specific modulators with desired modes of action. In light of these finding, the last paragraph of the discussion has been rephrased as:

"As a proof of concept, we focus on the PD, which corresponds to the binding site of MIPS52, a positive allosteric modulator (PAM) that increases Adenosine binding affinity in the orthosteric site (PB). Although PD is open in all conformational states the communication between PB and PD is enhanced along activation capturing the allosteric residues that were found to affect its PAM from the intermediate to the fully active. Based on this observation, we hypothesize that drugs that bind pockets and interact with PEN residues, which progresses towards regions of the receptor where function can be altered, may potentially affect the receptor activity through allosteric effects. Additionally, the pocket where the drug binds must be open at least in the conformation state that is targeted. As a practical aspect, virtual screening campaigns could use this information during the design procedure by selecting drug candidates that perform stronger interactions with the PEN contained in the pockets.

To further support this hypothesis, we explored the allosteric effects of ADO and MIPS52 PAM on the PEN. Interestingly, we observed that ADO is crucial for the formation of the extracellular center and its connection with TM6 pathway. Furthermore, MIPS52 PAM reinforces the pathway that connects PB and PD pockets and redistribute other connections. These alterations in the PEN can be related with their mode of action. ADO may increase the activity of the receptor through its communication with TM6 and the PAM may increase ADO binding affinity though stronger energy coupling between PD and PB pockets. These findings imply that the mode of action of allosteric drugs could be predicted depending on how they redistribute the PEN."

Accordingly, the last paragraph of the conclusions has been rephrased as:

"As a proof of concept, Adenosine and a previously experimentally determined positive allosteric modulator were found to enhance and redistribute the energy networks of the receptor in a manner that is consistent with their respective functions. The prediction of drug effects depending on how they redistribute the protein energy networks presents a promising avenue for drug discovery. All these system-specific structural dynamics understanding provides useful information to advance the design of A1R allosteric modulators on the basis of structure-based drug design. This computational approach can be also transferable to other GPCRs and related receptors, which is of interest for the design of novel allosteric drugs."

1. Free energy calculations:

a. A proof of convergence of the free energy calculations is missing. The authors argue that obtaining landscapes that do not change over time is proof of convergence, but this is incorrect in well-tempered metadynamics. The fact that the heights of the Gaussians decrease over time guarantees that the landscape will be stable over time, and the way to check convergence is to show that the collective variables become diffusive after convergence. In addition, to validate that the choice of collective variables (CV) is actually appropriate, they should check that CVs that were not biased are also diffusive. This would be best studied by looking at the behavior of microswitches that were not considered, such as ones describing the PIF motif, the NPxxY motif, the ligand binding pose, etc.

The goal of the well-tempered approach [Phys. Rev. Lett. 2008, 100, 020603] is to improve the convergence of the energy landscapes. This is achieved by gradually decreasing the height of gaussians over simulation. In this fashion, the height of the Gaussians is proportional to a decaying exponential function of the potential deposited in the currently visited point of the CV space. This technique has the added benefit of constraining reconstruction to the region of interest, reducing the risk of irreversible movement towards physically irrelevant regions of the CV space. As noted in the plumed tutorial (https://www.plumed.org/doc-v2.7/user-doc/html/master-_i_s_d_d-2.html), the fact that the Gaussian height is gradually decreasing should not be used as a measure of convergence. Rather, convergence can be assessed by monitoring the energy differences between chosen regions of the energy landscape over time (e.g. the inactive, intermediate and pre-active local energy minima used in our work). If this energy differences do not change significantly as a function of time, this can be taken as an indicator of convergence. We also would like to emphasize that our aim is to recover the major conformational states involved in the pathway of receptor activation rather than the study of subtle energy barriers and relative stability differences of the energy minima upon system perturbations. This objective has been made clear in the text.

We agree with the reviewers' comment that our measure of the convergence could be strengthened by additional analysis that verify the computed conformational pathway of receptor activation.

As suggested by the reviewers, we have plotted the CV1 values over time to asses convergence of our simulations (Figure S4). However, it should be noted that in this study, we have employed the walkers approach [J. Phys. Chem. B 2006, 110, 3533], which utilizes 10 replicas (walkers) to parallelize free energy reconstruction. Each walker simultaneously reconstructs the energy landscape by reading the Gaussian potentials deposited by the other walkers. Consequently, the correct time to stop the metadynamics simulation using the walkers approach becomes more problematic. To facilitate efficient sampling of the CV space and achieve convergence more rapidly, we utilized a sampling strategy that involved starting the simulation with walkers that spanned the entire CV space of interest. In this context, the fact that the walkers do not become trapped in the initial CV space and are able explore and cross into regions occupied by other walkers may serve as a useful indicator of convergence. This assessment of the convergence has been implemented before for Tryptophan Synthase, in which the resulting energy landscapes were consistent with experimental data. [J. Am. Chem. Soc. 2019, 141, 13049] and [ACS Catal. 2021, 11, 13733]

We have added the following paragraph in the convergence section including Figure S4:

"We have also assessed convergence by analyzing the CV1 values over simulation time. Figure S4A shows that during the first 100ns, walkers primary oscillates around their initial CV1 values. Subsequently, at around 200 ns walkers exhibit a higher frequency of crossing into regions occupied by other walkers. This is further supported by the exploration of W1 and W10, as shown in Figure S4B. These two walkers initially start the landscape reconstruction at the opposite extremes of the CV space. At 120 ns, they are able to escape from their respective basins and approach each other, sampling similar CV values (at approximately 240 ns). At this point of the simulation, only these two walkers have covered the entire conformational space of activation. Subsequently, they tend to return to previously sampled CV space. The observation that walkers do not become trapped in their initial CVs region, but instead explore and cross into other regions suggests that our sampling strategy, which involved starting the simulations with walkers that spanned the entire CV space of interest, has facilitated the exploration of the relevant conformational space. Although we cannot guarantee full convergence of the free energy landscape under these conditions, we successfully reconstructed the major conformational states of the receptor activation at 250 ns."

Accordingly, in the results section we have replaced "After 250 ns of accumulated time the FEL was considered to be converged (see Figure S3 and S4)" by "After 250 ns of accumulated time, we successfully reconstructed the major conformational states of the FEL (see convergence assessment in Figure S3 and S4)."

To further verify the accuracy of the collected conformational landscape, we have conducted additional analysis, which include the following:

"As a complementary analysis, we conducted the reweighting of the metadynamics simulations[28] to determine the free energy as a function of previously identified A1R micro-switches (ionic-lock, PIF motif, water-lock and toggle switch). The fact that we capture the distinct energy barriers associated with unbiased micro-switches highlights the accuracy of the metadynamics simulations in reproducing the pathway of activation and provides useful information to guide the selection of collective variables for future GPCR landscape calculations (Figure S5, S6 and S7)."

b. The authors should characterize the uncertainties/statistical errors on the measured free energy profiles to better evaluate the significance of change (e.g. for inspiration: https://www.plumed.org/doc-v2.7/user-doc/html/masterclass-21-2.html).

In response to the reviewers' comment, we have included a new sub-section in materials and methods together with an additional figure in the Supplementary Information (Figure S13), as follows:

"Error: We estimated the error on the 2D free energy landscape of the first collective variable (CV1), which is the TM3-TM6 intracellular ends distance (Figure S13) using the block averaging technique, as described in the PLUMED tutorial on calculating error bars (https://www.plumed.org/doc-v2.8/user-doc/html/lugano-4.html). We calculated the weights using the metadynamics bias potential obtained at the end of the simulation, and assuming a constant bias during the entire course of the simulation.[28] Specifically, we calculate the error using blocks of histograms of 25 ns each, covering the entire 250 ns simulation time."

c. In the cMD trajectories, a large part of phase space is sampled, which does not appear consistent with what one would expect based on the free energy landscapes. For instance, it does not seem reasonable to cover an almost complete conformational transition in 500ns when the barrier of the system is on the order of 5-8kcal/mol. The definition of CVs may have led to an overestimation of the free energy barrier. Hence an independent validation of the free energy barrier height is needed, by e.g. changing the CV definition.

We agree with the reviewers' that the cMD simulations cover a large part of the phase space. This is in part due to running simulations staring from both the inactive and active structures. For the latest, we removed the G-proteins from the receptor. This situation increases the flexibility of the receptor and induces a population shift respect to the starting point. However, as expected, we did not observe any complete transitions in our cMD simulations either staring from the inactive or active structures (see Figure 1C and S2).

Regarding the activation energy barrier, we would like to clarify that the aim of our study is not to compare subtle differences in energy barriers after system perturbations or compare them with experimental data. As far as we know, there is not NMR data available that confirms the exact time-scale of activation for A1R receptor, suggesting that A1R could be highly flexible. Notably, we report a rather low activation energy barrier of approximately 4 kcal/mol derived from the metadynamics simulations of A1R in the presence of adenosine. This is consistent with other computational studies of A1R where a complete transition from active to inactive [PNAS 2022, 119, E2203702119] and from inactive to pre-active [Nature 2021 597, 571] states is sampled in the course of nanosecond-scale Gaussian accelerated MD simulations. In addition, similar activation barrier values have been computed for the b2 adrenergic GPCR in the presence of adrenaline using different collective variables, as reported in [PLoS Comput. Biol. 2011, 7, e1002193] and [eLife 2021, 10, e60715]."

After careful consideration, we found relevant to validate our path of conformations by reweighing of our free energy into other collective variables. As previously stated, we have reweighted the original free energy into the ionic-lock, PIF motif, water-lock and toggle micro-switches (refer to the last paragraph of Essential revision 2.1 and Figure S5 and S6). Our analysis reveals that our original CV2 (TM6 torsion), the PIF motif, and the toggle display modest energy barriers, while our original CV1 (TM3-TM6 distance) presents a rather high energy barrier. Moreover, the ionic-lock and the water-lock exhibit the highest energy barriers. Thus, if we had chosen to use our initial CV1 and the PIF motif, we would have obtained a similar energy barrier, whereas choosing our initial CV1 and the water-lock would have resulted in a higher energy barrier, as predicted by our reweighting calculations (see Figure S7).

As mentioned in the manuscript, this data highlights the ability of our simulations to reproduce the activation pathway and provides interesting insights that can guide the selection of collective variables for future GPCR landscape calculations.

1. Configurations extracted from both conventional MD and wt-metadynamics are mixed in the analyses of the allosteric networks and the pockets. A more accurate way to integrate these datasets would be to modulate the weights of the configurations by their statistical weights, which can be retrieved from the metadynamics simulations.

We thank the reviewers for this suggestion and we will consider to use configurations by their statistical weights in future work. For this case, we aimed to include as much as configurations as possible from each conformational state. We then included all configurations from the inactive, intermediate and pre-active derived from the metadynamics and for the fully-active we applied a stride value in order to collect a similar number of structures.

1. Related to Figure S6, it is essential to compare the dynamics for all of the key class A activation motifs including the Na binding site, PIF motif, and NPxxY.

Based on the reviewers' comments, we have generated histograms for the relevant micro-switches corresponding to the inactive, intermediate, and pre-active states (see Figure S10). This analysis provides further support to the activation pathway derived from the metadynamics simulations. Notably, the population distributions of these micro-switches in the inactive, intermediate, and pre-active states exhibit a correlated progression that mirrors the receptor's activation pathway.

1. Please provide clarification on why 500 ns was chosen as the time-scale of the MD simulations and inclusion of the time course for the three independent MD simulations for each of the key structural features (e.g. TM6 torsion angles and TM3-TM6 distances).

To ensure adequate initial sampling of receptor activation, we performed three independent MD replicas of 500 ns each, starting from both the inactive and active structures. The purpose of these MD simulations is to provide an initial sampling of the receptor instead of describing the complete activation pathway. Based on this initial sampling, we selected a path of 10 conformations as starting points for the walker metadynamics simulations. We have added this information to the text for clarification.

"For each starting point we computed 3 replicas of 500ns, which is a reasonable simulation time to provide an initial sampling of the receptor activation."

1. The validation of the results in the form of previously published mutagenesis results does not appear completely convincing. Large parts of the protein are included in the allosteric network, making it likely that mutations in some of these residues will have an effect if mutated. In addition, the fact that mutations in ECL2 and ECL3 affect allostery is expected and does not constitute a good validation of the results. If no other results are included, we recommend that the language be toned down so as not to overstate the significance of the results.

Following the reviewers' comment, we have removed the following sentences from the results:"Among the PEN residues, we successfully captured most of the allosteric residues previously identified by mutagenesis studies, which highlights the reliability of the allosteric networks computed" and "The high predictive power of the PEN to identify allosteric residues highlights the reliability of the characterized allosteric pathways"

1. What is the justification for using an energy-based scoring for network analysis, given that a conventionally correlation-based approach has been used successfully in the field? The concern with an energy-based approach is that the interaction energy calculations do not consider the dielectric effect, i.e., if water molecules interfere with two interacting residues. Since the dynamic network is one of the critical aspects of this study, we believe the authors need to explore other tools such as the one implemented in VMD (https://www.ks.uiuc.edu/Research/vmd/plugins/networkview/) and compare the results.

We decided to perform protein energy networks analysis because our aim was to investigate how the networks change in different conformational states along activation. We selected this approach because energy networks can provide a more detailed insight into how communication within the protein changes during activation, as compared to cross-correlation networks, which are more suited to characterizing communication through correlated motions in the global ensemble. In order to compare both protein energy networks and correlation networks we performed the cross-correlation analysis in the global ensemble. Note that both approaches yield some similarities and provides complementary information.

We also would like to thank the reviewers for raising concerns about the methodology employed in our work.We acknowledge that gRINN, which we used to generate the pairwise residue mean interaction energy matrix, does not include water molecules and ligands during the matrix generation process. As a result, we are unable to capture communication pathways that involve water-mediated connections or interactions between ligands and residues. For example, in the pre-active ensemble, Y2005.58 and Y2887.53 from the NPxxY motif are connected by a hydrogen bond facilitated by a bridging water molecule (the son-called water-lock). However, such communication is not captured in our analysis due to the absence of water molecules (Figure 3). We highlight this major limitation in the text. Another example can be observed in the simulation of the PAM-ADO-A1R-Gi2 system, where communication between L242 and S246, two residues involved in the Positive Allosteric Modulator (PAM) binding site, is missing in the PEN (Figure 8). Since these residues must be connected through the PAM, our methodology cannot detect their communication. A promising tool to considered in future studies is webPSN v2.0 [Nucleic Acids Res. 2020, 48, W95], a protein structure network analysis that includes nucleic acids and more than 30,000 biologically relevant molecules and ions, which is highly advantageous to study the effect of drugs on the protein communication.

1. Provide generic residue numbers such as GPCRdb or Ballesteros Weinstein numbering for all mentioned residues in text and figures, as is standard for structural papers.

As the reviewers suggested, we have renumbered all residues mentioned in the text and figures according to the Ballesteros Weinstein numbering scheme.

Additional suggestions for the authors to consider:

1. For the PEN analysis it would be useful to digest these communication networks with respect to the established structural activation motifs of class A GPCRs (Na binding site, PIF, and NPxxY) that are present at the A1R.

In order to make the PEN analysis more digestive, we have revised the second paragraph of the "Energy Networks captures the dynamic allosteric pathways along A1R activation" results section. Specifically, we have highlighted the most relevant micro-switches captured in the PEN, with a particular focus on the ionic and water-lock switches, which are the most prominent for the protein communication.

1. It is unclear why the authors chose two largely correlated CVs (See comment 2c). In addition, the choice of CV is likely contributing to the distortion of S6, as displayed in Figure 1E. It has been shown that choosing a different CV set that describes the motion between states in a more distributed way is more likely to lead to a converged conformational ensemble. We suggest repeating the metadynamics simulations with a more distributed CV set that encompasses all of the microswitches in the receptor.

Regarding the concern raised by the reviewers about the distortion observed in the TM6 end, we want to clarify that it is not attributed to the selection of collective variables (CVs) since it is already explored in the initial conventional MD simulations (see Figure S2B, W5 structure). We selected two CVs that may seem largely correlated, but they actually describe different aspects of the TM6 inward-to-outward transition. The first CV, which measures the distance between the center of mass of TM3 and TM6, is more related to the dynamics of the ionic lock in the intracellular region. On the other hand, the second CV (TM6 torsion) is related to the forces sensed by the upper region of TM6, including the dynamics of the W2476.48 toggle switch. Therefore, we believe that the combination of these two CVs provides a comprehensive description of the TM6 transition.

It is also worth mentioning that a more distributed set of CVs may be beneficial to better reproduce the activation energy barriers of the receptor. In fact, as shown in the reweighting calculations of the metadynamic bias potential (see Figures S6 and S7), using the TM3-TM6 COM end distances and the Y2005.58-Y2887.53 distance (water-lock) as CVs appears to be a good choice for this purpose.

1. To support the vision on how the analysis of activation pathway, energy networks and transient pockets could be used "to ease the design of allosteric modulators for A1R" (last sentence of the abstract), it might be necessary to show that the combination of these methods can indeed be predictive for the binding and effect of known ligands. This might provide a first step towards establishing that molecules that bind to pockets "near allosteric networks" is a promising avenue for drug discovery.

This highly relevant point has been addressed in Essential revisions 1.

1. The specific TM3-TM6 residues should be specified in figures and text. Commonly used TM3-TM6 comparisons include the measured maximum distance between 2x46 to 6x37, which could be used here also (e.g. see https://docs.gpcrdb.org/structures.html#structure-descriptors).

The specific residues of TM3 and TM6 that were used in the analysis have been clearly specified in both the Materials and Methods section and the figure captions of Figure 1 and 4.

1. Even though the "A1R in complex with PSB36 (PDB 5N2S)" is an inactive structure, PSB36 is an agonist. Hence, the authors should consider using the DU172 antagonist-bound structure for comparison (PPDB 5UEN)

According to literature, PSB36 is selective antagonist for A1R. In fact, experimental data showed that PSB36 exhibit low inverse agonist activity [Chem. Med. Chem. 2006, 1, 891]. Although PDB 5N2S and PPDB 5UEN structures are almost identical, the bulkier DU172 ligand causes a displacement of TM2 in the extracellular region. Therefore, we chose to use the 5N2S structure in our study. However, we will consider using the 5UEN structure in future studies.

1. How does adenosine and MIPS521 binding impact the different conformational states and PEN.

This highly relevant point has been addressed in Essential Revisions 1. For a more detailed analysis, refer to Figure 8 and S12, which shows the impact of MIPS521 on the PEN and the conformational landscape, respectively.

1. It would be interesting to note how the findings from this study compare/contrast to a very recently published report by Li et al, PNAS, 2022 "The full activation mechanism of the adenosine A1 receptor revealed by GaMD and Su-GaMD simulations". Similarly with regards to the determination of allosteric binding pockets in this recent publication: "The pocketome of G-protein-coupled receptors reveals previously untargeted allosteric sites" (https://doi.org/10.1038/s41467-022-29609-6)

According to the reviewers' suggestion, we have compared our findings to those of a recently published study [PNAS 2022, 119, E2203702119], which explored the full activation mechanism of A1R using both su-GaMD and GaMD.

This published work serves to further confirm the combined activation mechanism that we observed for A1R in our study, which entails the formation of a pre-active state in the presence of Adenosine and the stabilization of a fully-active state in the presence of both Adenosine and G-proteins. Moreover, their study reports the pre-activation of the receptor from the inactive state within 150 ns of GaMD, indicating a rather low activation energy barrier of the receptor in presence of Adenosine. This is consistent with the approximately 4 kcal/mol activation energy barrier we calculated for A1R-ADO in our 250 ns metadynamic simulations.

We would also like to highlight an additional noteworthy point we have included in the results as: "Notably, a recently published study reported that the orthosteric pocket contracts after ADO binding, as demonstrated by shortened distances of the so-called vestibular lid (defined as the sum of length of the triangle perimeters formed by E17045.51-Y2717.36-E17245.53 interacting residues) and the E17245.53-K26567 salt bridge.[48] Remarkably, the TM7-ECL3-ECL2 enhanced pathway by PAM effect contains the vestibular lid and the E17245.53-K26567 salt bridge residues (Figure 8). This suggest that PAM promotes the contraction of PB, leading to the stabilization of the ADO-bound state. Thus, the enhanced energy coupling between PB and PD may be responsible for the increase in the binding affinity of ADO in presence of the PAM, as observed experimentally.[19]"

1. A major advantage of allosteric drugs is the potential to achieve higher selectivity. Expansion of this study to include other adenosine receptor subtypes or linking to other types of molecular pharmacology (e.g. biased signalling, subtype selectivity, etc.) would be a major benefit to the field.

We are grateful to the reviewers for recognizing the potential impact of our work in various applications beyond our initial scope. We will consider to incorporate their valuable suggestions in our future research endeavors.

1. Consider including an explanation of the physiological and pharmacological relevance of A1AR in the introduction.

According to this suggestion, we have incorporated a new sentence in the introduction section as follows:

"The adenosine A1 receptor (A1R) is a member of the class A G protein-coupled receptor (GPCR) family that preferentially couples with Gi/o proteins. It is widely distributed in multiple organs mediating a variety of physiological processes, including those in the brain and the heart. Thus, A1R has significant therapeutic potential in the treatment of numerous diseases and disorders.[18]"

1. Even if not entirely necessary for the results, it would be more consistent if the study would include metadynamics of the G-protein bound state.

Performing a metadynamics calculation of the G-protein bound state is challenging as it requires careful consideration of the G-protein binding process. As a reference work, Giulio Mattedi et al. successfully implemented this calculation for the glucagon receptor [Proc. Natl. Acad. Sci USA, 2020, 117, 15414]. However, in our study the effect of the G-proteins in the activation landscape is a minor remark. Our study places a greater emphasis on sampling the most stable conformations associated with the fully-active conformational state to compute the protein energy networks.

1. Methods: "In other words, once the free energy surface does not change significantly during a relatively long period of time in the last part of the simulation". What is "relatively long period of time" and "change significantly". The convergence, should be stated as a quantitative description of the observed energy differences.

We have addressed this technical issue in Essential revisions 2. It is worth noting that "the relatively long period of time" required for convergence may vary depending on the specific system under study. Nonetheless, we believe that observing a stable energy surface over the course of 50-100 ns, while the system explores different relevant regions of the CV space, provides a good criterion for convergence."

1. The authors should strongly consider making their analysis code and simulation data publicly available (e.g. on GitHub or Zenodo) so that others can replicate and build upon this work.

We thank the reviewers for this suggestion. We will make all the output files generated from the get Residue Interaction eNergies and Networks (gRINN) calculation available to the public.By doing so, users will be able to visualize the results in the gRINN visual interface and perform customized network analysis using the pairwise residue mean interaction energy matrices. Additionally, we will provide all Pymol sessions that include the protein energy networks as well as the Isosurface representation of the frequency maps of the transient pockets. We believe that these materials will provide better visualization compared to the current figures presented in the manuscript and supporting information, which will be helpful in guiding future structure-based drug design campaigns.

(This is a response to peer review conducted by Biophysics Colab on version 3 of this preprint.)

Consolidated peer review report (28 November 2022)

GENERAL ASSESSMENT

The objectives of the study:

This paper aims to characterize the dynamics that drive allostery of the adenosine A1 receptor (A1R) via computational analysis of its activation free energy landscape and measurements of the appropriate geometrical parameters. This is done by focusing on the allosteric signaling pathways in different activation states, from inactive to active states via intermediate and pre-active ones, as well as the characterization of putative drug-binding pockets. The long-term objectives are to eventually be able to aid drug discovery efforts for this therapeutically important GPCR.

Key findings and major conclusions:

Conventional MD does not enable the sampling of the complete conformational landscape of receptor activation. Instead, enhanced sampling MD simulations are required to achieve this. Using metadynamics, the authors decipher the activation pathway of A1R, decode the allosteric networks and identify transient pockets. The protein energy networks computed throughout the inactive, intermediate active, pre-active and active conformational states unravel the extra and intracellular allosteric centers and the communication pathways that couple them, whereby the pathways are reinforced in the activated state. These conformations primarily differ in the dynamics of the ionic lock motif that couples TM3 to TM6 in the inactive conformation and reveal that G-proteins are required to fully stabilize the active conformation. Support for these findings comes from prior mutagenesis work on the A1R that identified key allosteric residues that in many cases map to identified communication nodes. Finally, the authors identified allosteric pockets throughout the A1R in four different conformational states that support prior experimental and MD studies on the mechanism of the positive allosteric modulator MIPS521 and which could be targeted for the design of modulators. Overall, these findings provide complementary support to a structure-based mechanism of activation and allosteric modulation of A1R, and extend the findings to incorporate dynamics across the full activation pathway.

The perceived strengths and weaknesses:

This preprint employs a combination of computational techniques to successfully reconstruct and analyze the conformational ensemble of the A1R activation. The metadynamics simulations supported the aim of the study, the results are clearly presented and the work is very well written. The authors could improve the discussion of how the protein energy network analysis could further advance rational design of specific modulators with a desired mode of action. The computational approach needs to be refined to be robust, with a focus on characterizing the convergence of the free energy landscapes. Overall, A1R is a good choice as the target for this study as there is existing structural and pharmacological data to support preliminary findings. Moreover, the framework presented herein could be adapted and scaled to other GPCRs with structural templates, which might enable comparison of allosteric pathways across families and classes.

RECOMMENDATIONS

Revisions essential for endorsement:

1.    The paper could better demonstrate how the insights gained herein will or could lead to progress in the rational design of specific modulators with a desired effect. The authors should outline and discuss how they envision the modeling pipeline they have designed will be used towards this goal and tone-down or explain why “this information is essential to ease the design of allosteric modulators for A1R.”. A recent study on FFAR1, where the authors targeted a specific dynamic pocket could be helpful in this respect (https://www.pnas.org/doi/full/10.1073/pnas.1811066116). Specifically this might entail: How does specificity for a receptor correlate with pockets forming in a specific state? From this, how does one design an agonist vs. an antagonist vs. an inverse agonist? Does breaking a specific network select a function of the drug? How would another group follow up on this work, for example in a virtual screening campaign?

2.    Free energy calculations:

a.    A proof of convergence of the free energy calculations is missing. The authors argue that obtaining landscapes that do not change over time is proof of convergence, but this is incorrect in well-tempered metadynamics. The fact that the heights of the Gaussians decrease over time guarantees that the landscape will be stable over time, and the way to check convergence is to show that the collective variables become diffusive after convergence. In addition, to validate that the choice of collective variables (CV) is actually appropriate, they should check that CVs that were not biased are also diffusive. This would be best studied by looking at the behavior of microswitches that were not considered, such as ones describing the PIF motif, the NPxxY motif, the ligand binding pose, etc.

b.    The authors should characterize the uncertainties/statistical errors on the measured free energy profiles to better evaluate the significance of change (e.g. for inspiration: https://www.plumed.org/doc-v2.7/user-doc/html/masterclass-21-2.html).

c.    In the cMD trajectories, a large part of phase space is sampled, which does not appear consistent with what one would expect based on the free energy landscapes. For instance, it does not seem reasonable to cover an almost complete conformational transition in 500ns when the barrier of the system is on the order of 5-8kcal/mol. The definition of CVs may have led to an overestimation of the free energy barrier. Hence an independent validation of the free energy barrier height is needed, by e.g. changing the CV definition.

3.    Configurations extracted from both conventional MD and wt-metadynamics are mixed in the analyses of the allosteric networks and the pockets. A more accurate way to integrate these datasets would be to modulate the weights of the configurations by their statistical weights, which can be retrieved from the metadynamics simulations.

4.    Related to Figure S6, it is essential to compare the dynamics for all of the key class A activation motifs including the Na binding site, PIF motif, and NPxxY.

5.    Please provide clarification on why 500 ns was chosen as the time-scale of the MD simulations and inclusion of the time course for the three independent MD simulations for each of the key structural features (e.g. TM6 torsion angles and TM3-TM6 distances).

6.    The validation of the results in the form of previously published mutagenesis results does not appear completely convincing. Large parts of the protein are included in the allosteric network, making it likely that mutations in some of these residues will have an effect if mutated. In addition, the fact that mutations in ECL2 and ECL3 affect allostery is expected and does not constitute a good validation of the results. If no other results are included, we recommend that the language be toned down so as not to overstate the significance of the results.

7.    What is the justification for using an energy-based scoring for network analysis, given that a conventionally correlation-based approach has been used successfully in the field? The concern with an energy-based approach is that the interaction energy calculations do not consider the dielectric effect, i.e., if water molecules interfere with two interacting residues. Since the dynamic network is one of the critical aspects of this study, we believe the authors need to explore other tools such as the one implemented in VMD (https://www.ks.uiuc.edu/Research/vmd/plugins/networkview/) and compare the results.

8.    Provide generic residue numbers such as GPCRdb or Ballesteros Weinstein numbering for all mentioned residues in text and figures, as is standard for structural papers.

Additional suggestions for the authors to consider:

1. For the PEN analysis it would be useful to digest these communication networks with respect to the established structural activation motifs of class A GPCRs (Na binding site, PIF, and NPxxY) that are present at the A1R.

2. It is unclear why the authors chose two largely correlated CVs (See comment 2c). In addition, the choice of CV is likely contributing to the distortion of S6, as displayed in Figure 1E. It has been shown that choosing a different CV set that describes the motion between states in a more distributed way is more likely to lead to a converged conformational ensemble. We suggest repeating the metadynamics simulations with a more distributed CV set that encompasses all of the microswitches in the receptor.

3. To support the vision on how the analysis of activation pathway, energy networks and transient pockets could be used “to ease the design of allosteric modulators for A1R” (last sentence of the abstract), it might be necessary to show that the combination of these methods can indeed be predictive for the binding and effect of known ligands. This might provide a first step towards establishing that molecules that bind to pockets “near allosteric networks” is a promising avenue for drug discovery.

4. The specific TM3-TM6 residues should be specified in figures and text. Commonly used TM3-TM6 comparisons include the measured maximum distance between 2x46 to 6x37, which could be used here also (e.g. see https://docs.gpcrdb.org/structures.html#structure-descriptors).

5. Even though the "A1R in complex with PSB36 (PDB 5N2S)" is an inactive structure, PSB36 is an agonist. Hence, the authors should consider using the DU172 antagonist-bound structure for comparison (PPDB 5UEN)

6. How does adenosine and MIPS521 binding impact the different conformational states and PEN.

7. It would be interesting to note how the findings from this study compare/contrast to a very recently published report by Li et al, PNAS, 2022 “The full activation mechanism of the adenosine A1 receptor revealed by GaMD and Su-GaMD simulations”. Similarly with regards to the determination of allosteric binding pockets in this recent publication: “The pocketome of G-protein-coupled receptors reveals previously untargeted allosteric sites” (https://doi.org/10.1038/s41467-022-29609-6)

8. A major advantage of allosteric drugs is the potential to achieve higher selectivity. Expansion of this study to include other adenosine receptor subtypes or linking to other types of molecular pharmacology (e.g. biased signalling, subtype selectivity, etc.) would be a major benefit to the field.

9. Consider including an explanation of the physiological and pharmacological relevance of A1AR in the introduction.

10. Even if not entirely necessary for the results, it would be more consistent if the study would include metadynamics of the G-protein bound state.

11. Methods: "In other words, once the free energy surface does not change significantly during a relatively long period of time in the last part of the simulation". What is “relatively long period of time” and “change significantly”. The convergence, should be stated as a quantitative description of the observed energy differences.

12. The authors should strongly consider making their analysis code and simulation data publicly available (e.g. on GitHub or Zenodo) so that others can replicate and build upon this work

REVIEWING TEAM

Reviewed by:

Antonios Kolocouris, Professor, Department of Medicinal Chemistry Faculty of Pharmacy National and Kapodistrian University of Athens, Greece:

CADD and computational biophysics on adenosine receptors

David Thal, Senior Research Officer, Monash University, Australia:

structural biology and molecular pharmacology of allosteric mechanisms underlying Class A GPCRs

SciLifeLab Journal Club, Stockholm, Sweden (see Appendix for members)

Curated by:

Alexander S. Hauser, Associate Professor, University of Copenhagen, Denmark

APPENDIX

SciLifeLab Journal Club:

Feedback was generated in a meeting of the journal club involving:

Lucie Delemotte (Journal Club oversight), Associate Professor of Biophysics, KTH Royal Institute of Technology, Sweden: modeling and enhanced sampling of GPCRs and other membrane proteins.

Olivia Andén, PhD student, Stockholm University: cryo-EM and functional characterization of membrane proteins.

Cathrine Bergh, PhD student, KTH Royal Institute of Technology: enhanced sampling simulations of membrane proteins.

Koushik Choudhury, PhD student, KTH Royal Institute of Technology: membrane protein modeling, enhanced sampling simulations.

John Cowgill, postdoctoral scholar, Stockholm University: cryo-EM and simulations of membrane proteins.

Chen Fan, postdoctoral scholar, Stockholm University: cryo-EM and simulations of membrane proteins.

Nandan Haloi, postdoctoral scholar, KTH Royal Institute of Technology: membrane protein modeling, free energy calculations, structure refinement in cryo-EM maps.

Rebecca J Howard, researcher, Stockholm University: membrane protein structure-function, allosteric modulation.

Marie Lycksell, PhD student, Stockholm University: structure and simulations of membrane proteins.

Antoni Marciniak, PhD student, KTH Royal Institute of Technology: enhanced sampling simulations of GPCRs and other membrane proteins.

Darko Mitrovic, PhD student, KTH Royal Institute of Technology: membrane protein modeling, enhanced sampling, machine learning.

Alex Payne, PhD student, Memorial Sloan Kettering Center for Cancer Research: membrane protein modeling, cryo-EM structure determination, drug discovery.

Urška Rovšnik, PhD student, Stockholm University: cryo-EM and functional characterization of membrane proteins.

Akshay Sridhar, postdoctoral scholar, KTH Royal Institute of Technology: membrane protein modeling, enhanced sampling simulations.

Amanda Dyrholm Stange, PhD student, Aarhus University: membrane protein modeling, enhanced sampling simulations.

(This consolidated report is a result of peer review conducted by Biophysics Colab on version 3 of this preprint. Minor corrections and presentational issues have been omitted for brevity.)

I am teaching CISC 3325 ("Information Security") in Fall 2023. "Hacking" is defined as "gaining an unauthorized access to a tool/website," but in the given scenario, the only difference between monitored websites and ones that aren't monitored (as long as the users can't change the code behind the original webpage) is that anyone can post any comments they won't, so people are, by definition, authorized to comment on websites. As such, the rates of hacking a monitored webpage or one that isn't monitored shouldn't be significantly different. Given so, should the word "hackers" remain in this article?

remove

for -features - IndyLab

• Granuarly addressable HTML content
• Archived electronic discussions
• published papers and source code
• a Topic Map of all DKR content
• ontology and a glossary of DKR

If people whould like to use a bookmarklet to reload a webpage using via.hypothes.is/ you could use the following code. It works simular to the bookmarklet that is found on this page.

javascript:location.href='https://via.hypothes.is/'+location.href

The webhook end-point must return a status code

Reviewer #1 (Public Review):

This manuscript by Mahlandt, et al. presents a significant advance in the manipulation of endothelial barriers with spatiotemporal precision, and in the use of optogenetics to manipulate cell signaling in vascular biology more generally. The authors establish the role of Rho-family GTPases in controlling the cytoskeletal-plasma membrane interface as it relates to endothelial barrier integrity and function, and adequately motivate the need for optogenetic tools for global and local signaling manipulation to study endothelial barriers.

Throughout the work, the optogenetic assays are conceptualized, described, and executed with exceptional attention to detail, particularly as it relates to potential confounding factors in data analysis and interpretation. Comparison across experimental setups in optogenetics is notoriously fraught, and the authors' control experiments and measurements to ensure equal light delivery and pathway activation levels across applications is very thorough. In demonstrating how these new opto-GEFs can be used to alter vascular barrier strength, the authors cleverly use fluorescent-labeled dextran polymers of different sizes and ECIS experiments to demonstrate the physiological relevance of BOEC monolayers to in vivo blood vessels. Of particular note, the resiliency of the system to multiple stimulation cycles and longer time course experiments is promising for use in vascular leakage studies.

Given that dozens of Rho GTPase-activating GEFs exist, expanded rationale for the selection of p63, ITSN1, and TIAM1 in the form of discussion and literature citations would be helpful to motivate their selection as protein effectors in the engineered tools. Extensive tool engineering studies demonstrate the superiority of iLID over optogenetic eMags or rapamycin-based chemogenetic tools for these purposes. However, as the utility of iLID and eMags has been demonstrated for manipulation of a variety of signaling pathways, the iSH-Akt demonstration does not seem necessary for these systems.

The demonstration of orthogonality in GTPase- and VE-cadherin-blocking antibody-mediated barrier function decreases and is compelling, even without full elucidation of the role of cell size or overlap in barrier strength. The discussion section presents a mature and thoughtful description of the limitations, remaining questions, and potential opportunities for the tools and technology developed in this work. Importantly, this manuscript demonstrates a commitment to scientific transparency in the ways in which the data are visualized, the methods descriptions, and the reagent and code sharing it presents, allowing others to utilize these tools to their full potential.

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

Learn more at Review Commons

Referee #1

Evidence, reproducibility and clarity

Summary:

In zebrafish embryos, progenitor cells for both the prechordal plate and anterior endoderm reside at the dorsal margin in early gastrulation. Both cell populations are induced via signaling through the Nodal signaling pathway, however the mechanisms that send Nodal-exposed cells to one fate versus the other remain a matter of debate. Cheng et al use single-cell RNA sequencing to investigate the mechanistic origins of this developmental decision. They argue that both populations emerge from a common progenitor pool marked by the prechordal-plate marker gene goosecoid (gsc). By adding single-cell ATACseq analysis, they go on to argue that Nodal signaling encourages open chromatin states at target genes, and that this may underly the distinction between prechordal plate and endodermal fates. Finally, they suggest two potential regulators (gsc and ripply1) that may repress commitment to the endodermal fate.

Major Comments:

1. In lines 128-136, the authors describe a live imaging experiment to support the argument that anterior endodermal cells emerge from a gsc+ progenitor pool. The claim is that sox17+ cells (marked by RFP fluorescence) arise in gsc+ cells (marked by GFP fluorescence). From the presented data, I find it very hard to evaluate this claim. The GFP signal appears quite close to background in the highlighted cell. Additionally, the argument- as presented-turns on the behavior of a single highlighted cell. I think that this analysis should be clarified and extended to support the claim.

I suggest that the authors (1) plot average cell fluorescence over time rather than a 'line scan' across the cell, (2) draw cell borders from the mask used in each frame for clarity of presentation, and (3) plot the trajectories of gsc+/sox17+, gsc-/sox17- and gsc+/sox17-cells for comparison.

Alternatively, it could be helpful to extract fluorescence intensities for each cell in the field of view and scatter the RFP vs. GFP intensity for each cell. If the claim is true, three distinct subpopulations should be visible (i.e. gsc+/sox17-, gsc-/sox17- and gsc+/sox17+). Statistical analysis supporting the significance of these differences (e.g. comparing the means of each reporter within the populations) would be clarifying.

OPTIONAL: The live imaging experiment the authors present is quite ambitious, but perhaps overly difficult for the task at hand. I think this point could be more easily and clearly demonstrated by using two-color fluorescent in situs or HCR staining for gsc and sox17. Using an endpoint measurement would allow for deeper sampling across multiple embryos, and would likely yield clearer signals for cell type quantifications.<br /> 2. In the same section, I suggest that the authors address the possibility that the sox17+ cells observed don't go on to become part of the anterior endoderm. I commend the authors experimental work to support their scRNA-Seq data, however observation of the expression of a reporter gene (injected on a plasmid) is not equivalent to demonstrating that those cells adopt a given fate in the end. Is it not possible that the sox17 expression is transient, and these cells revert to prechordal plate fate? This point would be sealed by a formal fate mapping study (e.g. photoconversion of sox17::kaede cells), but I don't think this is a necessary bar for publication.<br /> 3. In Figure 1 M, the explant data does not seem to clearly support the claim that higher Nodal signaling intensities favor prechordal plate over endoderm. It appears that, for the endodermal panel, 2/3 replicates for 6 pg and 10 pg injections resulted in no endodermal cells observed. Could the authors clarify how this reflects the certainty of the conclusion? No statistical analysis is indicated on this panel or the one below.<br /> 4. OPTIONAL: The analysis presented in Fig. 1M strikes me as rather indirect (i.e. deconvolution of bulk RNA-Seq data to infer cell population proportions), and not strongly compelling. I think a stronger support of this point would be to inject Nodal into embryos and measure positive cell counts for gsc and an endodermal marker (e.g. sox32 or sox17) via HCR or in situ hybridization. This would yield a direct measurement of the cell counts in question. I think this would greatly support the claim, but I don't think should be considered a requirement for publication.<br /> 5. In Fig. 2H, the authors analyze responses to ectopic Nodal gradients in order to corroborate the results of their LIANA analysis. This experiment is a welcome addition to the argument, but has weak points that should be addressed.<br /> - a. The description of image analysis procedures used to construct the quantification plots are inadequate. It seems likely that the nuclei were segmented from the DAPI images, but this was not clear from the methods section. The authors should completely describe the segmentation pipeline and include sample code in the supplementary material.<br /> - b. The methods section seems to suggest that the analysis was performed exclusively on maximum intensity projections. I think this procedure may make the data hard to interpret and should be modified/support with additional analysis. For example, there is no reason that, at any given position in the image, the brightest DAPI and pSmad2 channel pixels occur in the same plane. Segmentation boundaries may therefore not reliably match between channels in the maximum intensity projection. The segmentation should be performed using the full Z-stack images. This can be done using widely-available software packages (e.g. CellProfiler).<br /> - c. The fluorescence images in 2H (specifically for the pSmad2 channel) look like they may contain some artifacts that carry through into the quantification. Specifically, there appears to be substantial non-specific background (both hazy and punctate) in the lft1 mutant that may artificially elevate the quantified intensity. This is evident in the quantification as a larger 'offset' to which the gradient decays than in the other presented images. This may be another explanation for the observation that pSmad2 staining is stronger in this background. I suggest that the authors (a) present all fluorescence images from the dataset in the supplement to allow for visual inspection, and (b) estimate the effect of fluorescence background on their quantifications to ensure that this artifact is not the source of the claimed difference.<br /> 6. In lines 267-284 and Fig. 4 L, the authors make the argument that ripply1 acts as a cell-autonomous repressor of endodermal fate. I find the argument for the cell autonomous character of its function hard to follow. Specifically, the authors lean on the experiment in which a plasmid with a sox17 promoter-ripply1 construct is injected, resulting in a decrease in endodermal cell count. Could the authors elaborate on how this proves a cell autonomous effect? Is it not possible that ripply1 expressed from this construct induces a signal that influences neighboring cells?<br /> 7. The suggestion that prechordal plate fate is favored (over endodermal fate) by higher Nodal signaling levels is interesting. This claim is supported by the derivation of a 'Nodal score' from RNAseq data. However, I don't see where the score is defined in the Methods section or in the supplementary materials. If this was accidentally omitted (my apologies if I am just missing it), it should be added. Additionally, I found the description in the main text to be opaque, and the paper would benefit from a more intuitive/friendly explanation of this metric.

Additionally, could the authors comment on what they believe-in terms of Nodal signaling history for a given cell- this score represents? Does it correlate with integrated Nodal exposure? Nodal exposure duration? Peak Nodal exposure? Given the results of Sako et al-that Nodal exposure duration is a critical determinant of prechordal plate fate- it would be useful to know if the authors believe their Nodal score findings point toward a different mechanism.

Minor Comments:

1. Line 84: The authors refer to the prechordal plate cells being 'more mature' than endoderm. It is unclear what the claim is here; some elaboration would be helpful.
2. The fluorescence images in Fig. S2 are virtually invisible in the PDF. The images should be rescaled to make them visible.
3. Fig. 2H would be easier to make sense of if the image panels were labeled. Please indicate which color corresponds to which stain.

Significance

I believe that this study fills in some details on the process of anterior endoderm specification that will be of interest to specialists in zebrafish Nodal signaling. I believe that the strongest and most novel section is the combined scRNA-Seq/ATAC-Seq analysis. This dataset is likely to be of interest to researchers who want to dig into potential mechanisms for the separation anterior endoderm and prechordal plate. Further, the singling out of ripply1 as a potential regulator of endodermal specification is interesting, and I hope that the authors follow this promising lead in future work.

While this study does provide a useful single-cell view of the specification of anterior endoderm, I didn't feel that it came to a concrete conclusion about the mechanism of separation of the anterior endoderm and prechordal plate. A few interesting processes/players are suggested by the findings- for example, Nodal/Lefty signaling between the populations or ripply1 expression could tip the balance- but I don't believe these hypotheses were tested clearly. The authors correctly point out that models for Nodal-driven endoderm/mesoderm separation have recently emerged in the literature, however the findings presented here don't rule out either of these models or compellingly support an alternative. I don't believe that this should preclude publication, however I do think it will limit the reach of the paper. Experiments that more concretely test the possible mechanisms hinted at here- for example, studying the separation of the two lineages in ripply1 mutants- would strengthen the paper's reach.

My enthusiasm for the paper is also somewhat reduced by the fact that some key findings of the paper can be found in earlier work. Acknowledgement of this prior work in the relevant sections could be improved. Specifically:

1. The finding that anterior endoderm cells emerge from a gsc-expressing population in the dorsal margin was strongly suggested in the classic Warga et al paper on the origin of zebrafish endoderm. There, fate mapping experiments demonstrate that dorsal marginal cells (in the first two cell tiers) in the late blastula can go on to form both endoderm and mesoderm. This strongly implies that anterior endoderm cells emerge from a gsc+ population, given that these cells are firmly within the gsc expression domain. I also note that the scRNAseq data from Fig. 2 in Farrell et al directly demonstrates that some sox17+ endoderm cells express gsc in their developmental trajectory. The findings in this paper are a welcome confirmation of these earlier observations, however this context should be discussed.
2. The observation that squint and lefty single mutants (either lefty1 or lefty2) can alter the propensity to adopt endodermal or mesodermal fates has also been observed previously. See for example Fig.1 in Norris et al, Figs 3 and 4 in Rogers et al, or Fig.1 in Chen et al. Acknowledging some of these earlier findings would benefit the paper.

As a reviewer, I feel most qualified to comment on the embryological aspects of the presented work. While I am generally familiar with the single-cell genomics toolkit, I am not in a position to rigorously assess the technical merit of that side of this work. Accordingly, I have tried to restrict my comments to the embryology side.

References:

1. Warga, R.M. and Nüsslein-Volhard, C., 1999. Origin and development of the zebrafish endoderm. Development, 126(4), pp.827-838.
2. Farrell, J.A., Wang, Y., Riesenfeld, S.J., Shekhar, K., Regev, A. and Schier, A.F., 2018. Single-cell reconstruction of developmental trajectories during zebrafish embryogenesis. Science, 360(6392), p.eaar3131.
3. Norris, M.L., Pauli, A., Gagnon, J.A., Lord, N.D., Rogers, K.W., Mosimann, C., Zon, L.I. and Schier, A.F., 2017. Toddler signaling regulates mesodermal cell migration downstream of Nodal signaling. Elife, 6, p.e22626.
4. Rogers, K.W., Lord, N.D., Gagnon, J.A., Pauli, A., Zimmerman, S., Aksel, D.C., Reyon, D., Tsai, S.Q., Joung, J.K. and Schier, A.F., 2017. Nodal patterning without Lefty inhibitory feedback is functional but fragile. Elife, 6, p.e28785.
5. Chen, Y. and Schier, A.F., 2002. Lefty proteins are long-range inhibitors of squint-mediated nodal signaling. Current Biology, 12(24), pp.2124-2128.
6. Sako, K., Pradhan, S.J., Barone, V., Ingles-Prieto, A., Müller, P., Ruprecht, V., Čapek, D., Galande, S., Janovjak, H. and Heisenberg, C.P., 2016. Optogenetic control of nodal signaling reveals a temporal pattern of nodal signaling regulating cell fate specification during gastrulation. Cell reports, 16(3), pp.866-877.

Reviewer #2 (Public Review):

This paper addresses an important computational problem in learning and memory. Why do related memory representations sometimes become more similar to each other (integration) and sometimes more distinct (differentiation)? Classic supervised learning models predict that shared associations should cause memories to integrate, but these models have recently been challenged by empirical data showing that shared associations can sometimes cause differentiation. The authors have previously proposed that unsupervised learning may account for these unintuitive data. Here, they follow up on this idea by actually implementing an unsupervised neural network model that updates the connections between memories based on the amount of coactivity between them. The goal of the authors' paper is to assess whether such a model can account for recent empirical data at odds with supervised learning accounts. For each empirical finding they wish to explain, the authors built a neural network model with a very simple architecture (two inputs layers, one hidden layer, and one output layer) and with prewired stimulus representations and associations. On each trial, a stimulus is presented to the model, and inhibitory oscillations allow competing memories to pop up. Pre-specified u-shaped learning rules are used to update the weights in the model, such that low coactivity leaves model connections unchanged, moderate coactivity weakens connections, and high coactivity strengthens connections. In each of the three models, the authors manipulate stimulus similarity (following Chanales et al), shared vs distinct associations (following Favila et al), or learning strength (a stand in for blocked versus interleaved learning schedule; following Schlichting et al) and evaluate how the model representations evolve over trials.

As a proof of principle, the authors succeed in demonstrating that unsupervised learning with a simple u-shaped rule can produce qualitative results in line with the empirical reports. For instance, they show that pairing two stimuli with a common associate (as in Favila et al) can lead to *differentiation* of the model representations. Demonstrating these effects isn't trivial and a formal modeling framework for doing so is a valuable contribution. Overall, the authors do a good job of both formally describing their model and giving readers a high level sense of how their critical model components work, though there are some places where the robustness of the model to different parameter choices is unclear. In some cases, the authors are very clear about this (e.g. the fast learning rate required to observe differentiation). However, in other instances, the paper would be strengthened by a clearer reporting of the critical parameter ranges. For instance, it's clear from the manipulation of oscillation strength in the model of Schlichting et al that this parameter can dramatically change the direction of the results. The authors do report the oscillation strength parameter values that they used in the other two models, but it is not clear how sensitive these models are to small changes in this value. Similarly, it's not clear whether the 2/6 hidden layer overlap (only explicitly manipulated in the model of Chanales et al) is required for the other two models to work. Finally, though the u-shaped learning rule is essential to this framework, the paper does little formal investigation of this learning rule. It seems obvious that allowing the u-shape to collapse too much toward a horizontal line would reduce the model's ability to account for empirical results, but there may be other more interesting features of the learning rule parameterization that are essential for the model to function properly.

There are a few other points that may limit the model's ability to clearly map onto or make predictions about empirical data. The model(s) seems very keen to integrate and do so more completely than the available empirical data suggest. For instance, there is a complete collapse of representations in half of the simulations in the Chanales et al model and the blocked simulation in the Schlichting et al model also seems to produce nearly complete integration. Even if the Chanales et al paper had observed some modest behavioral attraction effects, this model would seem to over-predict integration. The author's somewhat implicitly acknowledge this when they discuss the difficulty of producing differentiation ("Practical Advice for Getting the Model to Show Differentiation") and not of producing integration, but don't address it head on. Second, the authors choice of strongly prewiring associations in the Chanales and Favila models makes it difficult to think about how their model maps onto experimental contexts where competition is presumably occurring while associations are only weakly learned. In the Chanales et al paper, for example, the object-face associations are not well learned in initial rounds of the color memory test. While the authors do justify their modeling choice and their reasons have merit, the manipulation of AX association strength in the Schlichting et al model also makes it clear that the association strength has a substantial effect on the model output. Given the effect of this manipulation, more clarity around this assumption for the other two models is needed.

Overall, this is strong and clearly described work that is likely to have a positive impact on computational and empirical work in learning and memory. While the authors have written about some of the ideas discussed in this paper previously, a fully implemented and openly available model is a clear advance that will benefit the field. It is not easy to translate a high-level description of a learning rule into a model that actually runs and behaves as expected. The fact that the authors have made all their code available makes it likely that other researchers will extend the model in numerous interesting ways, many of which the authors have discussed and highlighted in their paper.

The course code - course title is a good way to make a direct link to the course catalogue and the students' transcripts/HEAR report as codes and short titles always appear - so good for cross-referencing

Why are we talking about authorization code here?

It should be on the lines "the redirect uri passed in the request is incorrect."

is this error specific to only authorization code or can occur if other required params are not passed? If so then we should mention all the required params.