{"id":"d4c743c2-2ec6-4944-8771-55e4f221a8c8","arxiv_id":"2501.00278","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A data-driven biophysical model of the C. elegans premotor network reproduces observed switching between forward and reversal locomotion and suggests distinct roles for gap junctions and chemical synapses.","lead":"This paper builds a mathematical model of the 15 neurons that control forward and backward movement in the worm C. elegans, driven by signals from over 100 other neurons. It shows the model can mimic the worm's natural back-and-forth movement patterns and explores how different nerve connections shape behavior.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported agreement is in-sample: the same 22 datasets used to fit A_ij, d_i, beta, tau, and the 1.4x magnification factor are the datasets on which model-data resemblance is demonstrated, so the 'reproduces' claim is not yet independently supported.","rationale":"The reader's designated weakest assumption is the correctness of the regression-derived synaptic signs. That is a legitimate concern for mechanistic interpretation, but the in-sample validation problem is more directly load-bearing for the central claim. Even if every synaptic sign were exactly correct, the reported agreement would not establish predictive or reproductive power unless the comparison datasets were independent of the fitting datasets. Conversely, if synaptic signs are partly wrong but out-of-sample agreement holds, the model could still reproduce the observed dynamics, though its mechanistic conclusions might need adjustment. The paper does average parameters over 22 datasets and applies one general parameter set to all simulations, which mitigates per-dataset overfitting; however, averaging over datasets that include the validation datasets still constitutes in-sample evaluation. The manuscript never describes a train/test split or a leave-one-out analysis, and Figures 3 and 11 do not state which datasets were used in fitting. The absence of this control is what makes the strong phrasing of the abstract ('reproduces', 'correctly predicts') unsupported as written. The reader's rationale already mentions in-sample validation as a weakness, so there is partial agreement, but because the reader's formal 'weakest assumption' is the synaptic signs, I identify a different primary concern. The reader's CONDITIONAL verdict remains appropriate: the model is a plausible and interesting fit, but the central claims require out-of-sample validation before they can be accepted as stated.","tokens_in":18033,"tokens_out":4708,"duration_ms":53348,"concrete_test":"Perform a leave-one-out retraining: for each of the 22 fitting datasets, refit A_ij, d_i, and beta on the remaining 21 datasets, fixing tau and the magnification factor on the training subset, then simulate the held-out dataset using its own signal-neuron traces. Compare the held-out switching time series, dwell-time histograms, pairwise correlations, and activity PDFs to the reported in-sample matches. If the held-out results remain comparable (e.g., mean dwell times within the same 0.2-0.33 min range and correlation patterns unchanged), the central claim survives. If they degrade substantially, the claim should be weakened from 'reproduces/predicts' to 'fits the training data'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the model's resemblance to experimental data is a genuine property of the model, not a consequence of fitting it to those same data. The Methods state that A_ij and d_i are estimated by linear regression on 22 datasets from Ref. [1], that beta is chosen by a parameter sweep over the same datasets, and that tau and the magnification factor are fit by simulating the datasets shown in Figure 11. The validation in Figures 3 and 11 uses datasets from the same source, and the paper never states that any validation dataset was held out from the fitting procedure. Under the default reading, the fitting and validation sets overlap or coincide. Because the signal-neuron inputs in the simulations are also taken from the same recordings, and because Eq. 4 regresses the observed core-neuron derivatives onto those inputs, the fitted parameters are effectively chosen to reproduce the observed core traces; close agreement in an in-sample simulation is therefore expected to some degree and cannot by itself establish that the model captures the underlying dynamics. The abstract's stronger wording—'correctly predicts behavior such as dwelling versus roaming'—rests on datasets 2023-01-09-22 and 2023-01-09-15, which are not identified as held out. This is a missing reporting control rather than an internal inconsistency, but it is the weakest link in the central claim: if the agreement is only in-sample, the model has demonstrated a fit, not a reproduction or prediction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a data-driven biophysical network model of the C. elegans premotor circuit. A core group of 15 neurons is described by a driven dynamical system (Eq. 2) with intrinsic cubic dynamics fit to voltage-clamp data, gap-junction weights taken from the connectome, and synaptic weights and biases estimated by linear regression on 22 whole-brain calcium imaging datasets. The model is validated by comparing simulated premotor activity and the resulting forward/reversal behavioral sequence to the same type of recordings, and the authors report reproduction of stochastic switching, higher-order statistics, and dwelling versus roaming behavior. The model is then used to propose distinct roles for gap junctions versus synapses and to investigate how activating specific sensory neurons affects simulated behavior.","tokens_in":18301,"tokens_out":5090,"duration_ms":49552,"significance":"If the central claims are upheld, the model would be a notable contribution: a biophysical, minimally parameterized network model that produces complex, non-limit-cycle switching dynamics, with parameters anchored in connectomic and imaging data rather than tuned ad hoc. The paper also offers a clear mechanistic hypothesis about the division of labor between gap junctions and synapses, and it makes its code publicly available, which aids reproducibility. However, the lack of a demonstrable train/test split and the direct dependence of the behavioral 'predictions' on fitted parameters currently prevent the abstract's 'reproduces' and 'correctly predicts' claims from being fully supported. The modeling framework and analyses are promising, but the evidence presented is not yet sufficient to establish that the model captures the underlying dynamics independently of the fitting procedure.","major_comments":[{"comment":"The validation is in-sample. The same datasets used to estimate A_ij, d_i, beta, tau, and the 1.4x magnification factor are used in Figures 3 and 11 to demonstrate agreement with data. The Methods state that beta is chosen by a parameter sweep over the 22 datasets, that A and d are averages of regressions over these same 22 datasets, and that tau and the magnification factor are fit by simulating the datasets shown in Figure 11. Since the signal-neuron inputs in the simulations are also taken from the same recordings, the resemblance in Figures 3 and 11 is expected to some degree by construction. Please report explicitly which datasets (by filename) are used for fitting versus validation, and provide an out-of-sample evaluation (e.g., leave-one-dataset-out or a reserved test set) before claiming that the model reproduces the observed dynamics.","section":"Methods: Parameter fit; Premotor network simulation; Results: Model validation"},{"comment":"The sensory-neuron 'predictions' are direct readouts of the fitted synaptic weights, not independent predictions. The paper states that ASH and AWC have excitatory synapses with the reversal cluster (Fig. 2(b)) and then finds, unsurprisingly, that activating them promotes reversal; the ASE and OLQ results likewise follow from fitted weights and the connectome structure. Because A_ij are estimated by regression on the same whole-brain imaging data that are later used for validation, calling these results 'predictions' overstates their evidentiary value. The dwelling versus roaming demonstration uses datasets 2023-01-09-22 and 2023-01-09-15, which are not identified as held out from the fitting procedure; the authors should either show these are held out or qualify the abstract's 'correctly predicts behavior' claim.","section":"Results: Manipulation of sensory inputs; Abstract"},{"comment":"The mechanistic conclusions, including the distinct roles of gap junctions and synaptic connections, rest on the regression-derived signs of the synaptic weights. The Discussion documents that several sign assignments conflict with previous studies (e.g., RIM to AVA is inhibitory here and in Ref. [22] but excitatory in Ref. [5]). The paper asserts that the robustness of switching behavior suggests these discrepancies are not consequential, but no sensitivity analysis is provided. A perturbation analysis (e.g., flipping signs of a subset of fitted weights, or refitting with different sign priors) is needed to establish that the central switching result and the gap-junction/synapse dichotomy are robust to the sign uncertainty acknowledged in the manuscript.","section":"Discussion, Fig. 7; Results: Gap junctions versus synaptic connections"},{"comment":"The quantitative comparisons are informal. While the correlations in Fig. 3(c) look similar, there are no error bars or statistical tests for the correlations, probability distributions (Fig. 3(d)), or dwell-time distributions (Fig. 3(e)). The text notes that simulated dwell times are slightly longer, but no confidence intervals or effect sizes are given. To support the abstract's 'closely resembles' wording, the paper should quantify the agreement over datasets (e.g., summary statistics, permutation or bootstrap tests) rather than relying on visual comparison.","section":"Results: Model validation, Fig. 3"}],"minor_comments":[{"comment":"There are several typos and grammatical slips, e.g., 'C. elegan' for 'C. elegans' in the Introduction, and 'The whole-brain imaging data [19, 1] connects neuronal activity to behavior' with a subject-verb mismatch. A careful proofreading pass is recommended.","section":"Throughout"},{"comment":"The caption states 'Because the system is receiving partial signals we also magnify A by 1.4x. Increasing the amount of stimulus to the premotor neurons increases their average activity levels.' This is ambiguous: it should specify that the magnification factor is applied to the synaptic input term only and that this factor was selected to bring the simulated activity to the correct level, as described in the Methods.","section":"Figure 10 caption"},{"comment":"The criteria for selecting the 22 datasets are given, but the actual list of dataset filenames is not. To allow reproducibility and to make the train/test distinction transparent, a supplementary table listing the datasets used for fitting, and those used for validation (especially in Figures 3, 5, and 11), should be provided.","section":"Methods: Parameter fit"},{"comment":"The paper uses 'GCamp z-score' without defining the normalization precisely. It would be helpful to state whether the z-score is computed per neuron, per dataset, or globally, since this affects the interpretation of the state variables and the thresholds used for behavioral classification.","section":"Methods: Whole-brain imaging datasets"}],"recommendation":"major_revision","confidential_remarks":"This manuscript presents a useful modeling framework, and the in-sample validation issue is fixable with a proper train/test split and quantitative comparisons. I recommend requesting a major revision rather than rejecting, because the central concern is a missing reporting control and not an internal inconsistency. The 'prediction' language in the abstract and sensory-neuron sections should be softened until out-of-sample evidence is available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this before reading further: the paper's stated claims that the model 'reproduces' and 'correctly predicts' C. elegans premotor dynamics are not as strong as they look. The synaptic weights A_ij, biases d_i, the coupling scale beta, the timescale tau, and the 1.4x input magnification are all fit to the same 22 whole-brain datasets (Atanas et al.) on which the model's agreement is then demonstrated. The 'roaming vs. dwelling' example uses two of those same datasets. So the resemblance is partly a fit, not an independent reproduction. That is a missing reporting control rather than an internal inconsistency, but it is the load-bearing weakness of the central claim.\n\nWhat is actually new: a driven dynamical-systems architecture—a 15-neuron recurrent core forced by 112 signal neurons whose activity is taken directly from calcium imaging—is a reasonable way to bridge detailed connectome data and tractable dynamics. Fitting signed synaptic weights by linear regression on whole-brain recordings is a sensible strategy given that the connectome does not give signs, and the authors are transparent about the resulting uncertainty, even comparing their polarity estimates against several prior studies. The analysis distinguishing gap junctions from chemical synapses—gap junctions synchronize, synapses drive switching—is a useful mechanistic hypothesis. The paper ships code and model weights on GitHub, which is real support for reproducibility.\n\nThe soft spots beyond the in-sample issue: the sensory 'predictions' (ASH/AWC promote reversal, ASE/OLQ promote forward) are essentially readouts of the fitted weight signs, which the paper itself calls 'unsurprising.' The dwelling/roaming separation is also derived from the same driving inputs used in the simulation, so it doesn't function as an independent test. The quantitative comparisons are mostly qualitative—no error bars, no statistical test of goodness of fit. I would not call any of these fatal, because the modeling framework is still informative and the mechanistic conclusions are plausible, but the abstract overstates what has been established.\n\nWho should read it: people working on C. elegans neural circuits or data-driven network models will find the architecture and code valuable, and the discussion of synaptic sign ambiguity is honest. It deserves serious peer review, but a referee should insist on out-of-sample validation—say, holding out several datasets entirely, or at minimum clearly separating fitted from non-fitted quantities—and a toned-down abstract. If the authors can show the switching and statistics appear on held-out data, this becomes a much stronger paper. I would send it out, with major revision expected.","headline":"A promising driven-network model of the C. elegans premotor circuit whose central 'reproduces/predicts' claim is undermined by in-sample validation; the fitting and validation datasets are the same.","tokens_in":756,"tokens_out":841,"would_cite":false,"duration_ms":25403,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C20","37N25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a minimal biophysical model of the C. elegans premotor network, driven by real whole-brain calcium imaging, reproduces the worm's stochastic forward-reversal switching and predicts behavioral states like dwelling…","keywords":["C. elegans","premotor network","biophysical network model","stochastic switching","gap junctions","whole-brain calcium imaging","driven dynamical system","dwelling versus roaming"],"falsifier":"Use paired electrophysiology or optogenetic activation to measure the sign of a single identified premotor synapse, for example RIM to AVA. The model's fitted weights require RIM to AVA to be inhibitory, while a published gene-expression-based prediction classifies it as excitatory; observing excitation would show the polarity set is wrong. Separately, optogenetically activating OLQ should lengthen forward runs if the model is right; observing the known head-withdrawal or nose-touch avoidance instead would reject the model's forward-promoting assignment.","tokens_in":17735,"feed_emoji":"🧠","tokens_out":8466,"duration_ms":79558,"temperature":0.7,"pith_summary":"The paper claims that the seemingly random switching between forward crawling and reversal in C. elegans can be reproduced by a compact, data-driven biophysical model: 15 core premotor neurons, recurrently connected according to the connectome, driven by the recorded activity of 112 presynaptic \"signal\" neurons. Using whole-brain calcium imaging to fit synaptic weights and biases, and voltage-clamp data for intrinsic dynamics, the model produces forward/reversal cluster activity whose switching, pairwise correlations, activity distributions, and dwell times resemble real recordings. The point is to move beyond phenomenological Markov or limit-cycle descriptions toward a mechanistic account of how identified neurons cooperate. If the claim holds, the model offers an analyzable tool for explaining how gap junctions synchronize neural clusters, how synaptic input drives switches, and how sensory activity shapes behavior such as roaming versus dwelling.","feed_headline":"Model of worm brain reproduces random forward-reversal switches","feed_subtitle":"Biophysical model of 15 neurons reproduces worm's switches and predicts roaming vs dwelling.","key_machinery":"The central object is the driven dynamical system of Eq. (2): 15 core premotor neurons with bistable cubic intrinsic dynamics (low state at $x=-0.8$, high state at $x=1$), gap-junction coupling $W_{ij}$ read from the connectome, and rectified synaptic coupling $A_{ij}\\sigma(x_j)$ with $\\sigma=\\text{ReLU}$; the forcing comes from 112 signal-neuron activity traces taken directly from whole-brain imaging data. The load-bearing design decision is that the core group is small enough to be analyzed while the unmodeled presynaptic population is treated as external input, which lets the parameters $A_{ij}$ and $d_i$ be fit by linear regression rather than by tuning an entire connectome. The named mechanism is the division of labor: gap junctions equalize and synchronize cluster members, synaptic inputs drive the irregular high/low transitions, and the interaction of the two produces the observed stochastic switching.","core_discovery":"The central claim is that a driven dynamical system—a recurrent network of the 15 premotor interneurons most closely tied to forward and reversal locomotion, forced by time series of 112 signal neurons taken from whole-brain imaging—reproduces the main statistical features of the worm's premotor activity. Each core neuron's calcium level $x_i$ evolves by $\\tau \\dot{x}_i = f_i(x_i) + \\beta \\sum_j W_{ij}(x_j-x_i) + \\sum_j A_{ij}\\sigma(x_j)$, with $f_i$ a cubic giving low- and high-activity attractors, $W_{ij}$ gap-junction strengths from the connectome, and $A_{ij}$ signed synaptic weights estimated by linear regression over 22 imaging datasets ($\\beta=10$, $\\tau=0.2$, input magnification 1.4). With these parameters, simulated forward and reversal clusters switch between high and low states much as recorded neurons do, and the model's correlations, activity distributions, and dwell times match data. The paper further reports that gap junctions alone synchronize but do not switch, synapses alone switch but do not synchronize, and both are needed together; and that fitted synaptic signs make ASH and AWC reversal-promoting, ASE mildly forward-promoting, and OLQ strongly forward-promoting, consistent with their known roles in avoidance and chemotaxis.","pith_inferences":["The regression-derived synaptic polarities are model outputs, not established biology: the connectome does not record signs, and published studies disagree on several connections (e.g., RIM to AVA is inhibitory here but excitatory in a gene-expression-based prediction). If those signs are wrong, the specific mechanistic story may not transfer, even though the paper notes the switching behavior is ","The model's success suggests that the apparent randomness of C. elegans switching may be deterministic in origin—bistable recurrent dynamics driven by fluctuating inputs—rather than requiring intrinsic noise as in Markov-chain models.","The OLQ prediction is likely an oversimplification: real OLQ also mediates head withdrawal and nose-touch avoidance, so a single forward-promoting pathway may fail under optogenetic activation, pointing to parallel pathways the current core network omits.","Treating signal neurons as fixed inputs ignores core-to-signal feedback and the omitted PVC/DVA and motor-neuron connections; adding them could change switching statistics and is a natural next test of the driven-system approximation."],"forward_implications":["With one parameter set, the model reproduces forward/reversal switching across many imaging datasets, so the same circuit equations can serve as a common substrate for studying different behavioral episodes.","Removing either gap junctions or synapses in the model destroys the realistic switching, indicating that both connection types are jointly necessary for the emergent switching dynamics.","The model predicts that activating ASH or AWC sensory neurons lengthens and increases reversals, while activating ASE or OLQ lengthens forward runs, giving experimentally testable predictions about single-neuron effects on behavior.","Simulated locomotory paths built from the model's switching sequences distinguish dwelling-like from roaming-like datasets, suggesting behavioral state can be read from the forward/reversal cluster dynamics.","The driven core-network design can be exported to other circuits or organisms that have connectomic and large-scale imaging data, providing an analyzable middle ground between whole-connectome simulation and phenomenological models."],"supporting_citations":[{"why":"Supplies the whole-brain calcium imaging datasets used to fit synaptic weights and to provide signal-neuron input traces.","marker":"[1]"},{"why":"Provides the connectome: gap-junction weights and the presence or absence of synaptic connections that set the network architecture.","marker":"[47]"},{"why":"Supplies voltage-clamp measurements of C. elegans neuronal current-voltage relations used to define the intrinsic cubic dynamics.","marker":"[12]"},{"why":"Supplies the RMD steady-state I-V curve from which the cubic intrinsic dynamics are approximated.","marker":"[32]"},{"why":"Provides gene-expression-based synaptic polarity predictions that the paper compares against its regression-derived signs.","marker":"[5]"},{"why":"A Markov-chain model of stochastic switching that this model contrasts with as purely random.","marker":"[38]"},{"why":"A recurrent network model whose limit-cycle dynamics the paper contrasts with its irregular switching.","marker":"[22]"},{"why":"Whole-brain imaging showing global motor-command encoding, used to identify premotor neurons and motivate the model.","marker":"[19]"}],"fun_headline_variants":["Worm brain model reproduces forward-reversal switches","15-neuron model of worm premotor network matches real switching","Biophysical worm brain model predicts roaming vs dwelling behavior","Data-driven model of C. elegans premotor network recreates behavior","Worm locomotion switches explained by a 15-neuron biophysical model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's conclusions depend on the assumption that the synaptic strengths and signs estimated by regression from calcium imaging—which connections excite and which inhibit—are true of the real worm, even though the connectome records no signs and published studies disagree on several of them.","fun_headline_variants_meta":{"raw":{"variants":["Worm brain model reproduces forward-reversal switches","15-neuron model of worm premotor network matches real switching","Biophysical worm brain model predicts roaming vs dwelling behavior","Data-driven model of C. elegans premotor network recreates behavior","Worm locomotion switches explained by a 15-neuron biophysical model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1611,"prompt_tokens":1089,"completion_tokens":522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":705,"completion_tokens_details":{"reasoning_tokens":437}},"tokens_in":705,"tokens_out":522,"duration_ms":5486,"temperature":1.0,"reasoning_tokens":437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:54:08.033291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use paired electrophysiology or optogenetic activation to measure the sign of a single identified premotor synapse, for example RIM to AVA. The model's fitted weights require RIM to AVA to be inhibitory, while a published gene-expression-based prediction classifies it as excitatory; observing excitation would show the polarity set is wrong. Separately, optogenetically activating OLQ should lengthen forward runs if the model is right; observing the known head-withdrawal or nose-touch avoidance instead would reject the model's forward-promoting assignment.","supporting_citations":[{"cited_title":"Atanas, Jungsoo Kim, Ziyu Wang, Eric Bueno, McCoy Becker, Di Kang, Jungyeon Park, Talya S","cited_arxiv_id":null,"evidence_quote":"Supplies the whole-brain calcium imaging datasets used to fit synaptic weights and to provide signal-neuron input traces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the connectome: gap-junction weights and the presence or absence of synaptic connections that set the network architecture."},{"cited_title":"Goodman, David H","cited_arxiv_id":null,"evidence_quote":"Supplies voltage-clamp measurements of C. elegans neuronal current-voltage relations used to define the intrinsic cubic dynamics."},{"cited_title":"Biophysical modeling of C","cited_arxiv_id":null,"evidence_quote":"Supplies the RMD steady-state I-V curve from which the cubic intrinsic dynamics are approximated."},{"cited_title":"Fenyves, Gábor S","cited_arxiv_id":null,"evidence_quote":"Provides gene-expression-based synaptic polarity predictions that the paper compares against its regression-derived signs."},{"cited_title":"Roberts, Steven B","cited_arxiv_id":null,"evidence_quote":"A Markov-chain model of stochastic switching that this model contrasts with as purely random."},{"cited_title":"A recurrent neural network model of C","cited_arxiv_id":null,"evidence_quote":"A recurrent network model whose limit-cycle dynamics the paper contrasts with its irregular switching."},{"cited_title":"Kaplan, Tina Schrödel, Susanne Skora, Theodore H","cited_arxiv_id":null,"evidence_quote":"Whole-brain imaging showing global motor-command encoding, used to identify premotor neurons and motivate the model."}],"review_version":1}