{"id":"edc430ce-2545-4f2b-bdce-51f6d7ec5da2","arxiv_id":"2507.17700","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A GVP-based graph neural network predicts committor functions and rate constants directly from Cartesian coordinates for four molecular systems, with atom-level sensitivity analysis.","lead":"This paper uses an equivariant graph neural network to predict the probability that a molecular configuration will reach a target state before returning, using only atomic positions as input. The same architecture works for peptides, a chemical reaction, and a fast-folding protein, and it flags the atoms that dominate each transition.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is that Eq. 4 reweights a time-dependent WTM-eABF bias as if it were a static potential; the SI's own warning that bias CVs can shift learned sensitivities means the 'CV-free' claim is not yet supported.","rationale":"The reader's verdict is CONDITIONAL, and my analysis agrees. The NANMA comparison with shooting data and the Diels-Alder comparison with Ref. 17 provide real support that the architecture can represent committors, and the identical architecture across systems plus the lambda-robustness in the SI are additional positive signals. The weak point is not the GNN itself but the data-generation step: every training set comes from a biased simulation along user-selected CVs, and the only bridge to the unbiased committor is Eq. 4. That bridge is presupposed rather than demonstrated. The SI's warning in Section 4 is an explicit admission that bias CVs can alter learned sensitivities, which is exactly the failure mode that would invalidate the 'CV-free' and 'without prior assumptions' claims. A toy-system test with a known committor and a deliberately poor bias CV would settle the question directly. Until that validation is provided, CONDITIONAL acceptance is appropriate, so the reader's verdict does not change; the condition should name this reweighting validation explicitly.","tokens_in":20631,"tokens_out":7045,"duration_ms":84102,"concrete_test":"Run the identical qGNN pipeline on a low-dimensional model with an analytic committor (for example, a double well with a slow mode perpendicular to the naive coordinate), using WTM-eABF along a deliberately incomplete CV and along a complete CV set. Train with Eq. 2 and Eq. 4 exactly as in the paper, and compare both learned committor maps to the exact committor. If the incomplete-CV model differs from the exact committor by more than the statistical error of the complete-CV model, then the reweighting is not curing the time-dependent bias and the CV-free claim fails; if the two models agree within error, the quasi-static assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that qGNN learns the committor from Cartesian coordinates without hand-crafted CVs—requires that the loss in Eq. 2, evaluated on WTM-eABF trajectories, equals the unbiased variational loss. Equation 4 is a static-bias reweighting formula: it assumes delta-W is a fixed potential of mean force and that the biased ensemble is stationary. WTM-eABF, however, deposits a time-dependent bias throughout the simulation. The Methods section states only that the learning stage uses the part of the simulation corresponding to a 'converged free energy,' but no convergence diagnostic is given, and no check shows that delta-W varies negligibly over the lag tau or over the training segment. If the bias is not effectively static, the reweighted two-time correlation is not the unbiased C[q; tau], and the minimized loss can return a committor of the biased dynamics rather than the true one. Section 4 of the SI explicitly cautions that adding CV-dependent biases can influence the model's predicted degrees of freedom; trialanine and Trp-cage have no reference committor, and all four systems were trained on data generated with hand-picked bias CVs (RMSDs, bond distances). A demonstration that the learned q is invariant to the choice of bias CVs is therefore necessary to support the abstract's CV-free claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces qGNN, a graph neural network based on geometric vector perceptrons, trained to approximate the committor function directly from Cartesian atomic coordinates. The training loss is the committor variational principle of Eq. (2), with a boundary penalty and a reweighting of biased trajectories via Eq. (4). The architecture is kept identical across four systems: NANMA isomerization, trialanine conformational equilibrium, a Diels–Alder cycloaddition at the GFN2-xTB level, and Trp-cage folding using an unbiased 208 µs trajectory. The authors report committor maps consistent with earlier ANN and shooting results, atom-level sensitivity analyses that identify chemically sensible atoms, and rate constants within a factor of roughly 2–4 of published values. The paper claims that the method bypasses hand-crafted collective variables (CVs) and offers automated reaction-coordinate discovery.","tokens_in":20882,"tokens_out":5178,"duration_ms":58105,"significance":"If the central reweighting and Markovianity assumptions are validated, the paper would be a useful methods contribution: it demonstrates that a single equivariant GNN can map raw coordinates to a physically reasonable committor, with external benchmarks against shooting (NANMA), a prior ANN map (Diels–Alder), and published rate constants. The SI's λ-stability analysis and the explicit use of an unbiased long trajectory for Trp-cage are strengths, and the sensitivity analysis provides interpretable mechanistic output. However, the headline claim of 'CV-free' learning is not yet supported: every non-Trp-cage system is trained on WTM-eABF data biased along hand-picked CVs, and the paper's own Discussion and SI Section 4 concede that the learned degrees of freedom can depend on the bias CVs. The significance of the contribution therefore hinges on additional validation that the learned committor is invariant to the bias protocol.","major_comments":[{"comment":"The reweighting formula in Eq. (4) is a static-bias reweighting: it treats δW as a fixed potential of mean force and assumes a stationary biased ensemble. WTM-eABF, however, deposits a time-dependent bias throughout the simulation. The Methods state that only the part of the trajectory corresponding to a 'converged free energy' is used, but no convergence diagnostic is reported, and no check shows that δW varies negligibly over the lag τ or over the training segment. The SI's own Section 4 warns that adding CV-dependent biases can influence the model's predicted degrees of freedom, which is direct evidence this concern is not merely hypothetical. Please provide (i) a convergence measure for δW over the training window, and (ii) an invariance test, e.g., training on WTM-eABF data generated with different bias CVs for NANMA or trialanine and showing that the committor map and sensitivity profile are unchanged within statistical error.","section":"Methods, Eq. (4)"},{"comment":"The abstract's claim that the method 'bypasses the need for hand-crafted collective variables' and learns 'without relying on prior assumptions' is contradicted by the Discussion's admission that 'to some extent, the learning rests upon our choice of the CVs used for biased sampling' and by the SI's warning that CV-dependent biases can alter the predicted primary degrees of freedom. For NANMA, trialanine, and Diels–Alder, the training data were generated with hand-picked RMSD or interatomic-distance biases, and the main comparison for Trp-cage uses biased data only in the SI. Please either soften the abstract and Introduction to state that the qGNN representation is CV-free while biased sampling is not, or provide a direct invariance demonstration (e.g., repeated training with different bias CVs yielding the same committor).","section":"Discussion, final paragraph; SI Section 4"},{"comment":"The variational principle used in the loss requires the chosen time lag τ to lie in the Markovian regime, but no Markovianity validation is reported for any of the four systems. The rate fits in Figures S1–S4 each use a different τ range, and the SI states the Markovian assumption without testing it. If τ is not in the Markovian regime, the minimized functional is not the true committor functional, and the rate constants derived from C[q;τ] are conditional on the lag. Please add a diagnostic (Chapman–Kolmogorov test, implied-timescale plateau, or equivalent) and, ideally, a scan of the learned q with respect to τ for at least one system.","section":"Methods, Eqs. (1)-(3); SI Section 1.1"},{"comment":"The graph construction is not specified: Algorithm 1 and the Methods define G = GNN[x] but do not state how nodes are connected (e.g., radial cutoff, k-nearest neighbors, or bond perception). Because the graph connectivity constrains the function class the qGNN can represent, the claim that the model learns 'directly from Cartesian coordinates' is only as strong as this representation choice. Please specify the edge construction in the Methods or SI and test the sensitivity of the learned committor to that choice.","section":"Methods, Algorithm 1"}],"minor_comments":[{"comment":"In the captions of Figures S5 and S6, the panel for λ = 10000 is labeled '(C)' but should be '(D)', since panels (A)–(C) are already used for λ = 0.5, 10, and 100.","section":"Figures S5, S6"},{"comment":"References 31 and 39 are the same work (Ajaz et al., J. Org. Chem. 2011) and should be consolidated; this also removes the duplicate entry in the reference list.","section":"References"},{"comment":"The caption title of Fig. S7 says 'from the unbiased simulation,' while the caption text and the surrounding paragraph describe a 10-ns biased simulation; please reconcile the wording.","section":"SI Section 4, Fig. S7"},{"comment":"The abstract describes 'precise estimates of the rate constants,' but Table 1 shows deviations of factors of about 3–4 from the reference values (e.g., Diels–Alder 4.40×10⁻³ vs 1.03×10⁻³ ps⁻¹). Suggest replacing 'precise' with 'consistent with' and reporting uncertainty estimates on the GNN rates.","section":"Abstract; Table 1"},{"comment":"Both availability statements say data and code 'can be provided upon request' without a repository or versioned archive; for reproducibility of a methodological paper, a persistent public repository with trained models and trajectories is strongly preferred.","section":"Data and Code Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is a concise methods brief, but the central 'CV-free' claim is materially stronger than the evidence and the authors' own caveats. The relationship to Ref. [17] (VCN) is close—the main change is replacing the ANN feature layer with a GVP-GNN—so the novelty framing should be precise. The reweighting and Markovianity concerns are the main technical blockers; the proposed invariance and convergence tests are feasible within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper shows that a GVP-GNN trained with the variational committor loss can recover the committor from Cartesian coordinates for NANMA and a Diels–Alder reaction, and the sensitivity analysis is a nice touch. But the headline claim that it bypasses collective variables is not supported by the actual workflow, and the reweighting of the time-dependent bias needs more scrutiny.\n\nWhat's genuinely new is the specific combination: an equivariant graph network fed raw heavy-atom coordinates, trained with the committor variational principle, plus a node-sensitivity layer for interpretability. That is a useful step beyond the CV-based ANNs in Refs. [17,28]. The NANMA committor map agrees with shooting results and with the earlier ANN map, and the Diels–Alder separatrix matches expectations. The SI's λ-stability tests are also reassuring.\n\nThe soft spots are real but not fatal. Equation (4) reweights the biased ensemble as if the WTM-eABF bias were a static potential, and the Methods only say the learning segment corresponds to a \"converged free energy\" without showing a convergence diagnostic. If the bias is still evolving over the training window, the learned q could reflect biased dynamics. That said, the agreement with shooting on NANMA is strong external evidence that the procedure is not badly broken, at least for that system.\n\nThe bigger problem is the language of the abstract. Biased trajectories are generated with hand-picked CVs (RMSDs, distances), and the Discussion concedes that learning \"rests upon\" that choice. The SI, in the Trp-cage section, explicitly warns that adding CV-dependent biases can influence which degrees of freedom the model flags. That directly undermines the \"CV-free\" claim as stated. Trialanine and Trp-cage also have no reference committor, and the rate constants are quoted without error bars and agree with literature only to factors of 1.5–4.3. Code and data are \"upon request,\" which is not ideal.\n\nBottom line: the paper deserves a serious referee, but I would not let the current abstract stand. A revision that tones down the CV-free claim, adds a bias-stationarity check, reports rate uncertainties, and ideally releases code/data would make this a genuinely useful contribution. As it is, read it for the NANMA and Diels–Alder demonstrations and for the sensitivity-analysis idea, but treat the broader claims with caution.","headline":"GVP-GNN committor learning is a promising and partly validated idea, but the paper overstates CV-freedom and under-reports bias-stationarity checks.","tokens_in":21449,"tokens_out":2920,"would_cite":false,"duration_ms":29622,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph neural network learns the committor directly from atomic coordinates, without hand-crafted collective variables.","keywords":["committor function","graph neural networks","geometric vector perceptrons","collective variables","reaction coordinate","transition path theory","molecular dynamics","rare events"],"falsifier":"Train qGNN on a system with an accurately known committor (for example, NANMA from shooting or a low-dimensional model solved by finite differences), using WTM-eABF data biased along very different CV sets, then compare separatrix and rate constants: if the learned q changes systematically with the choice of biasing CV, or if it disagrees with the shooting result in the barrier region, the central claim fails. A lighter check would be to compare two Trp-cage qGNNs, one from the 208 µs unbiased trajectory and one from the 10 ns biased trajectory; the SI already reports different node-sensitivity rankings, and matching committors would be required under the paper's claim.","tokens_in":20405,"feed_emoji":"🧬","tokens_out":8547,"duration_ms":72143,"temperature":0.7,"pith_summary":"This paper claims that the committor function—the probability that a configuration will reach product state B before returning to reactant state A—can be learned directly from Cartesian atomic coordinates by a graph neural network, with no hand-crafted collective variables or system-specific featurization. The network, called qGNN, is trained with a variational principle that minimizes the time correlation of its output under Markovian dynamics while enforcing boundary conditions. The authors demonstrate this on four systems of increasing complexity: conformational isomerization of two short peptides, a Diels–Alder cycloaddition, and reversible folding of Trp-cage. In each case the learned committor yields a separatrix consistent with reference results, a sensitivity analysis that pinpoints the atoms and distances that drive the transition, and rate constants within a small factor of published values. If the claim holds, reaction-coordinate discovery in molecular simulation no longer has to start from intuition about which collective variables matter.","feed_headline":"Committor learned from raw atomic coordinates, no hand-picked CVs","feed_subtitle":"One architecture reproduces transition rates across four systems and names the atoms that drive each change.","key_machinery":"The load-bearing mechanism is the geometric vector perceptron (GVP) grafted onto message-passing graph neural networks. A GVP maps tuples of scalar features and 3D vector features to new tuples while preserving rotation and reflection equivariance of the vectors, so the network can read geometric information directly from the atomic arrangement rather than from preselected distances or angles. The qGNN builds a molecular graph from coordinates, propagates node and edge features through GVP layers with message passing, and pools the resulting node embeddings to a scalar committor value. Training minimizes the variational loss of Eq. (2)—a reweighted time-correlation function of the committor plus boundary penalties—using a Siamese setup that compares the network output at time 0 and time $\\tau$. That is the object whose optimization identifies the committor.","core_discovery":"The central claim is that a single GVP-GNN architecture, fed only the Cartesian coordinates of heavy atoms (or Cα atoms for the protein), approximates the committor as a permutation-invariant, rotation-equivariant function of the whole configuration. Training uses the committor variational principle: the loss is the time-correlation functional $C[q;\\tau] = \\frac{1}{2}\\langle (q(\\tau)-q(0))^2\\rangle$ plus a penalty enforcing $q=0$ in A and $q=1$ in B, and biased-sampling data is reweighted by Eq. (4). Across NANMA, trialanine, the ethylene–vinylacetylene Diels–Alder reaction, and Trp-cage folding, the model produces committor maps whose separatrix matches variational-network and shooting references, node-sensitivity analyses that recover the known $\\phi/\\psi$ dihedrals, the newly formed bond distances, and physically meaningful end-to-end and helix distances in Trp-cage, and rate constants within about a factor of 1.5–4 of reference values. The authors are explicit that the method is not fully CV-free: the biased sampling that generates the training data still uses chosen CVs, and SI Section 4 warns that CV-dependent biases can influence the learned sensitivities.","pith_inferences":["A direct test of the reweighting assumption: train qGNN on the same system with different biasing CV sets and compare the resulting committor maps; the paper's own SI shows sensitivity differences between biased and unbiased Trp-cage runs, hinting that part of the learned function may inherit the bias.","The variational loss should also work with purely unbiased trajectory data, as in the Trp-cage case; if unbiased training succeeds on more systems, the residual dependence on hand-picked biasing CVs could be removed entirely, which would fully deliver the paper's title.","Equivariance suggests a strong testable symmetry: predictions should be exactly invariant under global rotations and translations of the input coordinates, and sensitivity gradients should rotate accordingly; any violation would point to a training artifact rather than the architecture.","A calibration against an exactly solvable low-dimensional model with a known committor, using the same reweighted biased data, could separate errors from the GVP approximation, from the reweighting, and from the choice of time lag."],"forward_implications":["The same architecture, unchanged except for graph size and a hyperparameter, yields committors for systems as different as a vacuum peptide isomerization, a semiempirical quantum-chemical cycloaddition, and a solvated protein folding; no per-system feature engineering is required.","Node- and edge-sensitivity analysis turns the trained network into an automated mechanism detector, naming the heavy atoms and pairwise distances that matter most for a transition.","Committor values can be converted into rate constants by estimating the first-passage flux; the reported rates land within a factor of about 1.5–4 of published experimental or theoretical values across all four systems.","Because the network never reduces the configuration to a few collective variables, the transition ensemble can be examined directly in full coordinate space, as done in the Trp-cage clustering near the separatrix."],"supporting_citations":[{"why":"Supplies the variational principle for the committor time-correlation functional that defines the loss.","marker":"[6]"},{"why":"Establishes the committor as an optimal reaction coordinate, motivating the variational estimation.","marker":"[12]"},{"why":"Defines the variational committor-network loss and reweighting, and supplies reference committor maps and rate constants for comparison.","marker":"[17]"},{"why":"Provides the GVP-GNN architecture and the equivariance and universal-approximation properties the qGNN builds on.","marker":"[26]"},{"why":"Introduces the Siamese variational training scheme and iterative pathway ideas the qGNN adapts.","marker":"[28]"},{"why":"Provides the 208 µs unbiased Trp-cage trajectory used as the folding testbed.","marker":"[33]"},{"why":"Supplies the GFN2-xTB semiempirical method used for the Diels–Alder energies and forces.","marker":"[38]"},{"why":"Supplies the WTM-eABF enhanced-sampling algorithm that generates the biased training data.","marker":"[51]"}],"fun_headline_variants":["Committor direct from atoms: GNN bypasses CV selection","Raw coordinates to kinetics: GNN learns committor","Skipping collective variables: GNN predicts committor","Atom-level committor learning skips collective variables"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction rests on the assumption that the time-dependent WTM-eABF bias used to generate training data can be handled by a quasi-static reweighting formula over a converged portion of the run, and that the chosen time lag $\\tau$ falls in the Markovian regime; if either fails, the network learns the biased dynamics rather than the true committor.","fun_headline_variants_meta":{"raw":{"variants":["Committor direct from atoms: GNN bypasses CV selection","Raw coordinates to kinetics: GNN learns committor","Skipping collective variables: GNN predicts committor","Atom-level committor learning skips collective variables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2486,"prompt_tokens":924,"completion_tokens":1562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1493}},"tokens_in":540,"tokens_out":1562,"duration_ms":14068,"temperature":1.0,"reasoning_tokens":1493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:18:57.891594+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train qGNN on a system with an accurately known committor (for example, NANMA from shooting or a low-dimensional model solved by finite differences), using WTM-eABF data biased along very different CV sets, then compare separatrix and rate constants: if the learned q changes systematically with the choice of biasing CV, or if it disagrees with the shooting result in the barrier region, the central claim fails. A lighter check would be to compare two Trp-cage qGNNs, one from the 208 µs unbiased trajectory and one from the 10 ns biased trajectory; the SI already reports different node-sensitivity rankings, and matching committors would be required under the paper's claim.","supporting_citations":[],"review_version":1}