{"id":"dd103c91-cc5d-4a19-9077-50898f661601","arxiv_id":"2412.01947","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An iterative neural network method learns the committor and a transition path aligned to its gradient, tested on model potentials and two biomolecular systems.","lead":"Scientists built a neural network that learns both the probability of a molecular transition and the most likely path connecting two states, updating the path over several rounds. The method reproduces known routes in simple models and finds multiple folding pathways in a small protein.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central step—asserting that minimizing Cqq(τ) learns the true committor and that the softmax CCS inherits its gradient—is not demonstrated at finite τ for the biased, reweighted trajectories actually used, so the committor-consistency claim may be circular.","rationale":"The reader's weakest_assumption correctly focused on the Markovian embedding/CV-sufficiency premise. I add two concrete refinements: first, the paper's own SI proof only establishes geometric alignment of the softmax expansion, not physical committor consistency; second, the unbiasing scheme (Eq. 7) requires small τ while the variational principle needs a large enough τ to isolate the reactive eigenmode, and this tension is never checked. These are correctness risks, not mere 'outside consensus' issues: the central claim that the CCS is 'committor-consistent' and can be used to estimate accurate rate constants—which fails numerically on chignolin and triple-well 100 K—rests on exactly this unverified premise. The paper does independent support through reproducible benchmarks (BS, MB, NANMA) and a clear iterative protocol, so a full rejection is not justified. A CONDITIONAL verdict with a request for the CV-completeness test and a quantitative check of the τ/rewighting trade-off is the appropriate recommendation.","tokens_in":29668,"tokens_out":1666,"duration_ms":208202,"concrete_test":"Run the double-well/BS test with trajectories projected onto a deliberately incomplete CV (e.g., only x for the BS potential with γx/γy=10) and check whether the learned committor/CCS and the rate constant match the known anisotropic solution; if the method reproduces the wrong gradient-driven path and rate, that confirms the CV-sufficiency assumption is load-bearing. Additionally, on chignolin, add a third hydrogen-bond distance (e.g., Asp3N-Thr6O) as a CV and compare the learned q and CCS two-dimensional projection to the two-CV result; if the two-CV committor shifts appreciably, the reported pathway is not committor-consistent.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The variational committor principle (Eq. 1, Eqs. 2–3) is valid for the true committor only when {z(x)} is a Markovian embedding and τ is large enough that the two-point correlation Cqq(τ) is dominated by the slow reactive eigenmode. For chignolin, the only CVs are the two hydrogen-bond distances (Asp3N-Gly7O, Asp3N-Thr8O), and the paper itself admits the method 'depends on a careful choice of CVs' and that larger CV sets are intractable in practice. If these two distances do not capture all slow channels (e.g., Asp3-Thr6 interactions seen in the transition-state medoids), the learned F(z) is a projection of the true committor, and Eq. 5's qs, which is forced to interpolate the values of F at the string images through a softmax basis on the same CVs, cannot recover the true reactive flux. The proof in SI that ∇qs ∝ string tangent only shows alignment of the expansion to its own softmax basis, not consistency with the physical committor. A second issue is that the biased simulations are reweighted with Eq. 7, which the SI states is valid only for 'sufficiently small' τ, while the variational principle requires a finite/large τ to isolate the reactive eigenmode; the paper never quantifies this conflict. Finally, Table 1 shows the PCCANN rate for chignolin (8.76e-5 ps^-1) is ~50x the MFPT reference (1.6e-6 ps^-1), which is not explained by the claimed sampling-convergence caveats.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PCCANN, an iterative neural-network method that couples a variational committor network (VCN) with a softmax-expanded 'committor-consistent string' (CCS). Starting from a guess path, the method runs biased MD simulations along a path-collective variable, learns a committor q(z) by minimizing the two-point correlation Cqq(τ), constructs a string whose tangent is aligned with the gradient of a Gaussian expansion qs(z) of the learned committor, and iterates until convergence. The method is applied to a 1D double-well, the Berezhkovskii–Szabo (BS), Müller–Brown (MB), and triple-well (TW) potentials, to NANMA, and to the mini-protein chignolin. The authors report recovery of known pathways and rate constants for the model potentials, detection of bifurcating reactive tubes in TW and chignolin, and a chignolin folding free-energy difference consistent with prior work.","tokens_in":30073,"tokens_out":7443,"duration_ms":70890,"significance":"If the central claims are correct, the PCCANN would be a valuable addition to the enhanced-sampling toolbox: it provides a way to iteratively refine both a committor and a reaction path from biased simulation data, and it explicitly handles anisotropic diffusion and multiple reactive pathways. The BS and MB benchmarks are genuinely convincing, and the iterative use of reweighted biased trajectories is an appealing practical strategy. However, the quantitative rate predictions for TW at 100 K and for chignolin deviate from the unbiased references by factors of roughly 70 and 55, respectively, which is hard to reconcile with the abstract's claim of reproducing 'established dynamics and rate constants'. Moreover, the physical consistency of the CCS hinges on the unverified assumption that the chosen CVs form a Markovian embedding and on the accuracy of the VCN committor, which is not independently validated for the molecular systems. These gaps currently prevent acceptance of the general claims.","major_comments":[{"comment":"The reweighting expression Eq. (7) is stated to be valid for 'sufficiently small' values of τ, while the variational principle in Eq. (1) requires τ to be large enough that the dynamics in the CV subspace is Markovian and the slow reactive eigenmode dominates Cqq. The manuscript never quantifies this conflict or shows that the τ values used for the TW and chignolin systems satisfy both conditions. Because all loss functions and rate constants are computed from reweighted Cqq, this is a load-bearing gap. I ask for a concrete numerical test, e.g., computing the committor and kAB for a system with accessible unbiased dynamics (BS or MB) from both unbiased and reweighted biased trajectories at identical τ and demonstrating agreement or a clear convergence in τ.","section":"Methods, Eq. (7) and SI 'Unbiasing Scheme'"},{"comment":"The biased PCCANN rate constants are 4.00e-3 vs 5.70e-5 ps^-1 for TW at 100 K and 8.76e-5 vs 1.6e-6 ps^-1 for chignolin, i.e., discrepancies of factors of about 70 and 55. The SI attributes the chignolin discrepancy to possible convergence problems in WTM-eABF, but this does not explain the systematic overestimation, and the TW discrepancy is not discussed quantitatively. Since the abstract claims accurate rate constants, these deviations undermine a central claim. The authors should either provide a quantitative explanation (e.g., a study of kAB versus tube width, or versus the convergence of δW) or explicitly limit their accuracy claim to the systems where agreement is demonstrated.","section":"Table 1, TW 100 K and chignolin rows"},{"comment":"The 'committor-consistency' of the CCS is imposed by construction: qs(z) in Eq. (5) is a softmax interpolant of the learned committor values at the images, and the loss Ltr in Eq. (6) explicitly penalizes misalignment between ∇qs and the string tangent. The SI proof therefore shows only that this particular expansion aligns with its own basis for large α, not that qs equals the physical committor. The paper should state this distinction clearly and provide an external validation of the learned committor (for example, by comparison with a committor estimated from unbiased shooting trajectories) for at least one system beyond the analytical potentials. As it stands, the physical relevance of the CCS rests entirely on the accuracy of the VCN committor, which is not independently verified for chignolin.","section":"The Committor-Consistent String, Eqs. (5)-(6), and SI Eqs. (S1)-(S2)"},{"comment":"The variational principle in Eq. (1) is valid only if the chosen CVs form a Markovian embedding. The Discussion acknowledges that the method 'depends on a careful choice of CVs', but for chignolin the model uses only two hydrogen-bond distances, while the transition-state clusters in Fig. 6 and SI Fig. 12 involve Asp3–Thr6 and Asp3–Thr8 interactions not represented by these CVs. No test of Markovianity or of CV-set completeness is provided. This is the weakest load-bearing assumption for the biological application. I request either a numerical test (e.g., comparing the two-CV committor with one obtained from a larger CV set or from a Markov state model) or an explicit statement that the chignolin results are conditional on this CV set and may be biased by the projection.","section":"Discussion and Chignolin section"}],"minor_comments":[{"comment":"The double-well potential in Eq. (8) has two identical Gaussian terms, both centered at x = -2; one of them is presumably intended to be centered at x = +2.","section":"SI, Eq. (8)"},{"comment":"There is a typo in 'Cartersian coordinates'; it should read 'Cartesian coordinates'.","section":"Methods, first paragraph"},{"comment":"The spelling of 'Berezhkovskii–Szabo' varies, including 'Berezhkovskii-Sazbo' in Table 1; please standardize. Similarly, 'Müller–Brown' is rendered inconsistently.","section":"Throughout and Table 1"},{"comment":"The manuscript does not report the actual time-lag values τ used for each system. Since the validity of the reweighting and of the variational principle depends on τ, a table of the τ values (or at least a range) should be included.","section":"PCCANN and Iterative Learning Procedure"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially useful contribution that builds on the authors' own VCN and string-method work. The model-potential benchmarks are convincing and the iterative protocol is well motivated. However, the quantitative rate errors for TW at 100 K and chignolin are too large to dismiss as caveats, and the Markovian-embedding assumption for the two-CV chignolin model is not tested. I would encourage the editor to seek a revision that either fixes these issues with additional numerical evidence or carefully narrows the claims. The paper is within the scope of the journal and, if revised, could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper combines the variational committor network (VCN) and the committor-consistent variational string method into an iterative scheme, PCCANN, that learns a committor and a string aligned to its gradient in one pass. The benchmark results on the Berezhkovskii–Szabo and Müller-Brown potentials are solid, and the demonstration that the CCS picks up anisotropic diffusivity where the minimum free-energy path does not is a genuinely nice result. The method also handles multiple pathways in the triple-well potential and chignolin, and nuisance variables in NANMA with four dihedrals. The writing is clear and the SI has real detail.\n\nThe soft spots are in the quantitative claims. Table 1 reports biased PCCANN rate constants that deviate from the unbiased references by factors of roughly 50–70 for the triple-well potential at 100 K and for chignolin. The paper blames sampling convergence, but the deviations are large enough that \"accurate rate constants\" is an overstatement. The more conceptual issue is the committor-consistency step. The SI proof that ∇q_s aligns with the string tangent is an asymptotic argument about the softmax expansion; it shows internal consistency of the expansion with its own basis, not that q_s equals the physical committor. And the reweighting of biased trajectories, equation (7), is stated to be valid only for small τ, while the variational principle needs a finite τ to isolate the reactive eigenmode. The paper never quantifies this tension. That said, the method is validated against external references where it works, and the authors openly admit the dependence on careful CV choice. So the circularity is partial rather than fatal.\n\nThe lack of released code and data is a minor annoyance: \"provided upon request\" is not a reproducibility standard.\n\nWho should read this: anyone working on rare-event MD with machine-learned committors. It deserves a serious referee. The core idea is worth engaging with, and the benchmark successes justify review time even though the rate-constant section needs scrutiny. I would send it out, with a request to temper the kinetic claims and to release the code.","headline":"A plausible hybrid of VCN and variational string method that works on benchmarks but overstates rate accuracy and leaves the committor-consistency proof partly heuristic.","tokens_in":30617,"tokens_out":3333,"would_cite":true,"duration_ms":34806,"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":"Neural network learns transition paths that obey the committor, in a single iterative loop.","keywords":["committor","transition path","rare events","artificial neural network","string method","enhanced sampling","molecular dynamics","free-energy calculations"],"falsifier":"Run the PCCANN on a system with a known full-dimensional committor, such as a two-dimensional potential with anisotropic diffusion, but train it on a deliberately incomplete CV, for example $x$ alone, and compare the predicted rate constant and string with exact values from long unbiased simulation. If the incomplete-CV CCS gives the same rates and paths, the Markovian-embedding assumption is not doing the work; if it gives different rates, the claim that the method recovers the true reactive path without adequate CVs is refuted. A sharper test: compute the angle between the converged string tangent and the true committor gradient from a long unbiased trajectory, since systematic misalignment near the separatrix would contradict the central consistency condition.","tokens_in":29468,"feed_emoji":"🧬","tokens_out":5934,"duration_ms":53385,"temperature":0.7,"pith_summary":"The paper proposes an iterative machine-learning scheme that solves a chicken-and-egg problem in rare-event molecular dynamics: finding a good transition pathway requires knowing the committor, while knowing the committor requires sampling near the pathway. Its neural network, the path-committor-consistent artificial neural network (PCCANN), learns the committor from biased trajectories and simultaneously refits the path so that the string tangent aligns with the committor gradient. The updated path defines a path-collective variable for a new biased simulation, and the cycle repeats until the string and committor converge. On benchmark potentials, a peptide isomerization, and chignolin folding, the converged strings cross the proper saddle points, reproduce rate constants from unbiased simulation, and expose bifurcating reactive tubes. If correct, the method gives enhanced-sampling simulations a reaction coordinate that is consistent with transition-path theory rather than merely convenient.","feed_headline":"Neural network learns transition paths that obey the committor","feed_subtitle":"Aligning paths with committor gradients yields consistent rates and pathways from peptides to chignolin.","key_machinery":"The machinery is variational learning of the committor plus a string layer forced to be consistent with it. The committor $q(z)$ is the probability that a trajectory from $z$ reaches $B$ before $A$, and the variational principle minimizes $C_{qq}(\\tau) = \\langle (q(\\tau)-q(0))^2 \\rangle/2$, the two-point time correlation whose minimum gives the reactive flux. The network is a Siamese ANN with shared weights; it receives $z(t)$ and $z(t+\\tau)$, and its output $F(z)$ is constrained by basin losses. The string is expanded in normalized Gaussians, $\\sigma_k(z) \\propto \\exp(-\\alpha (z - \\hat z_k)^2)$, and the committor $q_s(z)$ is the least-squares fit of such an expansion to the network committor at the images. For large $\\alpha$ the gradient of $q_s$ at the midpoints between neighboring images is parallel to the string tangent, so the loss $L_s$ learns, by construction, a path that follows the committor gradient; the iterative loop then resimulates along a path-collective variable defined by the current string, producing new biased data from which a better committor and string are learned.","core_discovery":"The central claim is that one can learn the committor $q(z)$ and a transition pathway from the same network, and that the pathway should be built as a committor-consistent string (CCS): a chain of images whose tangent is parallel to $\\nabla q_s$, the gradient of a softmax-expanded committor $q_s(z) = \\sum_k b_k \\sigma_k(z)$. The expansion coefficients $b$ are fixed by least-squares interpolation to the network committor at the images, and the string positions are optimized by minimizing a total loss $L_s = 2C_{q_s q_s} + k_b L'_b + k_{eq}L_{eq} + k_{tr}L_{tr}$, where $C_{q_s q_s}$ is the two-point correlation whose minimization is the variational committor principle and $L_{tr}$ enforces tangent-gradient alignment. Because biased simulation corrupts the dynamics, the time correlations are computed with an exponential reweighting that the paper states is valid for small lag $\\tau$. Starting from a rough path, computing a committor-averaged path, and iterating defines the workflow. The paper reports that for the Berezhkovskii-Szabo potential the CCS reproduces anisotropic-diffusion effects that the minimum free-energy path misses; for Muller-Brown and the triple well it matches reference paths and captures temperature-dependent bifurcation; for NANMA it recovers known transition-state correlations; and for chignolin it identifies two folding pathways whose transition-state ensembles agree with an independent study.","pith_inferences":["Beyond the paper, the consistency condition suggests a practical convergence diagnostic for string methods generally: report the angle between the string tangent and $\\nabla q_s$ along the path, rather than only checking free-energy convergence.","A direct stress test the authors do not report in full is to enlarge the CV set used as network input and check whether the CCS and rate constants drift; the NANMA comparison between $(\\phi,\\psi)$ and $(\\phi,\\psi,\\theta,\\omega)$ is a partial version of this, and applying it to chignolin with additional hydrogen bonds would separate CV-choice error from learning error.","The softmax width $\\alpha$ is doing double duty, defining both the PCV and the committor expansion, and the paper notes different values are needed; an automatic schedule for $\\alpha$ tied to the instantaneous mean squared displacement between images would make the method less dependent on hand-tuning.","The small-$\\tau$ reweighting assumption could be tested head-on by comparing $C_{qq}$ computed from biased trajectories with $C_{qq}$ from an unbiased reference on a model with known rates; the BS potential results already provide a natural setting for this check."],"forward_implications":["Converged CCS paths provide a one-dimensional reaction coordinate for biased simulations that is consistent with transition-path theory, not merely a lowest-free-energy route.","Rate constants can be extracted from the learned committor through the flux formula $k_{AB} = C_{qq}/(p_A \\tau)$, even when the sampling is biased, using the small-lag reweighting.","Bifurcating reactive tubes can be separated by clustering isocommittor slices, yielding a distinct CCS per pathway, as shown for the triple well and chignolin.","Anisotropic diffusion shifts the optimal string away from the MFEP, and the CCS captures that shift, so path-collective-variable simulations along a CCS should explore the physically dominant reactive tube."],"supporting_citations":[{"why":"Supplies the committor-consistent variational string method that the CCS construction generalizes.","marker":"[21]"},{"why":"Introduces the variational committor network whose loss function, reweighting, and basin constraints the PCCANN builds on.","marker":"[22]"},{"why":"Defines path collective variables, the functional form used to turn the CCS into a one-dimensional reaction coordinate for biased simulation.","marker":"[13]"},{"why":"Provides the finite-lag variational principle and the relation between the two-point correlation and reactive flux.","marker":"[19]"},{"why":"Supplies the transition-rate and spectral framework, including the separatrix criterion that guides string direction.","marker":"[20]"},{"why":"Independent recent study used as the comparison for chignolin transition-state ensembles and pathway bifurcation.","marker":"[27]"},{"why":"Provides the Dijkstra-based minimum free-energy path algorithm used as a reference path in several benchmarks.","marker":"[48]"},{"why":"Supplies the WTM-eABF biased-sampling algorithm used throughout the iterative procedure.","marker":"[46]"},{"why":"Defines the Berezhkovskii-Szabo potential benchmark with anisotropic diffusivity that the CCS must reproduce.","marker":"[43]"}],"fun_headline_variants":["Neural net learns committor-consistent paths for rare biomolecular events","Iterative neural networks align paths with committor gradients","AI finds transition pathways by matching committor gradients","Neural nets refine transition paths to obey committor consistency","Committor-consistent pathways from iterative neural network learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the hand-chosen collective variables $z(x)$ form a Markovian embedding: the committor depends only on these variables, and the time-lagged dynamics in them is Markovian at the lag $\\tau$; the paper itself states in the Discussion that the method depends on a careful choice of CVs.","fun_headline_variants_meta":{"raw":{"variants":["Neural net learns committor-consistent paths for rare biomolecular events","Iterative neural networks align paths with committor gradients","AI finds transition pathways by matching committor gradients","Neural nets refine transition paths to obey committor consistency","Committor-consistent pathways from iterative neural network learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2367,"prompt_tokens":997,"completion_tokens":1370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1285}},"tokens_in":613,"tokens_out":1370,"duration_ms":9122,"temperature":1.0,"reasoning_tokens":1285,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:00:34.274858+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the PCCANN on a system with a known full-dimensional committor, such as a two-dimensional potential with anisotropic diffusion, but train it on a deliberately incomplete CV, for example $x$ alone, and compare the predicted rate constant and string with exact values from long unbiased simulation. If the incomplete-CV CCS gives the same rates and paths, the Markovian-embedding assumption is not doing the work; if it gives different rates, the claim that the method recovers the true reactive path without adequate CVs is refuted. A sharper test: compute the angle between the converged string tangent and the true committor gradient from a long unbiased trajectory, since systematic misalignment near the separatrix would contradict the central consistency condition.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the finite-lag variational principle and the relation between the two-point correlation and reactive flux."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the transition-rate and spectral framework, including the separatrix criterion that guides string direction."}],"review_version":1}