{"id":"64048378-fdb7-49b8-98ec-a0fb5ecff80e","arxiv_id":"2603.25237","paper_version":2,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A review showing that combining committor-trained neural networks with LIME/SHAP attribute analysis identifies physically meaningful reaction coordinates for alanine dipeptide isomerization and NaCl ion-pair dissociation in water.","lead":"This paper reviews a deep-learning framework that trains neural networks to predict the committor — the probability that a molecular configuration reaches product before reactant — and then uses explainable AI (LIME and SHAP) to identify which collective variables matter most. For scientists studying molecular transitions, it offers a systematic way to find reaction coordinates from simulation data instead of relying on trial and error.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NaCl case fails the committor histogram test the framework itself uses; the claimed 'well-defined separatrix' rests on grid-averaged p*_B, not on a demonstrated valid RC.","rationale":"The reader's weakest_assumption focused on the convergence of committor estimates from short 1 ps trajectories, which is indeed a general concern. My stress test identifies a more specific and directly documented failure: the NaCl case explicitly fails the sharp-peak committor test near q=0, yet the abstract and Section III C claim a well-defined separatrix. This is not merely a sampling-convergence risk; it is an acknowledged inconsistency between the model's predicted RC and the standard criterion for RC quality. The concrete test would settle whether the broad distribution is a finite-sampling artifact or a real failure. Because the paper is a review of prior work rather than a new primary claim, the verdict remains somewhat unverdictable as a novel contribution, but the central methodological claim should only be accepted conditionally, with the NaCl shortcoming either resolved or clearly presented as a limitation, not as a successful demonstration.","tokens_in":27036,"tokens_out":3633,"duration_ms":37688,"concrete_test":"Recompute the committor histogram for all NaCl test-set configurations with |q|<0.2 using 1000 independent shooting trajectories per configuration (or trajectories extended to 10 ps) to distinguish finite-sampling noise from genuine scatter. If the distribution still lacks a sharp peak at p*_B=1/2, then the learned q fails the committor test, and the separatrix line in Figure 11 should be regarded as an artifact of grid-averaging rather than evidence of a valid reaction coordinate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the framework enables identification of CVs that yield 'well-defined boundaries' is directly contradicted by the paper's own NaCl result. In Section III C and Figure 9, the text states: 'the committor values near q=0 are widely distributed between 0 and 1 and do not exhibit a clear sharp peak at p*_B=1/2.' The standard committor histogram test, which the paper itself endorses in Section II A, requires a sharp peak at p*_B=1/2 for a good reaction coordinate. The claim of a 'well-defined separatrix line' in Figure 11 is based on grid-averaged p*_B contours, and averaging can produce a smooth p*_B=0.5 contour even when individual configurations have p*_B scattered near 0 and 1. Thus the identification of G5_58 and G5_1217 as dominant CVs, and the implication that they form a valid RC together with r_ion, is not supported by the committor criterion. The paper speculates that the failure is due to the large number of input variables and might be overcome by hyperparameter tuning, but this is untested. Since the NaCl application is one of the two core demonstrations, this unresolved inconsistency weakens the framework's advertised capability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a review of the authors' own explainable deep-learning framework for identifying reaction coordinates (RCs) from committor values. It describes the methodology—committor sampling, cross-entropy minimization as a loss function, a deep neural network mapping candidate collective variables to an RC, and LIME/SHAP for feature attribution—and surveys applications to alanine dipeptide isomerization, hyperparameter tuning, and NaCl ion-pair dissociation in water. The central claim is that combining deep learning of the committor with XAI enables identification of dominant collective variables and shows that the committor distribution on the surface of those variables is separated by well-defined boundaries. The review explicitly states that no new data were generated.","tokens_in":27376,"tokens_out":5603,"duration_ms":52754,"significance":"If the framework performs as advertised, it would offer a practical, interpretable route to RC identification in high-dimensional molecular systems. The alanine dipeptide application is convincing: the learned RC gives a sharp p*_B = 1/2 peak near q=0, and LIME/SHAP consistently identify the dihedral angle θ (alongside φ) as dominant, with the attribution shifting to θ near the transition state. The hyperparameter study is a useful robustness check, showing that different architectures yield similar RCs and consistent feature attributions. However, the NaCl case—one of the two core demonstrations—fails the very committor histogram test the paper itself endorses, and the claimed 'well-defined separatrix' is derived from grid-averaged data rather than from configuration-resolved committor validation. As it stands, the broad claim in the abstract is not supported by the full body of evidence.","major_comments":[{"comment":"The abstract claims that the approach 'demonstrates that the committor distribution on the surface using important CVs is separated by well-defined boundaries.' This is contradicted by the paper's own NaCl result: Section III C states that 'the committor values near q=0 are widely distributed between 0 and 1 and do not exhibit a clear sharp peak at p*_B=1/2.' According to the committor histogram test described in Section II A, a valid RC must produce a sharp peak at p*_B=1/2. The speculation that this failure is due to the large number of input variables and would be overcome by hyperparameter tuning is untested. The NaCl demonstration therefore does not support the advertised capability.","section":"Abstract; Section III C, Figure 9"},{"comment":"The 'well-defined separatrix line' in Figure 11 is obtained by dividing the (r_ion, G5) plane into a 200×200 grid, averaging p*_B within each cell, and plotting the p*_B=0.5 contour after cubic interpolation. Grid-averaging can produce a smooth p*_B=0.5 contour even when individual configurations have p*_B scattered near 0 and 1, which is exactly the behavior shown in Figure 9 near q=0. The manuscript does not report per-configuration committor histograms on either side of the separatrix. Without such validation, the claim that G5_58 and G5_1217, together with r_ion, form a valid RC is not established.","section":"Section III C, Figure 11"},{"comment":"The learned RC q is a function only of the 1,296 ACSF inputs; r_ion was not included as an input feature, as stated in Section III C ('two types of ACSFs ... were employed as CVs for the neural network inputs'). Nevertheless, the paper concludes that 'G5_58 or G5_1217 will appropriately represent the RC together with the interionic distance r_ion.' This inference is not supported by the model: the DNN never saw r_ion, so any role of r_ion in the RC is an external assumption. The 2D PMF plots in Figure 11 superimpose committor data on (r_ion, G5) but do not demonstrate that the learned q is a function of this pair. The manuscript should either include r_ion as an input feature or rephrase the conclusion to avoid claiming that r_ion is part of the identified RC.","section":"Section III C; Section II D"}],"minor_comments":[{"comment":"The sentence 'N_node most frequently converged to 5 and 3 in vacuum and in water, respectively' conflicts with the preceding description that N_node was searched from 100 to 5000. It appears that N_layer is meant. Please correct.","section":"Section III B"},{"comment":"There are numerous typos and encoding artifacts: 'depeptide' (Section III A heading), 'resepectively', 'adecuacy', 'hypearparameter', 'committer' for 'committor' in several places, 'rubust', 'are are' duplication, and 'Moveover'. Figure 12 contains placeholder characters '□' in place of Å^-3 and negative signs. These should be fixed.","section":"Throughout"},{"comment":"The journal reference lists 'J. Chem. Phys.164, 164, 094101' with a duplicated page number; should be 'J. Chem. Phys.164, 094101 (2026)'.","section":"Reference 141"},{"comment":"The phrase 'exhibits a second highest contribution following the RC' is unclear because r_ion was not an input. Clarify whether 'RC' refers to the predicted q or to a separately considered variable.","section":"Section III C"},{"comment":"The state definitions such as '(−150°,0°)≤(φ,ψ)≤(30°,180°)' are ambiguous; specify the intervals for φ and ψ separately to avoid confusion.","section":"Section III A"},{"comment":"The display of the SHAP kernel is mangled: 'MC|z''||z''|(M−|z''|)' should be typeset as the binomial-coefficient form. Please revise for readability.","section":"Section II E, Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is essentially a review of the authors' own published work, with no new data. That is acceptable for a review venue, but the review's abstract and summary overclaim a general capability that its own NaCl application refutes. The alanine dipeptide case is solid, and the hyperparameter robustness study is interesting, but the NaCl section needs either configuration-resolved committor validation or a clearly stated limitation. The authors' speculation that hyperparameter tuning will resolve the NaCl histogram failure is plausible but untested; as written, this is a load-bearing gap. I would recommend major revision rather than rejection because the issues are local and fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nQuick take: this is a review, not a new result. It consolidates the group's prior work on combining deep learning of committor values with LIME/SHAP for reaction coordinate identification. If you want a clear account of that framework and its alanine dipeptide application, this does the job. But the paper overstates what the NaCl case actually shows, and the overstatement is visible in the paper's own text.\n\nWhat's good: the methods section is well organized. The KL-divergence framing of the cross-entropy objective is standard and clearly explained. The alanine application is convincing: the histogram test gives a sharp peak at p*B=1/2 near q=0, and the LIME/SHAP features agree with the linear-regression coefficients and with the physical intuition that theta matters near the TS. The hyperparameter tuning section, from the authors' 2025 paper, is a useful practical message: many similar DNNs give the same RC, which is a genuinely nice robustness result.\n\nThe soft spot is the NaCl part. The paper's own Figure 9 shows p*B widely scattered between 0 and 1 near q=0, and the text explicitly says there is no clear sharp peak at 1/2. The histogram test, which the paper itself endorses in Section II.A, calls that a bad RC. The 'well-defined separatrix line' in Figure 11 is drawn from grid-averaged p*B contours, and averaging can produce a smooth 0.5 contour even when the underlying points are scattered. That is a real logical gap, not a nitpick. The paper waves it away with 'likely due to the larger number of input variables' and says future hyperparameter tuning might fix it. That is speculation. The abstract's blanket claim that the framework 'demonstrates' well-defined boundaries is not supported by the NaCl data.\n\nAlso minor: because the review reuses the authors' own published figures and text, an editor should ask how much new synthesis there is beyond the prior papers. The other-group work (Naleem et al., Jung et al.) is cited but not critically compared.\n\nWho's it for: a reader who wants a compact, reliable exposition of this particular pipeline, or a teaching reference. Not someone looking for independent validation, because there is none here.\n\nRecommendation: send to peer review if the venue publishes review articles, but with a clear instruction to the authors to either soften the NaCl claims to match their own Figure 9 or present a real committor histogram for the proposed RC. As is, the abstract overclaims.","headline":"A competent review of the authors' own committor+XAI framework, but the NaCl demonstration undercuts the central claim of 'well-defined boundaries' and the paper's own text admits it.","tokens_in":27902,"tokens_out":2055,"would_cite":false,"duration_ms":19973,"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 deep neural network trained on committor values, read with explainable-AI attribution, reveals the few collective variables that actually govern molecular transitions — and where the transition-state dividing surface sits.","keywords":["reaction coordinate","committor","deep learning","explainable AI","SHAP","LIME","atom-centered symmetry functions","molecular dynamics"],"falsifier":"Recompute committors for a few hundred configurations near q=0 using 10,000 instead of 100 velocity assignments and compare the p*_B distribution. If it does not sharpen into a peak at 1/2 — for NaCl, the paper already shows a wide 0-to-1 spread — the learned RC is not truly separating transition states. A second check: shoot trajectories from configurations lying on the claimed separatrix line of the (r_ion, G5_58) PMF; measure how often they commit to A vs B. An isocommittor surface would give 50/50 outcomes.","tokens_in":26919,"feed_emoji":"🧪","tokens_out":7243,"duration_ms":71801,"temperature":0.7,"pith_summary":"This review argues that the reaction coordinate for a molecular transition can be learned from committor probabilities without relying on physical intuition: a deep neural network maps candidate collective variables to a reaction coordinate, and explainable-AI tools (LIME, SHAP) then rank which input variables control the prediction. Applied to alanine dipeptide isomerization, the framework identifies dihedral θ as the dominant coordinate, with its importance peaking near the transition state in a way global linear models miss. Applied to NaCl dissociation in water, it identifies two atom-centered symmetry functions that, together with the interionic distance, separate associated from dissociated states by a well-defined p_B = 0.5 boundary. The review also shows that different neural-network architectures give nearly identical reaction coordinates and feature rankings, so the extracted mechanism does not depend on the specific model chosen. If right, the framework offers a data-driven path from simulation trajectories to interpretable molecular mechanisms in complex systems.","feed_headline":"Deep learning finds the variables that drive molecular reactions","feed_subtitle":"Alanine dipeptide and NaCl in water reveal which torsional and solvent coordinates set the transition state.","key_machinery":"The central object is the committor p*_B(R) — the probability of reaching state B before A with thermal velocities; the transition state is the p_B=1/2 surface. The machinery: sample near the saddle, estimate p*_B from short trajectories; train a multilayer perceptron mapping candidate collective variables to a reaction coordinate q, with cross-entropy loss enforcing p_B(q)=(1+tanh q)/2; then use LIME and SHAP to rank each input's contribution. This turns a black-box committor fitter into a mechanistic tool: the top-ranked CVs define low-dimensional PMFs whose isocommittor lines mark the TS.","core_discovery":"The paper's central claim: a deep neural network trained on committor values, then interrogated with LIME/SHAP, identifies which collective variables actually define the reaction coordinate. For alanine dipeptide, q reproduces the committor with a peak at p_B=1/2, and attribution shows dihedral θ (with φ) dominates, the θ contribution growing near the transition state in a way global linear models miss. For NaCl in water, SHAP singles out two atom-centered symmetry functions (G5_58: O–Na–O shell at 2.0 Å; G5_1217: Na–Cl–O angular term) that, with r_ion, yield a well-defined p_B=0.5 separatrix. The review also finds that different hyperparameters give nearly identical RCs and feature rankings","pith_inferences":["A natural stress-test: verify that the p*_B=0.5 separatrix in the identified 2D surface is actually an isocommittor surface (i.e., shooting from points along it reproduces p_B≈0.5 within error). The NaCl case already hints the learned RC may be imperfect, since committor values near q=0 are spread between 0 and 1 rather than sharply peaked.","The framework still depends on the preselected candidate CVs: if the true RC involves a coordinate not in the input set, the network cannot discover it. Integrating automated feature generation (e.g., graph-based descriptors) could close that gap.","The XAI attribution could be used online to guide adaptive sampling: focus new committor evaluations where the attributed dominant CVs are most uncertain, reducing the 100-trajectory cost per configuration.","The finding that G5_1217 increases toward the TS and then decreases implies a late-barrier solvent rearrangement; a time-resolved analysis of hydration-shell overlap near the separatrix would test whether this is a dynamic bottleneck."],"forward_implications":["For alanine dipeptide, the RC is dominated by the dihedral θ (with φ), not ψ; the θ contribution sharpens near the TS, predicting a tilted separatrix line on the (φ,θ) free-energy surface at p*_B = 0.5.","For NaCl in water, the SHAP-identified ACSFs G5_58 and G5_1217, together with r_ion, are sufficient to build a 2D PMF with a well-defined TS line; these descriptors correlate with the physical water-bridging variables ρ and N_B, connecting abstract features to mechanism.","Because different DNN architectures (depth, width, regularization) produce nearly identical RCs and feature rankings, the identified mechanism is not an artifact of a particular trained model.","The framework extends likelihood-maximization RC methods by replacing a linear parametric ansatz with a flexible nonlinear map, while the XAI step recovers interpretability lost in the nonlinearity.","The review's protocol — sample, train on committor, explain with XAI — is offered as transferable to other rare-event systems (nucleation, protein conformational change) where the RC is unknown."],"fun_headline_variants":["Explainable AI reveals the coordinates that define molecular reactions","Deep learning and XAI pinpoint the variables driving reactions","AI identifies the key collective variables behind molecular transitions","Neural networks and explainable AI decode reaction coordinates"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework rests on the committor estimates themselves: 100 one-picosecond trajectories with random thermal velocities must yield converged, unbiased p*_B labels for every sampled configuration; if those labels are noisy or biased, the learned RC and the XAI rankings inherit the error — and the NaCl results show that near the transition state the labels are widely scattered rather than sharply centered at 1/2.","fun_headline_variants_meta":{"raw":{"variants":["Explainable AI reveals the coordinates that define molecular reactions","Deep learning and XAI pinpoint the variables driving reactions","AI identifies the key collective variables behind molecular transitions","Neural networks and explainable AI decode reaction coordinates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000397,"raw_usage":{"total_tokens":1916,"prompt_tokens":747,"completion_tokens":1169,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1107}},"tokens_in":491,"tokens_out":1169,"duration_ms":11505,"temperature":1.0,"reasoning_tokens":1107,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:22:42.980314+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute committors for a few hundred configurations near q=0 using 10,000 instead of 100 velocity assignments and compare the p*_B distribution. If it does not sharpen into a peak at 1/2 — for NaCl, the paper already shows a wide 0-to-1 spread — the learned RC is not truly separating transition states. A second check: shoot trajectories from configurations lying on the claimed separatrix line of the (r_ion, G5_58) PMF; measure how often they commit to A vs B. An isocommittor surface would give 50/50 outcomes.","supporting_citations":[],"review_version":1}