REVIEW 4 major objections 5 minor 1 cited by
From Atoms to Dynamics: Learning the Committor Without Collective Variables
T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read A graph neural network learns the committor directly from atomic coordinates, without hand-crafted collective variables.
desk verdict GVP-GNN committor learning is a promising and partly validated idea, but the paper overstates CV-freedom and under-reports bias-stationarity checks. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (4)
- [Methods, Eq. (4)] 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.
- [Discussion, final paragraph; SI Section 4] 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).
- [Methods, Eqs. (1)-(3); SI Section 1.1] 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.
- [Methods, Algorithm 1] 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.
minor comments (5)
- [Figures S5, S6] 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.
- [References] 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.
- [SI Section 4, Fig. S7] 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.
- [Abstract; Table 1] 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.
- [Data and Code Availability] 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.
Circularity Check
No significant circularity: the variational loss is derived from transition path theory, and validation uses external shooting, VCN maps, and published rates.
full rationale
The derivation chain is not circular. The qGNN loss (Eq. 2) uses the committor variational principle C[q;tau] = <(q(tau)-q(0))^2>/2; SI Section 1.1 derives this principle from the TPT flux expression and shows that minimizing JAB recovers Eq. S2, the formal definition of the committor. The loss is therefore not defined in terms of the target committor, and the method does not fit committor labels; it optimizes a physical functional. The reweighting formula Eq. 4 is a standard static-bias identity, and the paper's assumption that it applies to the converged portion of WTM-eABF runs is an input assumption, not a circular reduction. Comparisons with NANMA shooting (OpenPathSampling), with VCN maps (Ref. 17), and with published rate constants (Refs. 17, 36, 39, 40, 32, 45) are external benchmarks that are not fitted. The per-system boundary weight lambda is tuned, but SI Section 3 reports stable results over a wide range, except for extreme lambda ~ 1e5. The only flagged concern is SI Section 4's warning that 'adding CV-dependent biases during the simulations can influence the model's prediction of the primary degrees of freedom' and the Discussion's admission that 'to some extent, the learning rests upon our choice of the CVs used for biased sampling.' This weakens the 'CV-free' claim as a correctness or limitation issue, but it does not make any equation reduce to its input by construction: the learned committor is not equal to the bias CVs or to any fitted parameter. Accordingly, no circular step is exhibited and the score is 0.
Assumptions & free parameters
free parameters (2)
- lambda (boundary weight in Eq. 2) =
not stated per system; SI tests 0.5, 10, 100, 10000
- tau (time lag in variational loss) =
not given explicitly; SI linear-regime fits span e.g.
assumptions (5)
- standard math The committor variational principle: minimizing C[q;tau] = 1/2 <(q(tau)-q(0))^2> with boundary conditions q=0 in A and q=1 in B recovers the true committor.
- domain assumption The reweighting formula (Eq. 4) for biased time-correlation functions is valid for WTM-eABF biased trajectories.
- domain assumption The molecular dynamics are Markovian on the tau timescale for each system.
- ad hoc to paper The graph constructed from Cartesian coordinates preserves all information relevant to the committor; the edge set is not detailed.
- standard math GVP-GNN can approximate the committor with the chosen hyperparameters.
Cite this review
Pith. "Pith review of From Atoms to Dynamics: Learning the Committor Without Collective Variables." pith.science (2026). https://pith.science/paper/UT75IZJC
@misc{pith2026250717700,
author = {Pith},
title = {Pith review of: From Atoms to Dynamics: Learning the Committor Without Collective Variables},
year = {2026},
howpublished = {\url{https://pith.science/paper/UT75IZJC}},
note = {Machine review of arXiv:2507.17700}
}
read the original abstract
This Brief Communication introduces a graph-neural-network architecture built on geometric vector perceptrons to predict the committor function directly from atomic coordinates, bypassing the need for hand-crafted collective variables (CVs). The method offers atom-level interpretability, pinpointing the key atomic players in complex transitions without relying on prior assumptions. Applied across diverse molecular systems, the method accurately infers the committor function and highlights the importance of each heavy atom in the transition mechanism. It also yields precise estimates of the rate constants for the underlying processes. The proposed approach opens new avenues for understanding and modeling complex dynamics, by enabling CV-free learning and automated identification of physically meaningful reaction coordinates of complex molecular processes.
Forward citations
Cited by 1 Pith paper
-
Accelerated descriptor-free path sampling for protein-ligand binding kinetics
Accelerated AIMMD, combining a descriptor-free PaiNN committor with a basin-restricted OPES bias, recovers protein–ligand unbinding rates within a small factor of experiment or unbiased-MD references, whereas standard...
Reference graph
Works this paper leans on
-
[1]
Roux, B.: Transition rate theory, spectral analysis, and reactive paths. J. Chem. Phys. 156(13), 134111 (2022) https://doi.org/10.1063/5.0084209
-
[2]
Chen, H., Roux, B., Chipot, C.: Discovering reaction pathways, slow variables, and committor probabilities with machine learning. J. Chem. Theory Comput.19(14), 4414– 4426 (2023) https://doi.org/10.1021/acs.jctc.3c00028
-
[3]
Meg ´ıas, A., Contreras Arredondo, S., Chen, C.G., Tang, C., Roux, B., Chipot, C.: Itera- tive variational learning of committor-consistent transition pathways using artificial neu- ral networks. Nat. Comput. Sci. (2025) https://doi.org/10.1038/s43588-025-00828-3
-
[4]
In: International Conference on Learning Representations (2021)
Jing, B., Eismann, S., Suriana, P., Townshend, R.J.L., Dror, R.: Learning from protein structure with geometric vector perceptrons. In: International Conference on Learning Representations (2021). https://doi.org/10.48550/arXiv.2009.01411
-
[5]
In: Proceedings of the 38th International Conference on Machine Learning (PMLR) (2021)
Satorras, V .G., Hoogeboom, E., Welling, M.: E(n) equivariant graph neural networks. In: Proceedings of the 38th International Conference on Machine Learning (PMLR) (2021). https://doi.org/10.48550/arXiv.2102.09844
-
[6]
In: Proceedings of the 38 Th International Conference on Machine Learning (PMLR) (2021)
Sch ¨utt, K.T., Unke, O.T., Gastegger, M.: Equivariant message passing for the prediction of tensorial properties and molecular spectra. In: Proceedings of the 38 Th International Conference on Machine Learning (PMLR) (2021). https://doi.org/10.48550/arXiv.2102. 03150
-
[7]
In: Proceedings of the 38 Th International Conference on Machine Learning (PMLR) (2021)
Jing, B., Eismann, S., Soni, P.N., Dror, R.O.: Equivariant graph neural networks for 3d macromolecular structure. In: Proceedings of the 38 Th International Conference on Machine Learning (PMLR) (2021). https://doi.org/10.48550/arXiv.2106.03843
-
[8]
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: A simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15(56), 1929–1958 (2014) https://doi.org/10.5555/2627435.2670313
arXiv 2014
Show all 18 references
-
[9]
IEEE Trans
Grattarola, D., Zambon, D., Bianchi, F.M., Alippi, C.: Understanding pooling in graph neural networks. IEEE Trans. Neural Netw. Learn. Syst.35(2), 2708–2718 (2024) https: //doi.org/10.1109/TNNLS.2022.3190922
2024
-
[10]
Weiss, G.H.: First passage times for correlated random walks and some generalizations. J. Stat. Phys. 37, 325–330 (1984) https://doi.org/10.1007/BF01011837
1984 doi
-
[11]
H ¨anggi, P., Talkner, P., Borkovec, M.: Reaction-rate theory: fifty years after Kramers. Rev. Mod. Phys. 62(2), 251 (1990) https://doi.org/10.1103/RevModPhys.62.251
1990 doi
-
[12]
Science 334(6055), 517–520 (2011) https://doi.org/10.1126/science.1208351
Lindorff-Larsen, K., Piana, S., Dror, R.O., Shaw, D.E.: How Fast-Folding Proteins Fold. Science 334(6055), 517–520 (2011) https://doi.org/10.1126/science.1208351
2011 doi
-
[13]
Huang, J., Rauscher, S., Nawrocki, G., Ran, T., Feig, M., De Groot, B.L., Grubm ¨uller, 12 H., MacKerell Jr, A.D.: Charmm36m: an improved force field for folded and intrinsi- cally disordered proteins. Nat. Methods 14(1), 71–73 (2017)
2017
-
[14]
Protein Eng
Barua, B., Lin, J.C., Williams, V .D., Kummler, P., Neidigh, J.W., Andersen, N.H.: The trp-cage: optimizing the stability of a globular miniprotein. Protein Eng. Des. Sel.21(3), 171–185 (2008) https://doi.org/10.1093/protein/gzm082
2008 doi
-
[15]
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V ., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V ., et al.: Scikit-learn: Machine learning in python. J. Mach. Learn. Res. 12, 2825–2830 (2011)
2011
-
[16]
Rousseeuw, P.J.: Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 20, 53–65 (1987) https://doi.org/10.1016/ 0377-0427(87)90125-7
1987
-
[17]
Cali ´nski, T., Harabasz, J.: A dendrite method for cluster analysis. Commun. Stat. 3(1), 1–27 (1974) https://doi.org/10.1080/03610927408827101
1974 doi
-
[18]
IEEE Trans
Davies, D.L., Bouldin, D.W.: A cluster separation measure. IEEE Trans. Pattern Anal. Mach. Intell. PAMI-1(2), 224–227 (1979) https://doi.org/10.1109/TPAMI.1979. 4766909 13
1979 doi
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.