acceptodds
Under review as a conference paper at ICLR 2027

Initializing Doob's Lagrangian with the Committor: A Freidlin–Wentzell Perspective on Rare Event Sampling

Abstract

Sampling rare transitions between metastable states is a fundamental challenge in computational chemistry because physically relevant transition paths occur on timescales far exceeding those accessible to direct simulation. Variational path optimization techniques based on minimizing a stochastic path action have been developed for decades, but most of these approaches require propagating the dynamics of multiple replicas of the system of interest. Doob's Lagrangian (DL) casts the problem as the minimization of an action over controlled path measures, with a neural control field and a training objective that does not require simulating the dynamics, but, in practice, it is sensitive to initialization and unstable under a misspecified transit time . In this work, we establish a mathematical equivalence between the DL framework and Freidlin–Wentzell (FW) large-deviation theory, showing that minimizing the Doob action over both the control field and is asymptotically equivalent to computing the quasipotential governing Arrhenius–Kramers transition rates. This equivalence motivates a principled initialization strategy based on the stationary committor function, for which we prove, when a single transition channel dominates, that the initial control lies exponentially close in to the optimal Doob control field and that the joint optimization is stable. On alanine dipeptide and chignolin at finite temperature, the committor initialization combined with the optimization of removes the collapse of the transit time to unphysically small values observed under linear initialization, and recovers physically meaningful transition paths where the original DL method fails.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Variational Transition Path Sampling, Large Deviation Theory, Doob's Lagrangian

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FfUUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper establishes the equivalence between the Doob’s Lagrangian framework and Freidlin–Wentzell (FW) large-deviation theory. Specifically, the paper proves that minimizing the Doob action asymptotically equals computing a quasipotential of Arrhenius–Kramers transition rates, motivating the initialization from a stationary committor function. The evaluation is on alanine dipeptide and Chignolin.

Critical strengths
  • The theory is solid and explains why fixing the transit time T leads to suboptimal minimization of the FW action, motivating the joint optimization of (v,T).
  • The connection to the FW action functional seems novel and is an interesting theoretical contribution that could lead to follow-up ideas and methods.
Critical weaknesses
  • It is questionable whether the training scheme will scale to larger systems, since it requires initialization with a stationary committor function (which may not be well defined for many systems), pretraining a generative model to match the stationary committor function, then minimizing T. Evaluation on larger systems beyond ALD and Chignolin are needed.
  • It is hard to tell whether the method will generalize, since the guarantees assume vanishing noise and a single transition channel. The current molecular experiments run at zero noise and the toy results suggest accuracy degrades at finite noise.
  • All molecular comparisons are limited to DL variants. Classical methods and learned path samplers ([1], [2], [3]) seem to be missing.

[1] Holdijk, L., et al. (2023). Stochastic optimal control for collective variable free sampling of molecular transition paths.

[2] Seong, K., et al. (2025). Transition path sampling with improved off-policy training of diffusion path samplers.

[3] Raja, S., et al. (2025). Action-minimization meets generative modeling: Efficient transition path sampling with the Onsager-Machlup functional.

Questions for authors
  • How does the method scale to larger proteins? How does the cost of committor estimation compare with the cost of DL training?
  • How much does the DL optimization improve over the committor-generated path it is initialized from?
  • How does the method perform with multiple competing transition channels, or at finite noise?
9 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer FfUU (Weakness 1)

We thank the reviewer for their comments. We respond point by point below. In the revised manuscript all modifications are highlighted in blue.

Critical Weaknesses

It is questionable whether the training scheme will scale to larger systems, since it requires initialization with a stationary committor function (which may not be well defined for many systems), pretraining a generative model to match the stationary committor function, then minimizing T. Evaluation on larger systems beyond ALD and Chignolin are needed.

The molecular experiments of the submission reproduce the setting of Du et al. (2024), our primary baseline, in which chignolin is the largest system considered. We extend the evaluation to a larger protein below, after some remarks on the requirements of the pipeline.

The stationary committor exists and is unique for any pair of disjoint states AA and BB. It is the solution of the backward Kolmogorov equation Lq~=0\mathcal L\tilde q=0 with q~=0\tilde q=0 on ∂A\partial A and q~=1\tilde q=1 on ∂B\partial B, and existence and uniqueness follow from the maximum principle for a nondegenerate generator (E and Vanden-Eijnden, 2006), independently of the complexity of the landscape. Theorem 3.2 is stated for this function. In practice the method uses the committor through the algorithm of Mitchell and Rotskoff (2024), which refines an estimate of the committor while sampling the transition region, and from which we take one reactive trajectory from AA to BB. What the initialization needs from the estimate is therefore a trajectory in the dominant reactive channel, and the optimization of the Doob action then refines the path within that channel. The two steps that follow add almost no cost. Fitting the model to the reactive trajectory is a regression on a few tens of frames, and optimizing TT adds no gradient evaluations, since the path is a continuous function of t∈[0,T]t\in[0,T] sampled with a fixed batch size.

We extended our evaluation to the folding of Trp-cage (20 residues, 284 atoms, about twice the size of chignolin), with the folded and unfolded structures of Seong et al. (2025), reported in the new Section 4.4 and Figure 4. The initialization did not become harder with size. The committor algorithm produced a reactive trajectory after 83 iterations, against 729 for chignolin, at a cost of 9.3×1069.3\times10^{6} force evaluations and 64 minutes, the fit took 47 seconds, and the DL optimization took 8.6×1078.6\times10^{7} evaluations and 21.8 hours. The resulting path is free of steric overlaps and has a single interior maximum 445.9 kJ/mol above its minimum, on the same scale as the 395.5 kJ/mol found for chignolin. The original DL with linear initialization and fixed TT, given the same budget, remains trapped in an unphysical path with overlapping atoms and a maximum of 7.8×10167.8\times10^{16} kJ/mol. The cost analysis and further details are given in our reply to Question 1.

AuthorsOpenReview ↗

Rebuttal to Reviewer FfUU (Weakness 2)

It is hard to tell whether the method will generalize, since the guarantees assume vanishing noise and a single transition channel. The current molecular experiments run at zero noise and the toy results suggest accuracy degrades at finite noise.

The vanishing-noise limit in our theory is the standard asymptotic regime of large-deviation theory. In this regime rare-event rates take the Arrhenius-Kramers form k≍e−V(A,B)/ϵk\asymp e^{-V(A,B)/\epsilon}, and reactive trajectories concentrate around the minimizers of the Freidlin-Wentzell action (Freidlin and Wentzell, 1984; Hänggi et al., 1990; Bouchet and Reygner, 2016). This is the setting in which the objective optimized by DL acquires a physical interpretation. Theorem 3.1 identifies the minimum of the Doob action over (v,T)(v,T) with the quasipotential that sets the rate exponent, and this identification motivates both the optimization of TT and the committor initialization. As in Kramers' theory, small noise means small relative to the barrier, so the limit describes the regime in which transitions are rare and path sampling is needed. The theory is stated in the limit ϵ→0\epsilon\to0, while the method itself runs at finite ϵ\epsilon.

All molecular experiments are run at finite temperature, 300 K for alanine dipeptide and Trp-cage and 500 K for chignolin, and the corresponding noise amplitude enters the Doob action that is optimized. The quantity we set to its minimum is the width of the Gaussian marginal. The optimization therefore returns the mean path, which is the most probable transition path at the simulation temperature and the object computed by minimum-action methods such as MAM and gMAM. We made this choice because in Cartesian coordinates small perturbations of the bond lengths produce large changes in energy, so the stochastic estimate of the action is dominated by bond vibrations and training becomes unstable. In order to address this comment we added experiments with non-zero sampling variance on alanine dipeptide, reported in Appendix C.1. With the path parameterized in internal coordinates, that is the full set of 3N−63N-6 bond lengths, angles and dihedrals from which the molecule is reconstructed, the model trains stably at finite width and returns an ensemble of reactive paths, all 64 of which reach BB (Figure 6).

The degradation in the toy results of Appendix B (Table 1) occurs when the noise becomes comparable to the barrier. On the double well, where ΔU=1\Delta U=1, the ratio SDL/V(A,B)S_{\mathrm{DL}}/V(A,B) equals 1.00 for every ϵ≤0.1\epsilon\le0.1, that is ΔU/ϵ≥10\Delta U/\epsilon\ge10. It grows to 1.48 and 3.64 at ϵ=0.5\epsilon=0.5 and ϵ=1\epsilon=1, where ΔU/ϵ≤2\Delta U/\epsilon\le2 and transitions are no longer rare. In that regime the quasipotential, a small-noise quantity, no longer describes the transition, and the Doob action at finite noise departs from it, consistently with Theorem 3.1 being a statement about the limit ϵ→0\epsilon\to0. The molecular systems lie well inside the first regime. For alanine dipeptide at 300 K the barrier of 37.9 kJ/mol gives a ratio ΔU/kBT\Delta U/k_BT of approximately 15, where the double-well ratio is 1.00.

The single-channel condition of Assumption 3.1 is a hypothesis of the convergence guarantee of Theorem 3.2, and the algorithm runs without it. With competing channels a single Gaussian follows the channel in which it is initialized. The Gaussian-mixture parameterization of Du et al. (2024), in which each component is initialized from a distinct committor-based path and carries its own transit time, recovers all of them. We report this extension in the new Appendix D (Figure 8). On a two-dimensional two-channel potential the learned channel weights match a grid reference to within 0.03, and on alanine dipeptide four components retain four distinct channels (see our reply to Question 3).

AuthorsOpenReview ↗

Rebuttal to Reviewer FfUU (Weakness 3)

All molecular comparisons are limited to DL variants. Classical methods and learned path samplers ([1], [2], [3]) seem to be missing.

We added a comparison on alanine dipeptide with the learned samplers suggested by the reviewer, reported in Appendix C.1 (Table 3). We ran TPS-DPS (Seong et al., 2025) and Onsager-Machlup action minimization (OMA, Raja et al., 2025), together with the original Doob's Lagrangian (Du et al., 2024). We could not obtain a working run of PIPS (Holdijk et al., 2023), which failed to produce reactive paths with two different seeds. Each method was run with the alanine dipeptide settings of its paper and repository on a common system, the amber14 force field in vacuum with our endpoint structures. Following Park et al. (2025), we evaluated 64 paths per method through the fraction of paths reaching BB, the RMSD of the final structure to BB, the barrier height along successful paths, and the computation time.

MethodHits BBRMSD (Å)Barrier (kJ/mol)Train (s)Sampling (s)
DL + committor + flexible TT (ours)64/640.0037.9±0.137.9 \pm 0.131,1401.9
DL, linear + fixed TT64/640.00933.2±0.4933.2 \pm 0.47,0591.9
TPS-DPS (seed 2)0/640.79n/a39,226202
TPS-DPS (seed 3)60/640.22125.4±14.1125.4 \pm 14.136,767198
OMA61/640.01107.1±185.8107.1 \pm 185.834,206337,395

Our method has the lowest barrier, every path reaches BB by construction, and the total training time, which includes the committor stage, is below that of TPS-DPS and OMA, while sampling is two to five orders of magnitude faster. It also corrects the failure of the original DL, whose paths reach BB only over a barrier of 933 kJ/mol. The barriers are not strictly comparable across methods, since ours generates configurations directly, TPS-DPS samples biased trajectories that carry thermal noise, and OMA minimizes the energy of every generated frame with L-BFGS. Even so, our paths have energies comparable to those of the minimized OMA structures, whose mean is 59.1±18.059.1 \pm 18.0 kJ/mol once five paths with barriers above 200 kJ/mol are excluded. The first TPS-DPS run produced no reactive path and a second seed was needed, and the final relaxation of OMA diverged for 3 of the 64 paths.

Reviewer_9ud2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper establishes an equivalence between Doob’s Lagrangian (DL) and Freidlin–Wentzell theory in the small-noise limit. Whereas the original DL uses a linear path initialization with a fixed time horizon, the proposed method initializes from a committor-based path and jointly optimizes the path and transit time. Experiments on alanine dipeptide and chignolin show lower energy barriers than those obtained with the original DL.

Critical strengths
  • The paper is generally well-written.
  • Connecting DL to Freidlin–Wentzell theory, which motivates optimizing the transit time T and initializing from the committor, is interesting.
  • The proposed method lowers the estimated energy barriers compared to the DL on molecular systems.
Critical weaknesses
  • In my view, the contribution is narrow for a general machine learning venue. The proposed method seems tailored to [1], while several other ML-based TPS methods have recently been proposed [2–5]. I wonder whether the proposed approach to initializing paths for TPS could also be used to initialize these related methods. In particular, similar path-space initialization may be useful for [4, 5].

  • The performance comparison is limited. For the molecular systems, the baselines include only DL variants, with no comparison to other ML-based TPS methods [2-5].

  • It is unclear what role the zero-noise limit plays in the proposed method. The theory relies on the small-noise limit, while the molecular experiments seems set the noise exactly to zero. As a result, the method returns a single deterministic path rather than an ensemble of transition paths, although the paper claims results at finite temperature. I wonder how the proposed method performs with nonzero noise.


[1] Du et al., Doob's Lagrangian
[2] Holdijk et al., Stochastic optimal control for collective variable free sampling of molecular transition paths.
[3] Seong et al., Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers
[4] Park et al., Functional Adjoint Sampler 
[5] Raja et al., Action-Minimization Meets Generative Modeling
Questions for authors
  • How much computation does it take to obtain the proposed committor-based initial path compared with the DL optimization itself?
  • (Minor point) The font used in the paper appears to differ from that used in the standard ICLR submission template.
6 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer 9ud2 (weakness 1)

We thank the reviewer for their comments, and in particular for the idea of applying our initialization strategy to other methods, which led to a new result. We respond point by point below. In the revised manuscript all modifications are highlighted in blue.

Critical Weaknesses

In my view, the contribution is narrow for a general machine learning venue.

The formulation of rare-event sampling as a problem of optimal control and large-deviation theory is classical, developed over several decades in applied mathematics and scientific computing. The role of ML here is one of representation, as neural networks provide a flexible and scalable parameterization of the functions these formulations require, in our case the control field and the committor. Our contribution is a strategy that uses this representation to optimize a physically meaningful objective over path space at scale, co-designed with the theory that justifies it. In this sense the paper re-examines a family of classical scientific-computing methods through a machine-learning lens.

At bottom the contribution is a result in stochastic analysis. It rests on a connection between three frameworks that are each used very broadly, the control of diffusion processes, Freidlin-Wentzell large-deviation theory, and Doob's h-transform. The h-transform is the general way of turning the conditioning of a diffusion into a control problem, and the quantity it turns on, the committor, is a first-passage probability, the chance of reaching one set before another. Maximizing the probability of reaching a target set is one of the most fundamental problems in the theory of Markov processes, and it can be posed for any diffusion with no recourse to physical dynamics at all. Our two results are statements about this general framework. The equivalence between Doob's Lagrangian and Freidlin-Wentzell theory (Lemma 3.1, Theorem 3.1, Corollary 3.1) shows that minimizing the Doob action over the control field and the transit time (v,T)(v,T) recovers the quasipotential and the Arrhenius-Kramers rate exponent (Hänggi et al., 1990). This places a learned generative sampler in the same family as classical minimum-action methods (E et al., 2004; Heymann and Vanden-Eijnden, 2008), while keeping it differentiable and applicable in high dimension. Theorem 3.2 is, to our knowledge, the first provable initialization and convergence guarantee for a neural sampler of this kind. A learned stationary committor (Rotskoff et al., 2022; Mitchell and Rotskoff, 2024) places the optimizer exponentially close, in TT, to the optimal control, with local convexity ensuring stable joint (v,T)(v,T) convergence. The advance is the connection itself, together with a provably initialized method for the generic problem these frameworks share.

There is direct precedent for contributions of this kind at this and comparable venues. The method we build on appeared at NeurIPS (Du et al., 2024), and learned samplers for the same problem at ICLR and ICML (Seong et al., 2025; Raja et al., 2025). The object we optimize, a learned control over a path measure, is also the one at the center of score-based diffusion, Schrödinger bridges and neural samplers (Song et al., 2021; de Bortoli et al., 2021; Vargas et al., 2023), so an initialization and convergence theory for this class of objectives is relevant beyond molecular systems. As we show in our next reply, the initialization derived from the theory already improves a method other than our own.

AuthorsOpenReview ↗

Rebuttal to Reviewer 9ud2 (weakness 2)

The proposed method seems tailored to [1], while several other ML-based TPS methods have recently been proposed [2–5]. I wonder whether the proposed approach to initializing paths for TPS could also be used to initialize these related methods. In particular, similar path-space initialization may be useful for [4, 5].

Our theory builds on the Doob's Lagrangian framework of Du et al. (2024), and we present our method as a theoretically grounded and substantial improvement. The two theoretical results that motivate the initialization are statements about the transition problem itself and do not depend on the DL parameterization. Theorem 3.1 identifies the minimizer of the action over paths and transit times with the Freidlin-Wentzell instanton, and Theorem 3.2 shows that the committor-based control v~=∇log⁡q~\tilde v=\nabla\log\tilde q lies exponentially close to the optimal control of the conditioned process. Both objects are defined by the dynamics and the two states, so any method that optimizes a path or a control for the same problem has the same target and can start from the same initialization.

We thank the reviewer for their suggestion to consider the initialization of Raja et al. (2025). Their truncated Onsager-Machlup action, written in continuous time for a gradient drift, is 14D∫0T(∣x˙∣2+ζ−2∣∇U∣2)dt\frac{1}{4D}\int_0^{T}\big(|\dot x|^2+\zeta^{-2}|\nabla U|^2\big)dt, which equals 14D∫0T∣x˙+ζ−1∇U∣2dt\frac{1}{4D}\int_0^{T}|\dot x+\zeta^{-1}\nabla U|^2dt up to a term that depends only on the endpoints. This is the Freidlin-Wentzell action at a fixed horizon TT. Their optimization is a gradient descent over a discretized path, which converges to the local minimizer in whose basin the initial path lies, and the authors observe that the choice of the initial path is crucial for the quality of the result. A committor-based path selects the channel that carries the dominant reactive flux, so it is a principled initial guess for that method.

We tested this transfer directly, and report it in the new Appendix C.4 (Table 6). We ran the Onsager-Machlup action minimization of Raja et al. (2025) on alanine dipeptide for 8 paths, initialized either from a straight line between the endpoints or from our committor-based path, with the same seed and all other settings unchanged.

OMA initial guessHits BBMedian barrier after action minimization (kJ/mol)Median barrier after final relaxation (kJ/mol)
Straight line8/8582168.7
Committor-based path8/858.743.3

All paths reach BB in both cases. With the committor-based initialization the median barrier at the end of the action-minimization stage is 58.7 kJ/mol, against 5821 kJ/mol from the straight line, and after the final relaxation of the method it is 43.3 kJ/mol against 68.7 kJ/mol. The results are also more uniform. From the straight line, two of the eight final paths remain at 334.8 and 548.7 kJ/mol, whereas from the committor-based path seven lie between 40.1 and 43.9 kJ/mol and one at 60.7 kJ/mol. The initialization motivated by our theory therefore improves a method other than the one it was derived for.

Our analysis also indicates that the horizon should be optimized together with the path, since the minimum over TT is what recovers the quasipotential (Theorem 3.1), whereas their horizon is a fixed parameter. To our knowledge the same holds for the other learned samplers we compare with, which prescribe the duration of the trajectory in advance (1 ps for TPS-DPS and 0.3 ps for OMA in our comparison, against a learned duration of 14.2 fs). Optimizing TT together with the path removes a hyperparameter whose misspecification forces the path across higher barriers.

For the methods that learn a bias force along simulated trajectories (Holdijk et al., 2023; Seong et al., 2025), the natural transfer is at the level of the control. The optimal bias for a bridge between two metastable states is expressed through the committor (Singh and Limmer, 2024), and Theorem 3.2 quantifies how close the stationary committor control is to it, so a learned committor could be used to initialize the bias force before training. We have not yet tested the transfer to these bias-force methods, which we regard as a direction opened by the theory.

AuthorsOpenReview ↗

Rebuttal to Reviewer 9ud2 (weakness 3)

It is unclear what role the zero-noise limit plays in the proposed method. The theory relies on the small-noise limit, while the molecular experiments seems set the noise exactly to zero. As a result, the method returns a single deterministic path rather than an ensemble of transition paths, although the paper claims results at finite temperature. I wonder how the proposed method performs with nonzero noise.

The molecular experiments are run at finite temperature, and we acknowledge that the wording "zero noise" in the submission is ambiguous. Two distinct noises enter the method, the thermal noise of the dynamics and the width of the variational Gaussian marginals, and the small-noise limit of the theory concerns the former. This limit is the standard asymptotic regime of large-deviation theory, in which rare-event rates take the Arrhenius-Kramers form k≍e−V(A,B)/ϵk\asymp e^{-V(A,B)/\epsilon} and reactive trajectories concentrate around the minimizers of the Freidlin-Wentzell action (Freidlin and Wentzell, 1984; Hänggi et al., 1990). As in Kramers' theory, small noise means small relative to the barrier. Theorem 3.1 identifies the minimum of the Doob action over (v,T)(v,T) with the quasipotential that sets the rate exponent, and this identification is what motivates the optimization of TT and the committor initialization. The theory is stated in the limit ϵ→0\epsilon\to0, while the method itself runs at finite ϵ\epsilon.

In the experiments the temperature is 300 K for alanine dipeptide and Trp-cage and 500 K for chignolin, and the corresponding noise amplitude enters the Doob action that is optimized. For alanine dipeptide at 300 K the barrier of 37.9 kJ/mol corresponds to approximately 15 kBTk_BT, well inside the regime described by the theory. The quantity we set to its minimum is the width of the Gaussian marginal, and the reviewer is correct that the method then returns a single path. This path is the mean path, which is the most probable transition path at the simulation temperature and the object computed by minimum-action methods such as MAM and gMAM (E et al., 2004; Heymann and Vanden-Eijnden, 2008). We made this choice because in Cartesian coordinates a finite width perturbs bond lengths and angles, so the stochastic estimate of the action is dominated by bond vibrations and training becomes unstable.

In order to address the performance at non-zero noise we added new experiments on alanine dipeptide, reported in Appendix C.1. We reparameterized the path in internal coordinates, that is the full set of 3N−63N-6 bond lengths, angles and dihedrals from which the molecule is reconstructed. With this parameterization the model trains stably at finite width, with the same committor-based initialization and trainable TT, and returns an ensemble of reactive paths. All 64 sampled paths reach BB, and they form a tube around the channel selected by the committor-based path (Figure 6). Paths through distinct channels are obtained with a Gaussian-mixture parameterization, which retains four channels on alanine dipeptide (Appendix D, Figure 8). The two-dimensional experiments are also run at finite noise. On the double well the action recovered by our method matches the quasipotential for every ϵ≤0.1\epsilon\le0.1, that is for barriers of at least 10 ϵ\epsilon (Appendix B, Table 1), and departs from it only when the noise becomes comparable to the barrier and transitions are no longer rare. We revised the wording on noise in the manuscript to remove the ambiguity.

Reviewer_PccqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is a modification of the Doob’s Lagrangian for transition path sampling, and in particular proposes a construction initialization for the action optimization method. The main theoretical result is connecting the Doob action to the Freidlin–Wentzell (FW) action, which yield theoretical motivation to initialize the molecular path from the committor function. Evaluation covers standard environments: Muller-Brown, alanine dipeptide, and chignolin.

Critical strengths
  • The presentation, in a field where mathematical presentation can vary, is clear and straightforward.
  • The environments evaluated on are standard in the field, allowing for ease of comparison for other researchers evaluating on these environments.
  • To my understanding of the paper's math, the math is all correct.
Critical weaknesses
  • There are many methods since the Doobs Lagrangian paper that propose methods for transition path sampling with similar baseline requirements, eg [1][2][3]. These need to be compared against fairly in their ability to find diverse, viable transition paths.
Questions for authors

Have you compared your method against other learning-based transition path sampling baselines, such as [1][2][3]? What are the standout benefits of this particular method (initialization+DB) that makes this uniquely useful compared to previous methods in the area?

4 comments
AuthorsOpenReview ↗

Rebuttal to Rev. Pccq (Weaknesses)

We thank the reviewer for their careful assessment of our paper and for their positive comments on its presentation and on the correctness of the mathematics. We respond point by point below. In the revised manuscript all modifications are highlighted in blue.

Critical Weaknesses

There are many methods since the Doobs Lagrangian paper that propose methods for transition path sampling with similar baseline requirements, eg [1][2][3]. These need to be compared against fairly in their ability to find diverse, viable transition paths.

We added a comparison with learning-based samplers on alanine dipeptide, reported in Appendix C.1 (Table 3). Of the methods suggested by the reviewer, we ran Onsager-Machlup action minimization (OMA, Raja et al., 2025), we could not obtain a working run of PIPS (Holdijk et al., 2023), which failed to produce reactive paths with two different seeds, and we could not find published code for the method of Lee et al. (2025). We added TPS-DPS (Seong et al., 2025) as a further baseline, together with the original Doob's Lagrangian (Du et al., 2024). To make the comparison fair, each method was run with the alanine dipeptide settings of its own paper and repository, on a common system, the amber14 force field in vacuum with the same endpoint structures. Following Park et al. (2025), each method generated 64 paths, which we evaluated through the fraction of paths that reach the target state BB, the RMSD of the final structure to BB, the barrier height along the successful paths, and the computation time.

MethodHits BBRMSD (Å)Barrier (kJ/mol)Train (s)Sampling (s)
DL + committor + flexible TT (ours)64/640.0037.9±0.137.9 \pm 0.131,1401.9
DL, linear + fixed TT64/640.00933.2±0.4933.2 \pm 0.47,0591.9
TPS-DPS (seed 2)0/640.79n/a39,226202
TPS-DPS (seed 3)60/640.22125.4±14.1125.4 \pm 14.136,767198
OMA61/640.01107.1±185.8107.1 \pm 185.834,206337,395

On viability, every path generated by our method reaches BB, and its barrier of 37.9 kJ/mol is the lowest of all methods. The original DL also reaches BB, but only over a barrier of 933 kJ/mol, which our two modifications correct. The first TPS-DPS run produced no reactive path, and a second seed, as recommended in its repository, gave 60 of 64 paths with a barrier of 125.4 kJ/mol. OMA fixes the endpoints by construction, but its final relaxation diverged for 3 of the 64 paths, and its mean barrier of 107.1 kJ/mol is inflated by five paths that remain above 200 kJ/mol. The total training time of our method, which includes the committor stage, is below that of TPS-DPS and OMA, and sampling is two to five orders of magnitude faster, because a path is obtained by evaluating the trained model, with no simulation and no relaxation. The barriers are not strictly comparable across methods, since ours generates configurations directly, TPS-DPS samples biased trajectories that carry thermal noise, and OMA minimizes the energy of every generated frame with L-BFGS. Even so, our paths have energies comparable to those of the minimized OMA structures, whose mean is 59.1±18.059.1 \pm 18.0 kJ/mol once those five paths are excluded.

On diversity, the paths generated by our method in this comparison are nearly identical, because the model is trained on the mean path and follows a single channel. We extended the method in two directions to address this. With the path parameterized in internal coordinates, that is the full set of 3N−63N-6 bond lengths, angles and dihedrals from which the molecule is reconstructed, the model trains at non-zero sampling variance and returns an ensemble of reactive paths, all 64 of which reach BB (Appendix C.1, Figure 6). The mean path of this model has a barrier of 45.4 kJ/mol, close to the 37.9 kJ/mol obtained in Cartesian coordinates. With a Gaussian-mixture parameterization, in which each component is initialized from a different committor-based path, the model recovers distinct reaction channels. On alanine dipeptide it retains four channels, with barriers of 66.01, 67.79, 63.58 and 39.41 kJ/mol, and on a two-dimensional potential with two channels the learned channel weights match a reference computed on a grid to within 0.03 (Appendix D, Figure 8).

AuthorsOpenReview ↗

Rebuttal to Rev. Pccq (Questions)

Questions for Authors

Have you compared your method against other learning-based transition path sampling baselines, such as [1][2][3]? What are the standout benefits of this particular method (initialization+DB) that makes this uniquely useful compared to previous methods in the area?

The comparison is described in our reply above (Appendix C.1, Table 3), where our method reaches the lowest barrier at a comparable training cost. Beyond these results, the following properties distinguish it from the other learned samplers.

The method comes with a theoretical guarantee. Theorem 3.2 shows that the committor-based initialization places the optimizer exponentially close, in TT, to the optimal control, in a region where the objective is locally convex. To our knowledge, none of the other learned samplers has an initialization or convergence guarantee of this kind. Theorem 3.1 moreover gives the objective a physical meaning, since its minimum over the path and the transit time is the quasipotential that sets the transition rate.

The transit time does not have to be supplied in advance. In our method TT is optimized together with the path and is an output of the model, whereas, to our knowledge, the other methods prescribe the duration of the trajectory, 1 ps for TPS-DPS and 0.3 ps for OMA in our comparison, against a learned duration of 14.2 fs. A misspecified duration is what forces the original DL across a barrier of 933 kJ/mol.

The initialization is useful beyond our own method. Used as the initial guess of OMA in place of a straight line between the endpoints, the committor-based path lowers the median barrier of that method from 68.7 to 43.3 kJ/mol over 8 paths (Appendix C.4, Table 6).

These properties carry over to a larger system. On Trp-cage, a protein of 20 residues added in the revision (Section 4.4, Figure 4), the method finds a path free of steric overlaps, with a single interior maximum 445.9 kJ/mol above its minimum. The original DL, given the same budget, remains trapped in an unphysical path with overlapping atoms and a maximum of 7.8×10167.8\times10^{16} kJ/mol, and the committor stage accounts for about 11% of the force evaluations of the optimization (Appendix C.3, Table 4).

References

Du, Y., Plainer, M., Brekelmans, R., Duan, C., Noé, F., Gomes, C. P., Aspuru-Guzik, A., and Neklyudov, K. (2024). Doob's Lagrangian: A Sample-Efficient Variational Approach to Transition Path Sampling. NeurIPS.

Holdijk, L., Du, Y., Hooft, F., Jaini, P., Ensing, B., and Welling, M. (2023). Stochastic Optimal Control for Collective Variable Free Sampling of Molecular Transition Paths. NeurIPS.

Lee, J., Plainer, M., Du, Y., Holdijk, L., Brekelmans, R., Beaini, D., and Neklyudov, K. (2025). Scaling Deep Learning Solutions for Transition Path Sampling. ICLR 2025 Workshop on Frontiers in Probabilistic Inference.

Park, B., Lee, J., and Liu, G.-H. (2025). Functional Adjoint Sampler: Scalable Sampling on Infinite Dimensional Spaces. arXiv:2511.06239.

Raja, S., Šípka, M., Psenka, M., Kreiman, T., Pavelka, M., and Krishnapriyan, A. S. (2025). Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional. ICML.

Seong, K., Park, S., Kim, S., Kim, W. Y., and Ahn, S. (2025). Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers. ICLR.

Thank you for the detailed rebuttal. After reviewing the new results, I am happy to raise my score.

We kindly thank the reviewer for their time and for acknowledging our response.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the reviewers and AC for their time and for their constructive comments. We appreciated that the reviews recognized the clarity of the presentation and the theoretical soundness of our work. We updated the manuscript with revisions highlighted in blue ink.

We summarize the main changes here and reply to each point in the individual responses.

We improved the paper by adding a comparison of our method with other learning-based path samplers on alanine dipeptide, where it reached the lowest barrier, 37.9 kJ/mol against more than 100 kJ/mol for the other methods, with every path reaching the target state and a comparable training time (Appendix C.1, Table 3).

We also ran experiments on Trp-cage, a larger protein, about twice the size of chignolin, which we added as a new section of the main text. In that experiment our method found a physical path, free of steric overlaps and with a single interior maximum 445.9 kJ/mol above its minimum, while the original algorithm remained at energies of order 101610^{16} kJ/mol, and the committor stage accounted for about 11% of the force evaluations of the optimization (Section 4.4, Figure 4, Appendix C.3).

We further showed that the method runs at non-zero sampling variance and returns an ensemble of reactive paths when the path is parameterized in internal coordinates (Appendix C.1, Figure 6).

We extended the method to landscapes with competing reaction channels through a Gaussian mixture, which recovered all channels on a two-dimensional potential and on alanine dipeptide (Appendix D, Figure 8).

Finally, following the reviewers' suggestion, we applied the committor-based initialization to the Onsager-Machlup action minimization of Raja et al. (2025), where it lowered the median barrier from 68.7 to 43.3 kJ/mol (Appendix C.4, Table 6). This confirmed that the initialization derived from our theory is useful beyond the method for which it was developed. Our method still reached the lowest barrier, 37.9 kJ/mol, with the transit time learned and without the energy minimization of every frame on which the Onsager-Machlup result relies.

The individual responses also clarify three points that do not require new experiments, namely the role of the small-noise limit in the theory as opposed to the finite temperature of the experiments, the cost of the committor relative to the optimization, and the scope of the theoretical results, which hold for a general diffusion and are relevant beyond molecular systems.

We hope these additions clarify the reviewers' concerns, and we are available for further clarification during the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.