acceptodds
Under review as a conference paper at ICLR 2027

Transition Path Sampling Using Koopman Operators and Exit-Time Optimal Control

Abstract

Sampling transitions between metastable states is a central problem in dynamical systems theory and molecular dynamics in particular. A key challenge is the existence of high free-energy barriers that separate the states, making transitions extremely rare. Recent machine learning-based methods cast transition path sampling (TPS) as optimal stochastic control (OSC) over a fixed time horizon, and parameterize the drift bias with a neural network trained by simulation-in-the-loop, where training requires repeated biased rollouts. We instead develop a new operator-theoretic approach for the problem based on the Koopman operator. Because Koopman operators are linear, their leading eigenfunctions reveal the metastable sets and provide an estimate of the committor function with no transition path information required. We formulate TPS as an OSC problem up to an exit time. Our time horizon is the first hitting time of the target set, and our running cost penalizes time spent in nonreactive regions by encoding the estimated committor function. We derive the optimal controller in closed form, and approximate it in a reproducing kernel Hilbert space (RKHS). This reduces the problem of constructing the optimal controller to solving a single equality-constrained convex quadratic program, whose solution is characterized by a linear KKT system. On the two-channel double well and alanine dipeptide, our controller increases the fraction of trajectories reaching the target from to within steps, and from to within , respectively.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion processes, transition path sampling, optimal stochastic control, Koopman operator

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_B7bcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Author propose an exit-time OSC approach for transition-path sampling with Koopman eigenfunctons and Gibbs measure.

Critical strengths

The exit-time formulation accommodates variable transition durations without prescribing a terminal time in the control objective.

Critical weaknesses
  1. Scalability beyond synthetic systems and alanine dipeptides remains unestablished.
  2. High target-hit rates alone do not establish accurate pathway probabilities or kinetics. Quantitative validation of the reweighted path ensembles for the physical interpretation.
Questions for authors
  1. How rare are the benchmark transitions under unbiased dynamics? Could you report mean first-passage times or transition rates, rather than only target-hit percentages at the selected TmaxT_\text{max}?
  2. The TPS-DPS release includes larger systems such as Trp-cage and chignolin. Consider expanding the evaluation systems beyond alanine dipeptide.
  3. How sensitive are the estimated committor and controller to the kernel? Have you compared other smooth kernels?
  4. What quantitative criterion selects mm in Section 4.1, especially when theres no clear spectral gap? How sensitive are the identified metastable cores and sampling results to this choice?
  5. How rare are the benchmark transitions under unbiased dynamics? H0=∫0τB(1−q(Xt))dtH_0 = \int_{0}^{\tau_B}(1-q(X_t))dt the equation penalizes residence time according to the committor, not state rarity. High target-hit rates do not establish coverage of rare intermediates, and alanine-dipeptide experiments do not address this.
4 comments
AuthorsOpenReview ↗

Response to Reviewer B7bc

We thank the reviewer for your time and comments. Changes in the revised manuscript are highlighted in blue.

Response to Critical Weakness 1: We sincerely agree with the reviewer that applying transition path sampling (TPS) to large biomolecules is an important problem, and most certainly a long-term goal of ours. At the same time, we want to stress that our work is mainly theoretical and meant to lay a new foundation for studying TPS through the use of Koopman operators. Our numerical experiments are there to check these results on systems where the ground truth is known and where many other methods have already reported solutions. We did not want to rush the simulations for chignolin, for example, since it would require a much more in-depth examination of folding/hairpin formation/competing folding results, and we mostly focused on deriving new theoretical claims as suggested in your W2 below.

Response to Critical Weakness 2: We have added new results: Theorem 2 in Section 4.3 shows that, for the approximate controller actually used, the weight (23) recovers the expectation under the original dynamics of every path statistic. Corollary 2 gives the transition probability by time tt (kinetics) and Corollary 3 the fraction of transitions through a channel or intermediate channel (pathway probabilities).

AuthorsOpenReview ↗

Response to Reviewer B7bc, continued

Response to Q1: We now report, for each benchmark, the mean first-passage time (MFPT) from the source to the target core, estimated from long unbiased simulations, together with the relaxation time 1/∣λ1∣1/|\lambda_1| of the slowest Koopman mode. For the two-channel double well at 1200 K, MFPT =(5.3±0.2)×104=(5.3\pm0.2)\times10^{4} and 1/∣λ1∣=2×1031/|\lambda_1|=2\times10^{3}, against Tmax⁡=10T_{\max}=10. For M"uller--Brown at 200 K, MFPT =390±26=390\pm26 ps and 1/∣λ1∣=82.51/|\lambda_1|=82.5 ps, against Tmax⁡=10T_{\max}=10 ps. For the four-well system at 200 K, MFPT =140±6=140\pm6 ps and 1/∣λ1∣=43.31/|\lambda_1|=43.3 ps, against Tmax⁡=10T_{\max}=10 ps. For alanine dipeptide at 300 K, no transition was observed in 100 unbiased runs of 5 ns each (500 ns in total), so the MFPT exceeds 5 ns, and 1/∣λ1∣=2.391/|\lambda_1|=2.39 ns, compared with Tmax⁡=1T_{\max}=1 ps. In all cases, the MFPT exceeds Tmax⁡T_{\max} by one to four orders of magnitude, so transitions within Tmax⁡T_{\max} are rare under the unbiased dynamics. We added these values to Appendix G.1–G.4.

Response to Q2: Please see our response to W1.

Response to Q3: The kernel must be twice continuously differentiable, since LL acts on it through (71). Kernel estimates of the leading Koopman eigenfunctions, on which our metastable sets and committor are built, are known to be accurate over a range of bandwidths, including for alanine dipeptide (Nüske & Klus, 2023). A more accurate kernel approximation yields a more effective controller, while the statistics of the original dynamics are recovered exactly by Theorem 2 in either case.

Response to Q4: We select mm at the spectral gap separating the slow eigenvalues 0=λ0>λ1≥⋯≥λm−10=\lambda_0>\lambda_1\ge\dots\ge\lambda_{m-1} from the rest, i.e., at the largest ratio of consecutive implied timescales tm−1/tmt_{m-1}/t_m, tk=1/∣λk∣t_k=1/|\lambda_k|. Since we consider metastable systems, such a separation is expected. Regarding sensitivity, in all our benchmarks, the spectral gap is clear, so mm is determined by the spectrum; a missing gap would indicate an inadequate estimate of the eigenpairs, and we increase the number of samples or adjust the kernel bandwidth until the leading eigenvalues stabilize.

Response to Q5: We have added new results in Section 4.3 and Appendix F, which show that no intermediate is lost. Specifically, by our new results (Theorem 2 and Corollary 2-3), any intermediate visited with positive probability by the original dynamics is visited with positive probability by the controlled dynamics, and its frequency is recovered exactly via Corollary 3.

Appreciate the acknowledgement of the scale of the system. The authors addressed my concerns. I will raise my score accordingly. However, next time, try not to post the rebuttal so close to the ddl.

Thank you very much for your helpful feedback and for raising your score. We apologize again for the late submission, and thank you for your patience.

Reviewer_DyaXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors consider the problem of transition path sampling (TPS), i.e., sampling transitions between metastable states that can be separated by large barriers. They formulate TPS as an optimal control problem until hitting time with a committor-based cost, and show that the optimal controller is ∇log⁡ρ\nabla\log\rho, where ρ\rho solves a linear elliptic PDE. They approximate this controller in three steps:

  • Eigendecomposition of the Koopman generator, followed by clustering in the slow eigenmodes to identify metastable regions.
  • Kernel approximations of the committor (using the Koopman eigenfunctions) and then of ρ\rho, from which they construct the control
  • Simulating the controlled dynamics until the paths reach the target.

A result shows how statistics of the original process can be derived by reweighting the controlled process. The method is validated on several examples.

Critical strengths
  • The mathematical foundation is rigorous and combines several interesting ideas from Koopman theory, control theory, TPS. Formulating control until first-hitting-time removes the arbitrary time horizon of previous methods and gives a time-independent closed-form controller built from interpretable quantities (metastable sets, committor).
  • From a practical perspective, the procedure seems to offer very fast TPS, the main argument for acceptance.
Critical weaknesses
  • The control results in Sec. 3 are standard, and hitting-time optimal control has been used for rare events before (e.g., Hartmann & Schütte, J. Stat. Mech. 2012; Schütte, Winkelmann & Hartmann, Math. Program. 2012), but this is not is cited. More importantly, the relation to the Doob h-transform is not cited. In the paper's framework, taking f=0f=0, Φ=+∞\Phi=+\infty on ∂Ai\partial A_i and Φ=0\Phi=0 on ∂Aj\partial A_j gives the Doob h-transform ∇log⁡q\nabla\log q, which samples exactly the reactive path ensemble, with no β\beta and no second PDE (Lu & Nolen, Probab. Theory Relat. Fields, 2015; Du et al., NeurIPS 2024). Thus, it is unclear if the paper contributes beyond a kernel implementation of known ideas.
  • The experiments do not show that the right paths are sampled. Target-hit-percentage only measures whether paths reach the target, which can always be achieved with a sufficiently strong bias. There is no comparison with reference statistics and the reweighting Lemma 4 is not tested.
Questions for authors

Clarifications:

  • What is new relative to prior work on exit-time optimal control for rare events in MD (e.g., Hartmann & Schütte 2012; Schütte, Winkelmann & Hartmann 2012), beyond the kernel approximation?
  • In the alanine dipeptide example, the controller is computed from an overdamped model, while the simulation is underdamped; the bias force is also multiplied by a hand-picked factor and capped. Can the authors justify these choices, and how sensitive are the results to them?
  • Eq. (90) is exact only for the optimal controller. How would the reweighting be done with the approximate controller?
  • The TPS-DPS results in Table 1 are taken from Seong et al. (2025), while the target set here is a Koopman core. Are the parameters the same in both cases? Otherwise the THP values are not directly comparable.
  • How are the samples for the eigenproblem generated, and how does the cost of the eigenproblem grow with dimension? Does one need samples in the transition region to resolve the committor, and if so, does this limit the method in high dimensions?
  • In Example 1, D=X∖AjD=X\setminus A_j is unbounded, and so are the drifts, so Assumptions 1 and 3 may not hold. Do the theorems apply to the examples?
  • The kernel approximations do not enforce positivity of ρ^\widehat{\rho}, although the controller involves ∇log⁡ρ^\nabla\log\widehat{\rho}. How is this handled?

Suggestions (results that would affect my assessment):

  • A comparison with the Doob controller ∇log⁡q^\nabla\log\widehat q (Lu & Nolen, Probab. Theory Relat. Fields, 2015; Du et al., "Doob's Lagrangian", NeurIPS 2024), and an explanation of what the β\beta-tilted path ensemble represents physically.
  • The authors should evaluate sampled path statistics in order to validate the quality of their method, not just THP (e.g., against reference statistics in the 2D examples). I also think that, given that the paper is methodological, it would be appropriate to include in the main text a more systematic comparison of the quality and efficiency of their TPS method with several popular alternatives.
  • Ablations against a constant running cost, which simply penalizes the time until exit, to justify the choice f=β(1−q)f=\beta(1-q), and against a direct kernel PDE solution without Koopman eigenfunctions.

Minor comments:

  • The paper is quite formal and dense, especially in Sec. 2 and 3. More intuition, more discussion of the construction in Sec. 4, and more analysis of the experiments would help. The introduction repeats several claims (e.g., about the closed-form controller).
  • Eq. (83) conflicts with Eq. (78) and introduces an undefined η\eta. Several cross-references and assumption numbers are incorrect. Figure 2 reports 99.0% target hits for β=1\beta=1, whereas Table 2 gives this for β=2\beta=2.
  • Typos: "masures" (l. 1210), "Ramachandran plain" (Fig. 4), "10.2pm.6" (Table 4).
5 comments
AuthorsOpenReview ↗

Response to Reviewer DyaX

We thank the reviewer for your time and the detailed, constructive comments. Changes in the revised manuscript are highlighted in blue.

Response to Critical Weakness 1: We have included all suggested references in the revision. We explain how our approach relates to prior work, and what is new, in our response to Q1. Here we address the Doob controller. As the reviewer notes, the Doob controller ∇log⁡q\nabla\log q corresponds to the limiting case Φ=+∞\Phi=+\infty on ∂Ai\partial A_i of our formulation, and it is also the limit of our closed-form factor, since ρ0=q^+(1−q^)r→q^\rho_0=\hat q+(1-\hat q)r\to\hat q as βT0→∞\beta T_0\to\infty in (21). It samples the reactive ensemble exactly only when the exact committor is available. In practice, the committor is approximated by q^\hat q, and keeping β\beta finite has several advantages. First, β\beta controls the trade-off between speed and fidelity to the original dynamics, with β=0\beta=0 recovering the original dynamics, and it has a variational meaning (Example 1 and our response to S1). Moreover, β\beta sets r=1/(1+βT0)r=1/(1+\beta T_0) in the closed-form factor: β=0\beta=0 gives ρ0≡1\rho_0\equiv1, while β→∞\beta\to\infty gives ρ0→q^\rho_0\to\hat q, so that ∇log⁡ρ0\nabla\log\rho_0 tends to the Doob controller. For finite β\beta, our controller is bounded, so paths can start inside the source set. In contrast, in the Doob controller, ∇log⁡q\nabla\log q is singular on ∂Ai\partial A_i. Second, by Theorem 2, we show that paths driven by our controller can be reweighted to recover the statistics of the original dynamics exactly; there is no counterpart of this result in the ``Doob's Lagrangian" (Du et al., NeurIPS 2024). The second PDE indeed adds little cost: it is solved through a single linear system (Appendix E), which takes 51 s on alanine dipeptide (Table 3). Empirically, with the same Koopman committor, the Doob controller reaches a THP of 73.3% versus 93% for ours on alanine dipeptide (see our response to S1 below).

Response to Critical Weakness 2: Beyond THP, the manuscript reports the energy along the paths (ETS, Tables 2 and 4) and the paths themselves (Figure 2). The results there show that a bias that is merely strong produces high-energy paths, as for TPS-DPS on M"uller--Brown, which crosses the high-energy barrier, whereas our paths follow the low-energy channel through the shallow well. To compare with reference statistics directly, we have made further changes. The new Theorem 2 in Section 4.3 shows that, for the controlled process we actually simulate with the approximate controller, the weighted average of any path statistic over the controlled paths equals its expectation under the original dynamics exactly. The new Corollaries 2--3 in Appendix F.3 then give the transition probability and the fraction of transitions through a channel or an intermediate state.

AuthorsOpenReview ↗

Response to Reviewer DyaX, continued

Response to Q1: We would like to thank the reviewer for drawing our attention to relevant literature, and we have incorporated all these references with appropriate attribution of ideas in the introduction. It seems that the connection between sampling and optimal control has been (re)discovered independently by several communities. The earliest work we are aware of, and one we have been building on, is Newton and Mitter (2003). The focus in that work was on path estimation in diffusion processes based on partial observation, but the key mechanism relating the form of the optimal control, the posterior density, and free-energy minimization via the Gibbs variational principle is already there. These ideas are also present in the papers of Hartmann & Sch"utte (2012) and Sch"utte, Winkelmann & Hartmann (2012), as well as in the survey paper of Hartmann, Richter, Sch"utte, and Zhang (2017). None of these works cite Newton and Mitter.

Another important common thread, which we will emphasize, is the fundamental connection to Doob's hh-path transform. This connection is made explicit in the paper of Lu and Nolen (2015). (Some of the related literature on the stochastic control approach to sampling, such as the works of Pavon and Dai Pra on the Schr"odinger bridge problem, also point out the connection to Doob's transform.) This paper is, indeed, important, and we should not have overlooked it. While it does not phrase the problem of sampling from the process governing the reactive trajectories as a stochastic control problem, it provides the key structural characterization of the underlying process law. This, in turn, gives the structure of the optimal control law once a suitable variational formulation is available, so that the process law identified by Lu and Nolen is attainable by some choice of an admissible control law.

There are several new aspects in our stochastic control formulation relative to the above works. One is that we use a direct argument justifying the Girsanov change of measure that bypasses the Novikov condition. The latter requires an exponential moment bound, which is hard to verify for metastable systems. In a related manner, we use a different parametrization of the added control term: σ(x)u\sigma(x)u in the works of Hartmann, Sch"utte, et al. vs. a(x)ua(x)u in ours, where a(x)=σ(x)σ(x)⊤a(x) = \sigma(x)\sigma(x)^\top. This alternative parametrization (also used by Newton and Mitter) is in line with the form of the drift correction identified by Lu and Nolen, u=∇log⁡qu = \nabla \log q in their notation. It is conceptually important both for the Girsanov change of measure and for future extensions of our approach that do not require the uniform ellipticity assumption (Assumption 2 in our paper), which is also imposed by Lu and Nolen. Further, Corollary 1 (Appendix B.2) shows that the controlled SDE has a finite expected exit time for every admissible control, which the verification argument of Theorem 1 requires and which is often assumed rather than proved. Theorem 2 (Section 4.3) recovers the statistics of the original dynamics exactly from the simulated ones.

Finally, on the connection to ``Doob's Lagrangian'' (Du et al., NeurIPS 2024). One important difference from our work is that Du et al. consider the setting where the time horizon is fixed. Moreover, the Lagrangian formulation, phrased in Theorem 1 of Du et al. as joint optimization over the flow of marginal densities (qt∣0,T)(q_{t \mid 0,T}) and drift vector fields (vt∣0,T)(v_{t|0,T}) for t∈[0,T]{t \in [0,T]} subject to the Fokker--Planck PDE constraint, should be distinguished from the Hamiltonian optimal control formulation in our work, where we optimize only over the admissible control law. These formulations are related, and we have added a discussion to the introduction. The differences have practical consequences. Since our horizon is the first hitting time of the target, the duration of each path is random, and our paths end in a target set rather than at a prescribed configuration. Moreover, the optimal controller u∗=∇log⁡ρu^*=\nabla\log\rho is stationary, and ρ\rho solves the linear elliptic boundary value problem (11), which we solve in an RKHS, without training a neural network. Also see our response to Reviewer KE6c, Q2.

AuthorsOpenReview ↗

Response to Reviewer DyaX, continued

Response to Q2: The reviewer is correct. We have now summarized pertinent explanations and details in Section 5 and given them in full in App. G.3. The simulation follows the alanine dipeptide benchmark of Seong et al. (2025), which uses underdamped Langevin dynamics (VVVR integrator in OpenMM), so that our results are directly comparable with TPS-DPS. The controller is computed from an overdamped model in position space, which is a standard approximation of underdamped molecular dynamics at long timescales (Lelièvre et al., 2010) that has also been used in kernel-based Koopman models of peptide dynamics (Nüske & Klus, 2023). As stated in Section 5, this experiment lies outside Assumption 2 and demonstrates the utility of our method beyond the theory.

We use the scaling term for the following reason. In the underdamped dynamics, the controller enters as an external force in the momentum equation, as is standard for biased molecular simulation. Since ρ^\hat{\rho} is dimensionless, u^=∇log⁡ρ^\hat u=\nabla\log\hat\rho has units of nm$^{-1}$, and an energy is needed to turn it into a force. We therefore apply Fbias=su^F_{\rm bias}=s \hat u, where ss has units of energy. This has the same form as the physical force, which can be written as −∇U=kBT∇log⁡e−U/kBT-\nabla U=k_BT \nabla\log e^{-U/k_BT}, the gradient of a log-density multiplied by an energy. For the cap, the value ρ\rho is very small inside the source basin and rises quickly at the boundary, so ∇log⁡ρ^\nabla\log\hat\rho, and hence the bias force, is large there. A sudden large force on a single atom, in particular a hydrogen, can move it too far in one 1fs step and break the hydrogen-bond constraints. We therefore cap the magnitude of the bias force on each atom. The cap only acts where the bias is large and leaves it unchanged elsewhere.

Response to Q3: We have added a new Theorem 2 in Appendix F.2, which gives an exact reweighting for the approximate controller. In particular, the weight requires only ρ^\hat\rho and Lρ^L\hat\rho along the simulated path. New corollaries 2 and 3 in Appendix F.3 specialize this to Pμ(τ≤t)P^\mu(\tau\le t) and to the fraction of transitions through a channel or intermediate.

Response to Q4: We thank the reviewer for pointing this out. The description in Appendix G.1 was incomplete: The double-well experiment does not use a Koopman core as the target. It uses the same target set as Seong et al. (2025), namely the Euclidean ball of radius 0.50.5 around the right minimum (5/2,0)(\sqrt{5}/2, 0). All other settings also match theirs: The potential, temperature 12001200K, overdamped dynamics with m=γ=1m=\gamma=1, step size Δt=0.01\Delta t=0.01, horizon Tmax⁡=10T_{\max}=10, and paths started at the left minimum. To make the comparison exact, we have now evaluated both methods under the same criterion, counting a path as a hit if it enters the target ball before Tmax⁡T_{\max}, using 10241024 paths sampled from a TPS-DPS (F) model trained with the released code. Our controller reaches 99.8%99.8\% (10001000 paths) and TPS-DPS reaches 100.0%100.0\%. We have revised Table 1 to report these values, and we have corrected Appendix G.1, and all table captions now state the target set and the hit criterion precisely.

Reviewer_KE6cOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the cost of recent ML-based transition path samplers, which learn a bias force through repeated simulation over a fixed horizon. The authors propose to formulate transition path sampling as an exit-time stochastic optimal control problem, in which paths stop at the first hitting time of the target set and a committor-based running cost β(1−q)\beta(1-q) penalizes time spent in non-reactive regions. The claimed contributions are: (i) a closed-form stationary optimal controller u∗=∇log⁡ρu^*=\nabla\log\rho, with a verification theorem showing that its path measure minimizes the free energy over all path measures; (ii) a construction of this controller from linear kernel (RKHS) solves, without neural-network training or simulation-in-the-loop; (iii) identification of the metastable sets and the committor from Koopman eigenfunctions, without collective variables or transition paths; and (iv) reweighting of the sampled paths to recover statistics of the original dynamics. Validation on three 2D potentials and alanine dipeptide in vacuum reports high target-hit rates (99.8% on the double well; 93% on alanine vs. 100% for TPS-DPS) at a much lower construction cost (about 180 s vs. 20.7 h).

Critical strengths

A Koopman-based alternative to neural transition path samplers. The paper brings the classical framework of optimal stochastic control up to an exit time, together with Koopman spectral analysis, to machine learning-based TPS. The metastable sets, the committor function and the controller are all obtained from the leading Koopman eigenpairs and linear KKT systems, without network training or simulation-in-the-loop. Compared with PIPS and TPS-DPS, this avoids the need for a properly chosen time horizon and yields a stationary controller built from physically meaningful quantities. To my knowledge, this is the first machine learning-based TPS method that derives the bias from a Koopman spectral analysis rather than training a neural network.

Critical weaknesses
  1. The main practical advantage is difficult to verify. The method obtains the controller at a much lower cost than neural samplers. However, the reported cost does not appear to include the generation of the data samples for the Koopman eigenpairs and the collocation points, and the paper does not describe where these samples come from, including how points in the transition region are obtained without transition paths. The comparison also includes only simulation-based samplers (PIPS, TPS-DPS) and does not include simulation-free transition path samplers such as Doob's Lagrangian (Du et al., NeurIPS 2024).

  2. The contribution of each component is not isolated. The Koopman-based construction offers several components, namely the committor estimate, the committor-based running cost and the residual correction ww, but the ablations that would show which of them matter are not included, such as a Doob h-transform controller built from the same Koopman-based committor estimate and the controller from ρ0\rho_0 alone. As a result, the contribution currently relies mainly on the cost advantage discussed above.

Questions for authors
  1. How were the data samples for the Koopman eigenpairs and the collocation points generated, how many were used, and at what computational cost? In particular, do these samples include configurations in the transition region between the metastable sets, and if so, how were they obtained without transition paths?
  2. How does the method compare with Doob's Lagrangian (Du et al., NeurIPS 2024), a simulation-free transition path sampler with an official implementation, in both the cost of obtaining the bias (including data generation) and THP on the same benchmarks?
  3. How does the controller ∇log⁡ρ0\nabla\log\rho_0 (without the residual ww) perform on the same benchmarks compared with the full controller? Since ρ0\rho_0 depends only on q^\hat q and the scalar rr, this comparison would show whether the committor-based running cost, through the residual ww, contributes beyond the committor estimate itself.
4 comments
AuthorsOpenReview ↗

Response to Reviewer KE6c

We thank the reviewer for your time and insightful comments. Changes in the revised manuscript are highlighted in blue. In what follows, we address your comments one by one.

Our response to Critical Weaknesses 1: We thank the reviewer and we have included the cost of data generation in Table 3 in the manuscript. Our method only requires data from the original, uncontrolled dynamics, so the data are generated once and can be reused to construct new controllers, e.g., for other source and target sets or other values of β\beta. Moreover, existing unbiased simulations of the same system can be used directly. A new controller then requires only Stages I--III (about 3 minutes as shown in Table 3), whereas TPS-DPS must be retrained for each new task. For our experiments, since we did not have access to unbiased data for this setup, we generated the data ourselves, which we describe in detail in Q1 below, and we have added a description in Appendix G.3. This takes 65.3 minutes on a 16-core AMD EPYC-Milan CPU with an 18GB slice of H200 NVL GPU, and we have added it to Table 3. Even including data generation, obtaining the controller takes about 68 minutes in total, compared with 20.7,h of GPU training for TPS-DPS. We have also added a comparison with Doob's Lagrangian (see Q2).

Our response to Critical Weaknesses 2: We ran the requested ablations on alanine dipeptide, with identical data, kernels and evaluation (Tmax⁡=1T_{\max}=1,ps), and the results are added to Appendix G.3. The full method reaches THP 93% (ETS 40.6±14.940.6\pm14.9,kJ/mol), the closed-form factor alone (∇log⁡ρ0\nabla\log\rho_0, w≡1w\equiv1) gives 68.3% (21.8±11.621.8\pm11.6), and the Doob controller ∇log⁡q^\nabla\log\hat q built from the same Koopman committor gives 73.3% (21.4±10.621.4\pm10.6). The ablations show that the residual ww contributes beyond the committor estimate itself: Solving for ww raises THP from about 70%70\% to 93%93\%. In addition, we have added new results in Theorem 2 and Corollary 2,3, which show that the statistics of the original dynamics can be recovered exactly by a reweighting.

AuthorsOpenReview ↗

Response to Reviewer KE6c, continued

Our response to Q1: We now describe data generation in Appendix G.3. Our method does not need transition paths: it only needs sample points in the state space that cover both basins and the transition region, which, as explained in W1, can be taken from existing unbiased simulations. In our synthetic two-dimensional examples, every point is a valid state, so the samples can be placed directly. For alanine dipeptide, they must be valid molecular structures, and since we did not have access to unbiased data for this setup, we generated them as follows. We first ran 500500 short biased MD runs of 2525ps at 300300K, 250250 starting from C5 and 250250 from C7ax: 55ps without bias, followed by 2020ps with a harmonic bias on the heavy-atom RMSD to the other basin’s structure, whose reference value decreases over time so that the molecule gradually moves towards the other basin. From these runs we selected 300300 snapshots, evenly spaced in time, as starting points covering both basins and the transition region; the biased runs are used only to select these starting structures, not as paths. From these initial states, we ran 300300 unbiased MD trajectories of 5050ps, with frames stored every 11ps, and all samples used by our method are drawn from these unbiased trajectories. In total, this amounts to 27.527.5ns of MD and takes 65.365.3 minutes on a 16-core AMD EPYC-Milan CPU and an 1818GB slice of an H200 NVL GPU, which is now included in Table 3.

Our response to Q2: We ran the official implementation of Doob's Lagrangian (Du et al., NeurIPS 2024) on alanine dipeptide, and we have added the comparison to Appendix G.3. The two methods differ in formulation: Doob's Lagrangian fixes the time horizon and optimizes jointly over the flow of marginal densities and drift fields subject to the Fokker-Planck constraint (Theorem 1 of Du et al.), whereas we pose an optimal control problem up to the first hitting time of the target and optimize only over the control law, which we obtain in closed form (Theorem 1). In practice, Doob's Lagrangian samples bridges between two prescribed atomistic configurations over a fixed horizon, so every sample ends at the target by construction and THP is not a meaningful comparison. Our method samples transition paths from a source set to a target set with a random, path-dependent duration. Following the definition of Seong et al. (2025), both sets are discs of radius 0.750.75 in the backbone dihedrals (ϕ,ψ)(\phi,\psi), centered at the energy-minimized C5 and C7ax structures. We therefore compare the maximum potential energy along the path (ETS) on alanine dipeptide. In Cartesian coordinates, our method attains 40.6±14.940.6\pm14.9kJ/mol, whereas Doob's Lagrangian attains 779.9±0.06779.9\pm0.06kJ/mol with our end structures and 1079.8±0.21079.8\pm0.2 and 971.5±0.04971.5\pm0.04kJ/mol with its own. Its ETS drops to about −16-16kJ/mol, only with two additional inputs that our method does not require: A molecule-specific internal-coordinate parametrization, and consistent atom labeling between the two prepared end structures. Even then, its ETS varies by only 0.010.01kJ/mol across paths, far below kBT≈2.5k_BT\approx2.5kJ/mol. In fact, its sampled paths are essentially identical: Any two of them differ by only 5×10−55\times10^{-5}nm in heavy-atom RMSD, about 600600 times less than two random thermal configurations of the C5 basin, and all reach their maximum energy at the same frame and the same (ϕ,ψ)(\phi,\psi). Our paths differ by 0.050.05nm on average, comparable to unbiased dynamics at 300300K, and our ETS varies by 14.914.9kJ/mol across paths, consistent with the thermal fluctuation of the potential energy of alanine dipeptide at 300300K: Each internal degree of freedom contributes a fluctuation of order kBTk_BT, and nn independent contributions add to about kBTn/2≈12k_BT\sqrt{n/2}\approx12kJ/mol for n=48n=48 (6666 Cartesian coordinates minus 66 for overall translation and rotation and 1212 for the constrained X–H bonds). If the goal is to recover the statistics of the original dynamics, such concentration is a limitation: Any estimate, reweighted or not, can only assign weight to paths that are sampled, so channels and transition states that the samples never visit receive probability zero. We showed in Theorem 2 and Corollaries 2--3 that the statistics with our method are recovered exactly by reweighting, which Doob's Lagrangian does not provide. The result is also sensitive to endpoint preparation, as a single swapped hydrogen label in the end structures raises its ETS to 234.3±0.02234.3\pm0.02kJ/mol.

Our response to Q3: Please also see our response to W2: we have performed this additional experiment, and our result shows that ∇log⁡ρ0\nabla\log\rho_0 alone gives THP 68.3%, compared with 93% for the full controller.

Thank you for the careful and detailed response. Reporting the data generation cost, running Doob's Lagrangian, and the ablation on the residual ww, which shows a clear gain from about 70% to 93% THP, resolve my main concerns. I have raised my rating accordingly.

Thank you very much for your positive reassessment and for raising your score. We greatly appreciate your time, constructive feedback, and support for our paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their valuable comments. Changes in the revised manuscript are highlighted in blue, and the major/key updates are as follows.

(1) We added a new Section 4.3 to the main manuscript, with proofs in Appendix F.2. Theorem 2 therein shows that for the actually simulated approximate controller, reweighting the simulated paths exactly recovers the path statistic of the original dynamics. Corollaries~2--3 specialize this to transition probabilities and the probability that a transition passes through a given channel.

(2) We added a discussion of related work to the introduction, covering Mitter & Newton (2003), Hartmann & Sch"utte (2012), Lu & Nolen (2015) and Doob's Lagrangian (Du et al., 2024). In particular, unlike Doob's Lagrangian, which fixes the time horizon and optimizes jointly over the flow of marginal densities and drift fields subject to the Fokker--Planck equation, and samples bridges between two prescribed atomistic configurations, we pose an optimal control problem up to the first hitting time of the target and optimize only over the control law. We obtain the latter in closed form, and sample transition paths between source and target sets identified from Koopman eigenfunctions, with a random, path-dependent duration.

(3) We added a comparison with Doob's Lagrangian on alanine dipeptide, the cost of data generation, and ablation results in Appendix G.3, referenced in Section 5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.