- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes a method called KC maps, which learns a finite-time transition density for an SDE from equilibrium samples and forces alone, without needing time-correlated samples like trajectories. It does so by training conditional normalising flows that map from the current state to the time-warped distribution. The loss is based on two-hop self-consistency, and the architecture smartly incorporates agreement between the known force and diffusion coefficients. Excluding training, it records higher throughput than MD at the cost of moderate error in the molecular systems.
Critical strengths
I think the theory is very elegant and easy to follow, and the paper is quite well written.
The method closes a literature gap. Furthermore, they modify the normalising flow coupling-layer architecture to satisfy most constraints before training, leaving only the two-hop consistency to the loss. For the toy potential, they also find they can use higher-temperature configurations to train the model, which is very promising for the future of a method like this. This would drastically speed up data acquisition, since data collection and training on these single systems make the method a bit moot right now. That said, this is of course understandable, given that this is early-stage method development. I wouldn't have expected the offer to train this on multiple systems, though that would have been a great plus.
Critical weaknesses
In my view, the biggest weakness is the lack of kinetics validation. Given that the data come from equilibrium simulations and the model outputs the same quantities, it would be interesting to include kinetics, especially since one can derive the rates from the reference data. One might imagine training data coming from umbrella sampling, which forsakes kinetics, thus already giving the method a use case, since it could supply those in principle.
Then the presented setup is a bit of a chicken-and-egg problem. Given that one now only trains on one system, it would have been nice to see the authors train on multiple systems, where a DL sampling method would make more sense, since the method would then amortise.
Questions for authors
Could you compute rates using this method? It should be possible, because you can get first passage times and the like just by sampling with this method.
Similarly, for the AL6 experiments, you reach a mode much faster than MD, and MD does not discover the state in 7 times as much simulation time. Is this a statistical accident, or is it related to any kinetic problems you might have? If so, where do they come from? The paper should expand on this. For instance, because it's mass-covering, the KL objective could make the kernel too wide, and clipping could smooth the landscape.
Can you set up a situation where you remove an entire mode from the training data? If so, can the method discover that mode? expectation is no, and that the method like this only shines once one amortises over different systems, but I'm still very curious to find out
There is an energy mismatch in the molecular systems, and I think a Metropolis step could alleviate it. In that case, how would the results change, and what would the acceptance rate be?
6 comments
Response #2 to Reviewer 7zHe
We continue addressing the reviewer concerns in the following:
Can you set up a situation where you remove an entire mode from the training data?
As the reviewer requested, we ran a longer simulation on AL6 for KC-Maps beyond the single GPU-hour and observed a mode recovered that was not present in the training data (sourced from [6]). To verify, we ran a larger MD run where we ran 16 chains of 2 microseconds per chain and indeed this mode is not an artifact of our method. To be clear, we kept the training data the same (sourced from Tan et al. [6]) which does not include that additional mode and illustrate this in Figure A1. As a result, we see our method is able to discover modes not present in the training data. We also show updated equilibrium metrics for AL6 using the new reference MD in Table A5.
There is an energy mismatch in the molecular systems, and I think a Metropolis step could alleviate it. In that case, how would the results change, and what would the acceptance rate be?
We agree that there is an energy mismatch for AL6 though we argue that energy distribution for AL2 and AL6 is well-matched; see figure 2). Though, we agree that a Metropolis step could help alleviate it. We provide the acceptance rates for all three systems (see Figure A3). We observe that, for AL6, it has low acceptance rates (< 1 %) for \tau > 20 fs so it makes MH-adjusted run not practical. However, we conducted an additional ablation where we incorporate the detailed-balance as a loss objective in the last 15% steps of training, and we observe that this addresses the energy issue seen in AL6 as well as improving the results in general. We report the results in Table 2 and add a dedicated section in Section 3.3.
Then the presented setup is a bit of a chicken-and-egg problem.
We agree that the trajectory-free setting indeed poses a "chicken and egg" problem by not specifying the source of the samples from the equilibrium distribution. Our main contribution is the method (KC-Maps) that can learn the kinetics only from independent samples at the equilibrium distribution. Therefore, we believe that our method would yield a practical solution when paired with any of the following approaches:
- Getting uncorrelated equilibrium samples from other sampling algorithms like Sequential Monte Carlo [1], Parallel Tempering MCMC [2], Annealed Importance Sampling [3] and then using the proposed method for the kinetics.
- Learning our method at a higher temperature and reducing the temperature for the inference (see Section 4.2 and Table 2 for Lennard-Jones experiment where training samples come from higher temperature data, but we learn the dynamics for the target temperature/distribution).
- Learning our method on partially available data, e.g. partial mode coverage. instead of the equilibrium distribution (see experiment in Appendix A.1 Figure A1)
- Amortizing our method across systems, i.e. it has been demonstrated previously that models can generalize to unseen systems [4, 5]. As suggested by the reviewer, we additionally demonstrate a small study on amortizing across dipeptides (see Appendix E.4).
Our paper does not provide a definitive answer, but we envision multiple possible solutions as described in what follows. Each of the proposed solutions deserves an independent study in future work.
[1] Del Moral, P., Doucet, A., & Jasra, A. (2006). Sequential Monte Carlo samplers. Journal of the Royal Statistical Society: Series B, 68(3), 411–436.
[2] Hukushima, K., & Nemoto, K. (1996). Exchange Monte Carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6), 1604–1608.
[3] Neal, R. M. (2001). Annealed importance sampling. Statistics and Computing, 11(2), 125–139.
[4] Tan, C. B., Hassan, M., Klein, L., Syed, S., Beaini, D., Bronstein, M. M., Tong, A., & Neklyudov, K. (2025). Amortized sampling with transferable normalizing flows. arXiv:2508.18175.
[5] Klein, L., Foong, A. Y. K., Fjelde, T. E., Mlodozeniec, B., Brockschmidt, M., Nowozin, S., Noé, F., & Tomioka, R. (2023). Timewarp: Transferable acceleration of molecular dynamics by learning time-coarsened dynamics. NeurIPS 2023.
[6] Tan, C. B., Bose, A. J., Lin, C., Klein, L., Bronstein, M. M., & Tong, A. (2025). Scalable equilibrium sampling with sequential Boltzmann generators. ICML 2025.
Thank you for the extensive revision and the many new experiments. The multi-system training on dipeptides, the AL2 kinetics, the unseen AL6 mode and the detailed-balance loss go well beyond what I asked for and address most of my concerns. Two points remain.
-
Mode discovery on AL6. My concern is about simulated time rather than compute. By the time KC-maps finds the third mode, it has simulated much less time than MD did over the full hour without finding it (roughly 25 ns against 183 ns). If both followed the same dynamics, this would be unexpected, so I suspect the original model crosses barriers too quickly. The AL2 kinetics use the new L_DB model, while Fig. 3 appears to still show the original one, so they don’t resolve this. Could you rerun the mode-discovery experiment with the L_DB model, ideally with a few replicas? If it no longer finds the mode much faster than MD, the corresponding claims in the abstract and introduction should be adjusted.
-
Acceptance with and without L_DB. Fig. A3 reports acceptance only for the model trained with Eq. (20). Since L_DB is the expected negative log Metropolis–Hastings ratio, the same curve without it would show directly what the new term contributes. Please also clarify which model the “<1 % for τ > 20 fs” in your response refers to.
Great work in general; I am considering raising my score.
Response #1 to Reviewer 7zHe
We thank the reviewer for their valuable feedback and greatly appreciate their comment on the theory being elegant and paper being easy to follow! We aim to resolve the reviewer’s concerns in the following:
As the reviewer requested, we provide results evaluating the kinetics for alanine dipeptide for KC-Maps against the reference MD where we measure the mean first passage times (see figure 4.b), autocorrelation of mode occupancy (see figure 4.c), and the transition probabilities between modes (see Appendix E.2 and Figure A2). We added a paragraph discussing the results (see Kinetics in Section 4.3) and a description of the metrics in Appendix D.4. From the results, we see that our method, with a larger step size of τ=30fs, is closely aligned with MD, especially in the larger modes where training data is abundant. We thank the reviewer for bringing up this point and for their suggestion as we believe this substantially strengthens our work.
Per the reviewer's request, we ran a preliminary study on amortization. We trained KC-Maps on 200 dipeptides from the ManyPeptidesMD dataset and evaluated it on 30 unseen dipeptides. Table A6 (Appendix E.4) reports the energy, torsion, and TICA W2, averaged over the 30 test dipeptides. The model generalizes well to unseen peptides. At τ=30fs, its average energy error is 0.48kBT. Its TICA error (0.112) is of the same order as MD’s (0.058), showing that the model recovers the slow conformational modes of molecules it has never seen.
We also compare against two transferable baselines, Timewarp [5] and Prose [4], with numbers taken from Table 2 of [4] (the paper does not report TICA for dipeptide). We see KC-Maps outperforms Timewarp on both energy (0.48 vs. 4.53) and torsion (0.55 vs. 0.84) W2. On energy it is competitive with Prose (0.48 vs. 0.37), even though Prose relies on self-normalized importance sampling (SNIS) and KC-Maps uses no reweighting. Finally, 16 MD chains converge quite well to the equilibrium within one GPU-hour, as we also observe for AL2 in the single-peptide experiments (Table 3). The advantage of KC-Maps appears on larger systems where MD mixes slowly. Namely, on AL4 and AL6, KC-Maps reaches lower TICA error than MD at the same compute budget (Table 3) and explores the reference modes faster (Figure 3). Given the time constraints, we only ran this study for dipeptides, and we leave a more focused study on amortization and scaling to future work.
We appreciate the question and hope to add clarification – the methods differ in simulated time and KC-Maps simulate for longer given the same compute budget (single GPU-hour on L40S). As we can see in figure 3, for 16 chains, we have ~505 total ns / GPU-Hour for KC-Maps (w/ step-size \tau=40fs) vs ~183 total ns / GPU-hour for MD, which results in a per-chain simulation time of ~31 ns and ~11 ns respectively. The increased throughput is primarily a result of two reasons: (i) the larger step size of KC-Maps, and (ii) the better utilization of the GPU of neural networks versus the force field. As for MD, the low per-chain simulation time can be explained by force fields poorly utilizing the GPU, especially with increasing number of chains as seen in figure 3.b where MD saturates at 4 chains for AL6.
In contrast, KC-Maps has a higher throughput ceiling as the number of chains continues to increase. This results in our method reaching a longer time horizon per chain whereas MD may not have just had enough time to reach that mode. Indeed, we see in the reference MD with longer trajectory in figure 3.a, MD eventually does recover that mode. Moreover, as we can see from the kinetic results in an earlier answer, we see the issue is not because of kinetics.