acceptodds
Under review as a conference paper at ICLR 2027

Structure-Constrained Koopman Generators for Flexible Time-Stride Molecular Dynamics Simulations

Abstract

Trajectory-based models can accelerate molecular dynamics (MD) simulations by predicting configurations over long temporal strides, but most are tied to a single prescribed temporal resolution during training, limiting their ability to leverage heterogeneous trajectory datasets and to infer dynamics at unseen time strides. To this end, we introduce FlexMD, a continuous-time Koopman framework that enables flexible time-stride prediction via structure-constrained equivariant latent dynamics. It encodes molecular configurations into an E(3)-invariant latent space and evolves them with a constrained linear generator. The generator preserves a learned quadratic invariant that regularizes latent evolution and prevents exponential amplification of error with positive definite matrix. The closed-form propagation of latent dynamics via matrix exponential enables a single model to learn across multiple temporal strides and perform inference at flexible temporal resolutions without iterative latent-space integration. We evaluate our method across various benchmarking datasets, including MD17, AdK, ATLAS, MD-CATH, and PET-MAD. Crucially, our model achieves accurate and stable long-horizon simulation, supports flexible inference across substantial training-test stride gaps, and benefits from multi-stride training that robustly improves generalization across temporal resolutions. These results demonstrate that the structure-constrained Koopman generator offers a principled and scalable path toward reliable long-timescale molecular simulation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular dynamics, Koopman operator learning, Geometric deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_b6EQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This method uses Koopman theory to generate molecular trajectories, learning across different temporal strides and providing a flexible time grain for sampling trajectories. The results on different benchmarking datasets also suggest the effectiveness of the method.

Critical strengths
  1. The paper is well-written and well-organized.
  2. Applying the Koopman theory with a matrix constraint sounds reasonable for predicting further frames efficiently.
Critical weaknesses
  1. This experiment lacks metrics related to ensemble and kinetics. Table 1 seems to miss several important Alphaflow metrics; I was wondering why the authors report only some of them instead of all of them. For kinetics, the paper only reports TICA in ATLAS. As a trajectory generation method, it would be better to report the performance on kinetics.

  2. Some important baselines are missing, like PVB (ICLR 2026), EBA(ICML25) BIoMD(ICLR26), and DynaMode (ICML 2026 Workshop), especially the relevant method [1], which also uses the Koopman theory. More detailed discussion and comparison will make the paper convincing.

  3. The conceptual novelty of the proposed method should be claimed more clearly.

[1]Sengar, Aditya, et al. "Beyond ensembles: simulating all-atom protein dynamics in a learned latent space." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

1.It seems that the Linear dynamics in the latent space is a Deterministic process; I was wondering if the model can get different replicas with the same initial conformation. 2. Did the A = JS constraint constrain the ability of the exploration to reach different meta states?

6 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer b6EQ

We thank the reviewer for the thoughtful and constructive assessment. We appreciate the recognition of the paper’s clear writing and organization, as well as the reviewer’s positive assessment of the constrained Koopman approach for efficient prediction of future frames. We have incorporated the corresponding revisions into the manuscript and highlighted them in blue.

Response to Weakness 1 [Ensemble and kinetic metrics]

This experiment lacks metrics related to ensemble and kinetics. Table 1 seems to miss several important Alphaflow metrics; I was wondering why the authors report only some of them instead of all of them. For kinetics, the paper only reports TICA in ATLAS. As a trajectory generation method, it would be better to report the performance on kinetics.

Response:

Thanks for this suggestion. We have added ensemble metrics from AlphaFlow and autocorrelation (AutoCor) comparisons with reference MD to the revised manuscript. AutoCor assesses temporal correlations across lags and complements the existing tICA metric. As shown below, FlexMD achieves the best joint PCA distance and weak-contact similarity, competitive transient-contact similarity, and the lowest AutoCor deviations at longer lags; these results have been added to the revised manuscript.

MetricMDGenFlashMDAlphaFlowConfRoverFlexMD
Joint PCA W2W_2 distance ↓3.084.192.242.332.14
Weak contacts Jaccard similarity ↑0.520.480.610.580.63
Transient contacts Jaccard similarity ↑0.280.250.410.350.39
AutoCor deviation at ℓ=40\ell=40 ↓0.060.280.120.080.06
AutoCor deviation at ℓ=80\ell=80 ↓0.260.230.370.240.10
AutoCor deviation at ℓ=120\ell=120 ↓0.470.520.410.320.12

Response to Weakness 2 [Missing baselines]

Some important baselines are missing, like PVB (ICLR 2026), EBA(ICML25) BIoMD(ICLR26), and DynaMode (ICML 2026 Workshop), especially the relevant method [1], which also uses the Koopman theory. More detailed discussion and comparison will make the paper convincing.

Response:

Thank you for suggesting these relevant baselines. We have added comparisons with PVB, EBA, BioMD, DynaMode, and Sengar et al. PVB uses a pretrained variational bridge; EBA aligns ensemble generation with physical feedback; BioMD combines forecasting and interpolation; and DynaMode generates trajectories through spectral diffusion. Sengar et al. is particularly relevant because it also draws on Koopman theory. To our knowledge, these methods have not explicitly demonstrated a single model trained across a range of physical lag times and conditioned on a selectable lag at inference. FlexMD addresses this setting using a continuous-time, quadratic-form-preserving Koopman generator.

We compare the methods on AdK and ATLAS.

For ATLAS, following the setting in Table 1, models are trained from scratch and evaluated at stride 40:

MethodPairwise RMSD r ↑Global RMSF r ↑Per-target RMSF r ↑Root mean W₂ ↓MD PCA W₂ ↓JSD ↓TICA ↑
PVB0.63 ± 0.020.65 ± 0.020.68 ± 0.032.95 ± 0.031.75 ± 0.050.62 ± 0.000.09 ± 0.02
EBA0.63 ± 0.020.71 ± 0.010.82 ± 0.032.53 ± 0.031.59 ± 0.040.46 ± 0.020.10 ± 0.00
BioMD0.70 ± 0.010.75 ± 0.010.85 ± 0.022.62 ± 0.051.56 ± 0.030.49 ± 0.030.12 ± 0.01
DynaMode0.56 ± 0.000.71 ± 0.040.74 ± 0.052.76 ± 0.041.64 ± 0.010.78 ± 0.030.14 ± 0.02
Sengar et al.0.34 ± 0.050.66 ± 0.010.69 ± 0.022.82 ± 0.011.69 ± 0.020.98 ± 0.010.10 ± 0.02
FlexMD0.72 ± 0.010.71 ± 0.030.86 ± 0.052.51 ± 0.011.42 ± 0.020.48 ± 0.030.17 ± 0.004

For AdK, following the setting in Table 4, models are trained from scratch at stride 5 and evaluated at the indicated test strides:

Method51025100
PVB0.922.462.618.52
EBA1.243.952.57—
BioMD0.932.322.55—
DynaMode2.533.98——
Sengar et al.1.031.892.489.54
FlexMD0.500.610.780.91

FlexMD achieves the lowest AdK prediction error at each reported test stride and the best mean performance on five of the seven ATLAS metrics. BioMD achieves higher global RMSF correlations, and EBA achieves a lower JSD; however, FlexMD still demonstrates excellent overall performance compared to these baselines. The comparison with Sengar et al. further shows that FlexMD performs well against another Koopman-based approach under these evaluation settings. We have added these experimental results to the revised manuscript.

AuthorsOpenReview ↗

[2/2] Response to Reviewer b6EQ

Response to Weakness 3 [Conceptual novelty]

The conceptual novelty of the proposed method should be claimed more clearly.

Response:

FlexMD proposed an EGNN-style molecular representation with a continuous-time A=JSA=JS latent generator, where S≻0S\succ0 preserves a learned quadratic form and gives an SS-norm bound on accumulation of latent discrepancies. The closed-form propagator can be evaluated at different lags, and the paper studies this framework across several molecular benchmarks. Our core contribution is proposing an organic framework with flexible-lag modeling, Koopman embeddings, and quadratic Hamiltonian structure and empirical evaluation on a variety of datasets.

Response to Question 1 [Different replicas]

It seems that the Linear dynamics in the latent space is a Deterministic process; I was wondering if the model can get different replicas with the same initial conformation.

Response:

Similar to most existing prior works, the same complete initial input produces the same trajectory, so the model cannot generate different replicas from an identical input. Different replicas could instead be obtained by varying the initial conformation or supplied momenta, or by introducing a stochastic component; we leave the exploration of these approaches to future work.

Response to Question 2 [Exploration of metastable states]

Did the A = JS constraint constrain the ability of the exploration to reach different meta states?

Response: Thank you for this question. On MD-CATH, where training includes MSM-guided pairs spanning metastable states, FlexMD achieves an MSM JSD of 0.17 ± 0.02, compared with 0.19 ± 0.00 for MARS-FM, suggesting that the constrained model can capture metastable-state distributions when relevant transitions are represented in the data. We conjecture that training-data coverage is an important factor in learning to reach different metastable states. At the same time, we think this constraint (A = JS) does not necessarily prevent transitions between metastable states: the encoder and decoder can place multiple physical states along the same latent trajectory.

Thanks for the response. I was wondering how you got the baseline results (mean ± std), especially for EBA, BioMD, and Sengar et. al. Are you training from scratch or using inference based on their checkpoint? Could you also report your results on the median, which aligns well with the baseline common setting?

Reviewer_KEVxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FlexMD, a model for predicting molecular configurations at different time strides. A graph encoder maps configurations into an invariant latent representation, which evolves under a continuous-time linear generator that preserves a learned positive-definite quadratic form. Matrix-exponential propagation allows predictions at arbitrary times without numerical integration in latent space, and the authors bound the accumulation of latent prediction errors. Experiments on MD17, AdK, ATLAS, MD-CATH, and PET-MAD evaluate prediction accuracy, structural stability, ensemble statistics, and transfer across time strides.

Critical strengths
  1. The protein results are promising. On ATLAS, FlexMD improves pairwise RMSD correlation from 0.53 to 0.72 and reduces coordinate Wasserstein distance from 2.63 to 2.51 relative to the best baseline for each metric (Table 1).

  2. The constrained-versus-unconstrained comparison in Table 8 directly tests the benefit of the generator constraint. At test strides 1,000 and 3,000, the unconstrained models become numerically unstable while the constrained models retain finite errors.

Critical weaknesses
  1. The methodological contribution is modest. Goyal et al. [1] already learn stable latent Hamiltonian dynamics with a positive-definite quadratic energy. FlexMD combines this structure with molecular geometric modeling and training on lagged configuration pairs. This is a useful application, but the experiments omit the closest methods for the central flexible-stride task: ITO [2] and TITO [3] already learn molecular transitions conditioned on the physical lag, with TITO also transferring across systems. I would like to see a comparison with one of these methods e.g. TITO since it is transferable.

  2. I think details are missing regarding how the decoder produces equivariant coordinates. The encoder pools scalar features into a latent vector invariant to rigid transformations (lines: 850-856). Appendix C.1 then says the decoder uses the initial edges and edge features, but does not describe how the decoder uses the inputs to construct the output (lines: 857-859).

  3. The evidence for accurate long-timescale kinetics is limited. The coordinate MSE results test prediction at specified times, but Table 2 evaluates ensemble properties; for example, its folding free-energy error does not measure folding rates. I could not find reported kinetic measurements, such as folding/unfolding rates or mean first-passage times. Could the authors compare autocorrelation curves with reference MD over lags ℓ=10\ell=10 to 30003000, following the evaluation in STAR-MD [5] (Section 10, Eq. 31)? This would test whether the generated trajectories reproduce the temporal correlations of MD across these lags.

  4. Figure 2a is unclear. FlexMD does not appear in the legend, and I am not sure whether “BondFF” is supposed to refer to it as it is not explained anywhere in the paper. I also think the comparison could be visualized more clearly by fitting TICA (where the model is fitted on C$\alpha$ coordinates of the reference MD trajectory) and plotting both the reference and generated trajectories in that same projection. Figure 1b of Timewarp [4] is an example of what I mean.

  5. Table 3 is concerning to me. First, the paper does not establish what Δt\Delta t is here to understand the timescale for each ℓ\ell. Second, I am not convinced that coordinate MSE establishes accurate time-stride generalization. The very low MSE at ℓ=5\ell=5 is not surprising given the short prediction interval e.g 2.5 fs if Δt=0.5\Delta t=0.5 fs for ℓ=5\ell=5. At this timescale, the possible future configurations remain close, so a deterministic prediction can achieve very low error. Yet the model trained on these very short transitions still achieves an MSE of 3.16×10−23.16\times10^{-2} at ℓtest=3000\ell_{\text{test}}=3000, compared to 2.23×10−22.23\times10^{-2} for the model trained at ℓ=3000\ell=3000. A model supervised on nearly unchanged configurations therefore scores comparably to one supervised on much longer transitions. This makes me question whether the metric is sensitive enough to the longer-time motions that flexible-stride prediction is supposed to capture. Since the current positions do not uniquely determine the future configuration, MSE against a single MD sample can favor an average prediction without showing that the model reproduces the dynamics.

[1] Pawan Goyal, Süleyman Yıldız, and Peter Benner. Deep Learning for Structure-Preserving Universal Stable Koopman-Inspired Embeddings for Nonlinear Canonical Hamiltonian Dynamics. 2023 preprint; Machine Learning: Science and Technology, 2025.

[2] Mathias Schreiner, Ole Winther, and Simon Olsson. Implicit Transfer Operator Learning: Multiple Time-Resolution Surrogates for Molecular Dynamics. NeurIPS, 2023.

[3] Juan Viguera Diez, Mathias Schreiner, and Simon Olsson. Transferable Generative Models Bridge Femtosecond to Nanosecond Time-Step Molecular Dynamics. October 2025 preprint.

[4] Leon Klein, Andrew Y. K. Foong, Tor Erlend Fjelde, Bruno Mlodozeniec, Marc Brockschmidt, Sebastian Nowozin, Frank Noé, and Ryota Tomioka. Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics. NeurIPS, 2023.

[5] Nima Shoghi, Yuxuan Liu, Yuning Shen, Rob Brekelmans, Pan Li, and Quanquan Gu. Scalable Spatio-Temporal SE(3) Diffusion for Long-Horizon Protein Dynamics. 2026 preprint.

Questions for authors
  1. If the decoder is given the coordinates it reconstructs, what prevents it from copying them while the encoder outputs zero? Equation (4) appears to allow both displayed losses to vanish in that case.

  2. In Table 4, were the stride-5 baselines applied 2, 5, and 20 times to predict at strides 10, 25, and 100 or are you doing one-shot inference to higher strides? Please also specify the training stride and future information used by each baseline in Table 3. These details are needed to interpret the reported failures at long strides and assess whether the methods solve the same prediction task.

  3. What varies across the ten independent inference runs on ATLAS and MD-CATH (lines: 1068-1069 and 1091-1093)? Please specify how starting configurations are selected and whether any randomness is introduced after initialization.

  4. What exactly does “TICA ↑\uparrow” in Table 1 measure? I am familiar with TICA-W2, where lower is better, but I could not find a definition of the reported score.

6 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer KEVx

We thank the reviewer for the thoughtful and constructive assessment. We appreciate the recognition of the promising protein results on ATLAS and the constrained-versus-unconstrained comparison, which demonstrates the benefit of the generator constraint for stability at large strides. We have incorporated the corresponding revisions into the manuscript and highlighted them in blue.

Response to Weakness 1 [Novelty and missing baselines]

The methodological contribution is modest ...

Response: Our contribution lies in proposing an equivariant molecular representation and continuous-time latent propagation. While Goyal et al. learn stable, structure-preserving embeddings for canonical Hamiltonian systems, we designed a quadratic-invariant latent generator with molecular geometric encoding and prediction from configuration pairs, without requiring the latent blocks to represent physical positions and momenta. While ITO learns stochastic, lag-conditioned molecular transition distributions using diffusion models and TITO extends generative transfer-operator learning across molecular systems, FlexMD uses a quadratic-invariant linear latent propagator whose matrix exponential provides predictions at flexible lags while reduced the error accumulation.

To address the requested comparison, we evaluated these methods on AdK. Models are trained at stride 5 and evaluated at strides 5, 10, 25, and 100; the table reports next-time-stride MSE, with lower values indicating better point prediction.

Test strideGoyal et al.ITOTITOATOMFlexMD
58.24±0.833.85±0.782.67±0.620.68±0.480.50 ± 0.51
1012.35±0.758.38±0.657.49±0.810.88±0.710.61 ± 0.10
2582.76±0.6910.28±0.829.85±0.751.92±0.680.78 ± 0.82
100102.73±0.8329.93±0.5210.23±0.692.84±0.870.91 ± 0.69

FlexMD achieves the lowest next-time-stride MSE across the evaluated lags, including those beyond the training stride. This result would support its effectiveness for flexible-lag prediction on AdK.

Response to Weakness 2 [Equivariant decoder]

I think details are missing regarding...

Response: We have revised Appendix C.1 to clarify that the encoder retains per-atom coordinate and scalar features. The decoder initializes each atom’s coordinate state from the first three propagated latent channels and uses the remaining channels as scalar node features. EGNN-style layers then update coordinates using scalar-weighted relative vectors, with the initial graph and invariant edge features.

Response to Weakness 3 [Kinetics and autocorrelation]

The evidence for accurate long-timescale kinetics is limited ...

Response: We compare autocorrelation errors against reference MD at lags ℓ=40,80,120\ell=40,80,120, following the STAR-MD evaluation. The requested range of ℓ=10\ell=10 to 30003000 is not available under our ATLAS evaluation because of the limited time-strides evaluated. STAR-MD obtains the lowest error at ℓ=40\ell=40, while FlexMD achieves the lowest error at ℓ=80\ell=80 and 120120. These results provide a limited assessment of temporal correlations.

Methodℓ=40\ell=40ℓ=80\ell=80ℓ=120\ell=120
MDGen0.060.260.47
FlashMD0.280.230.52
AlphaFlow0.120.370.41
ConfRover0.080.240.32
STAR-MD0.040.190.28
FlexMD0.060.100.12

Response to Weakness 4 [Figure 2a and TICA projection]

Figure 2a is unclear ...

Response:

Thank you for noting the unclear legend and suggesting the TICA visualization. We have corrected the legend in Figure 2a and added a TICA projection of the reference and generated trajectories to the revised manuscript.

Response to Weakness 5 [Physical timescale and MSE sensitivity]

Table 3 is concerning to me ...

Response: We agree that the low MSE at ℓtest=5\ell_{\mathrm{test}}=5 is partly attributable to the small displacement over a short interval and should not by itself be taken as evidence of accurate dynamics. The purpose of Table 3 is to assess whether a model trained at one stride can be evaluated at other strides. At ℓtest=3000\ell_{\mathrm{test}}=3000, the model trained at ℓ=5\ell=5 has an MSE of 3.16×10−23.16\times10^{-2}, compared with 2.23×10−22.23\times10^{-2} for the model trained at ℓ=3000\ell=3000 and outperformed other baselines. This shows that short-stride training permits longer-stride predictions.

We also agree that MSE against a single reference trajectory can favor an average prediction and does not measure the predicted distribution or transition kinetics. Our ATLAS evaluation therefore also reports RMSD and RMSF correlations and Wasserstein distances, while MD-CATH includes distributional divergences and folding free-energy errors. These provide complementary evidence beyond coordinate MSE.

AuthorsOpenReview ↗

[2/2] Response to Reviewer KEVx

Response to Question 1 [Decoder copying and collapse]

If the decoder is given the coordinates it reconstructs, what prevents it from copying them while the encoder outputs zero? Equation (4) appears to allow both displayed losses to vanish in that case.

Response:

In the implementation, the decoder receives coordinate-like features produced by the encoder, so the reconstruction loss alone does not rule out a copying shortcut. Moreover, because these features are part of the latent state, a zero scalar representation alone would not make the full latent dynamics loss vanish. Nevertheless, Equation (4) does not by itself guarantee a nontrivial latent representation or exclude degenerate solutions. We will clarify this and the decoder’s inputs in the revised manuscript.

Response to Question 2 [Baseline protocols]

In Table 4, were the stride-5 baselines applied 2, 5, and 20 times to predict at strides 10, 25, and 100 or are you doing one-shot inference to higher strides? Please also specify the training stride and future information used by each baseline in Table 3. These details are needed to interpret the reported failures at long strides and assess whether the methods solve the same prediction task.

Response: In Table 4, all baselines are trained only at stride 5 and use one-shot inference at strides 10, 25, and 100.

In Table 3, each FlexMD column corresponds to a model trained only at the indicated stride and evaluated at different strides. FlexMD does not use data from other strides during training or inference. For the baselines, we evaluated each time stride using models trained on all time strides and reported the best performance.

Response to Question 3 [Independent inference runs]

What varies across the ten independent inference runs on ATLAS and MD-CATH (lines: 1068-1069 and 1091-1093)? Please specify how starting configurations are selected and whether any randomness is introduced after initialization.

Response:

We trained FlexMD once for each dataset and performed ten evaluation runs using the same trained model. For MD-CATH, we evaluated each test domain at 450 K. For each run, we selected one of its five reference MD replicas randomly and used the first frame of that replica as the input configuration. For ATLAS, we selected one of the three reference MD replicas randomly and used the first frame as the input configuration. No additional randomness was introduced after initialization; inference was deterministic under the evaluation settings. We have clarified this protocol in the revised paper.

Response to Question 4 [TICA metric]

What exactly does “TICA ↑\uparrow” in Table 1 measure? I am familiar with TICA-W2, where lower is better, but I could not find a definition of the reported score.

Response: “TICA ↑\uparrow” follows the tICA correlation metric in STAR-MD. Separate tICA models are fitted to the reference and generated trajectories, and the score is the absolute Pearson correlation between their per-residue contributions to the top two tICA components, averaged across components and lag times. Higher values indicate more similar patterns of residue contributions to the slow modes.

Reviewer_KEVxOpenReview ↗

Response to Authors

I thank the authors for their response. I still have concerns regarding the paper.

First and foremost, I still think the choice of problem is a poor setting for your method since you train on stochastic molecular dynamics data using a deterministic model. For example, MD-CATH is generated using Langevin dynamics, but your method produces only one future for a given starting configuration and time lag. It therefore cannot represent the conditional distribution of possible future configurations for this stochastic process. I think this gap is unaddressable with the method as currently formulated. Below, I describe in more detail my other concerns:

To address the requested comparison, we evaluated these methods on AdK. Models are trained at stride 5 and evaluated at strides 5, 10, 25, and 100; the table reports next-time-stride MSE, with lower values indicating better point prediction.

I believe there are a few flaws in the evaluation. As I stated before, the MSE metric is not a good indicator of measuring the quality of the dynamics from the model. For these dynamics models, you need to compare the equilibrium distribution of long trajectories and the kinetics observables. The reported MSE error does not establish your method reproduces either.

Moreover, you compare to ITO and TITO where they generate conditional distributions of the future state whereas your method generates a deterministic prediction. MSE against a single reference future can favor a prediction close to the conditional mean while punishing models that output distributions. The baselines will produce variance, and even when that variance correctly reflects the possible MD futures, it can increase the reported MSE. Therefore, the lower MSE of your method does not establish better dynamics than ITO or TITO.

Response to Weakness 3 [Kinetics and autocorrelation]

Can you describe the autocorrelation setup in Figure 3? I only see errors at lags 40, 80, and 120, but it is unclear what these correspond to in physical time. STAR-MD shows autocorrelation curves computed for 100 ns trajectories. Your appendix also describes 100 ns ATLAS rollouts. Does Figure 3 use these rollouts? Please show the autocorrelation curves alongside reference MD.

Response to Weakness 4 [Figure 2a and TICA projection]

Looking at figure 2, it seems like FlexMD outputs just two gaussian blobs and does not resemble the reference distribution at all. Also, typically, tica plots are shown as density plots (see figure 1.b minus trajectory lines in [1]). I am not convinced the method samples from the equilibrium distribution reliably.

Response to Weakness 5 [Physical timescale and MSE sensitivity]

While your response acknowledges the weakness of the evaluation, it does not resolve it, and my weakness 5 still stands. The additional ensemble metrics do not establish that the flexible-stride predictions reproduce the conditional distributions or kinetics of MD.

[1] Leon Klein, Andrew Y. K. Foong, Tor Erlend Fjelde, Bruno Mlodozeniec, Marc Brockschmidt, Sebastian Nowozin, Frank Noé, and Ryota Tomioka. Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics. NeurIPS, 2023.

Reviewer_E4w7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce a new framework for MD learning in which molecular dynamics are linearised in a latent space, evolved in closed form, and decoded back into molecular configurations. By constraining the latent generator to preserve a quadratic invariant, they aim to improve long-term stability while allowing inference at arbitrary time strides and demonstrate strong empirical performance across five MD benchmarks.

Critical strengths

The experimental coverage appears very strong, with five quite different benchmarks covering small molecules, proteins, and materials. This is substantially more than prior works I am aware of in this space and reflects positively on the robustness of the proposed method. This reinforces the good long-horizon MSE and other results present in the paper. Overall I would say the empirical elements are the paper's strongest element.

The closed form solution appears to admit efficient inference of the method, as is supported by Table 1. I would have liked to see runtimes present in the other tables as well, however.

Critical weaknesses

The methodological novelty seems limited. In particular, ATOM (ICLR 2026) already introduced a molecular neural operator supporting training and inference across varying temporal resolutions, including unseen time horizons. Given that flexible time-stride prediction is a central contribution of this paper, its omission from both the related work and experimental comparisons is a substantial weakness. Furthermore, Goyal et al. (2025), whom the authors cite, already studied structure-preserving Koopman embeddings with quadratic Hamiltonians and stability guarantees.

The authors also continue to benchmark on the original MD17, despite its documented numerical noise. The creators of rMD17 explicitly caution against relying on the original benchmark, so I believe rMD17 should be used instead.

I appreciate that the authors acknowledge limitations around missing momenta and stochastic dynamics, but would like more discussion of what these actually mean for realistic MD trajectory generation.

References

Thompson, L., Guan, D., Shi, D., Matthews, S., Gao, J., & Han, A. (2026). ATOM: A Pretrained Neural Operator for Multitask Molecular Dynamics. ICLR 2026. https://arxiv.org/abs/2510.05482

Goyal, P., Yıldız, S., & Benner, P. (2025). Deep learning for structure-preserving universal stable Koopman-inspired embeddings for nonlinear canonical Hamiltonian dynamics. Machine Learning: Science and Technology, 6(1), 015063. https://doi.org/10.1088/2632-2153/adb9b5

Christensen, A. S., & von Lilienfeld, O. A. (2020). On the role of gradients for machine learning of molecular energies and forces. Machine Learning: Science and Technology, 1(4), 045018. https://doi.org/10.1088/2632-2153/abba6f

Christensen, A. S., & von Lilienfeld, O. A. (2020). Revised MD17 dataset. Materials Cloud Archive, 2020.82. https://doi.org/10.24435/materialscloud:wy-kn

Questions for authors
  1. Could the authors compare against ATOM and better position their work relative to TITO, both of which already support flexible time-stride molecular dynamics? At minimum, could they better clarify what FlexMD contributes beyond this existing literature?

  2. How much do the deterministic latent dynamics and missing momentum information limit the model's ability to generate realistic MD trajectories? Can any theoretical statements be made about this?

References

The TITO paper: Viguera Diez, J., Schreiner, M., & Olsson, S. (2026). Transferable generative models bridge femtosecond to nanosecond time-step molecular dynamics. Science Advances, 12(15), eaed2333. https://doi.org/10.1126/sciadv.aed2333

2 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer E4w7

We thank the reviewer for the thoughtful and detailed assessment. We appreciate the recognition of our broad experimental coverage across small molecules, proteins, and materials, as well as the strong long-horizon results. We also thank the reviewer for noting the inference efficiency enabled by the closed-form solution. Below, we address each concern and provide clarifications and additional results. We have incorporated the corresponding revisions into the manuscript and highlighted them in blue.

Response to Runtime Comparison

I would have liked to see runtimes present in the other tables as well, however.

Response:

We have updated runtimes for the major experiments in our revised manuscript.

Response to Weakness1 and Question 1 [Novelty and related work]

The methodological novelty seems limited. In particular, ATOM (ICLR 2026) already introduced a molecular neural operator supporting training and inference across varying temporal resolutions, including unseen time horizons. Given that flexible time-stride prediction is a central contribution of this paper, its omission from both the related work and experimental comparisons is a substantial weakness. Furthermore, Goyal et al. (2025), whom the authors cite, already studied structure-preserving Koopman embeddings with quadratic Hamiltonians and stability guarantees. Could the authors compare against ATOM and better position their work relative to TITO, both of which already support flexible time-stride molecular dynamics? At minimum, could they better clarify what FlexMD contributes beyond this existing literature?

Response:

ATOM uses a pretrained transformer neural operator with temporal attention to predict across time horizons. FlexMD instead propagates a geometric molecular representation using a structured continuous-time generator A=JSA=JS, without pretraining. Its closed-form propagator eAτe^{A\tau} defines a consistent family of latent evolutions across strides and preserves a learned quadratic invariant. ITO learns stochastic, lag-conditioned molecular transition distributions using diffusion models, while TITO extends generative transfer-operator learning across molecular systems. FlexMD takes a different approach: its quadratic-invariant linear latent dynamics provide flexible-lag predictions through a matrix exponential and bound the amplification of an initial latent perturbation in the associated SS-norm.

Goyal et al. (2025) study structure-preserving embeddings for nonlinear canonical Hamiltonian systems, using a symplectic transformation to relate physical and latent dynamics. FlexMD instead targets molecular trajectory prediction from geometric configurations with different quadratic constraint designs. FlexMD also incorporates molecular geometry through an equivariant graph-based encoder and decoder. Thus, they differ in their target problem and the constraints placed on their representations.

Our central contribution is therefore a geometric molecular prediction framework in which a structured continuous-time generator enables flexible-stride propagation and controls latent error amplification. This bound concerns perturbations under the learned latent dynamics; it does not by itself guarantee accurate decoded trajectories or eliminate model error.

To examine time-stride extrapolation, we compared ATOM, TITO, and FlexMD on Aspirin from MD17, with all models trained at a fixed stride of ℓ=10\ell=10. ATOM achieves the lowest error at the training stride, while FlexMD has lower error at each larger stride tested (this could also be observed in Experimental 4.3 of the original publication of ATOM):

Test Time-Stride (fs)ATOMTITOFlexMD
102.16e-068.73e-058.16e-06
305.83e-022.876.95e-04
1001.8611.922.19e-03
3003.6716.522.34e-02
1,00019.3457.893.21e-02
3,00036.5289.552.38e-02
10,00047.2897.674.25e-02

We also evaluated these methods on AdK. Models are trained at stride 5 and evaluated at strides 5, 10, 25, and 100; the table reports next-time-stride MSE, with lower values indicating better point prediction.

Test strideGoyal et al.ITOTITOATOMFlexMD
58.24±0.833.85±0.782.67±0.620.68±0.480.50 ± 0.51
1012.35±0.758.38±0.657.49±0.810.88±0.710.61 ± 0.10
2582.76±0.6910.28±0.829.85±0.751.92±0.680.78 ± 0.82
100102.73±0.8329.93±0.5210.23±0.692.84±0.870.91 ± 0.69
AuthorsOpenReview ↗

[2/2] Response to Reviewer E4w7

Response to Weakness 2 [MD17 benchmark]

The authors also continue to benchmark on the original MD17, despite its documented numerical noise. The creators of rMD17 explicitly caution against relying on the original benchmark, so I believe rMD17 should be used instead.

Response:

We agree that rMD17 provides a better benchmark, particularly for machine-learned force fields, as the creator clarified. However, the problem of using MD17 for trajectory prediction remains to be established. More importantly, we have evaluated FlexMD on AdK, ATLAS, MD-CATH, and PET-MAD, which extend the evaluation beyond the small-molecule systems in MD17 and rMD17.

To alleviate the reviewer's concern, we evaluated FlexMD on rMD17 and report the state-to-state (S2S) MSE alongside the published EGNO and ATOM results below (trained and tested on time interval τ=3000\tau=3000 fs). FlexMD achieves the lowest MSE on seven out of eight molecules.

ModelAzobenzeneEthanolMalonaldehydeNaphthaleneParacetamolSalicylicTolueneUracil
EGNO8.96 ± 0.0323.26 ± 0.0140.11 ± 0.051.42 ± 0.0028.08 ± 0.011.06 ± 0.0128.28 ± 0.010.88 ± 0.00
ATOM8.88 ± 0.0523.49 ± 0.1440.29 ± 0.131.36 ± 0.0030.12 ± 0.871.03 ± 0.0028.56 ± 0.040.86 ± 0.00
FlexMD7.27 ± 0.0520.17 ± 0.0440.21 ± 0.071.19 ± 0.0226.36 ± 0.430.97 ± 0.0125.73 ± 0.030.65 ± 0.02

We have added these results to the revised manuscript.

Response to Weakness 3 and Question 2 [Momenta and stochastic dynamics]

W. 3 appreciate that the authors acknowledge limitations around missing momenta and stochastic dynamics, but would like more discussion of what these actually mean for realistic MD trajectory generation. Q. 2 How much do the deterministic latent dynamics and missing momentum information limit the model's ability to generate realistic MD trajectories? Can any theoretical statements be made about this?

Response:

A molecular configuration alone does not uniquely determine its future: different momenta or thermostat states can produce different trajectories from the same coordinates. A deterministic coordinate-only predictor produces a single future for a fixed input and therefore cannot represent this conditional variability. The limitation can be stated formally: In classical MD, the future configuration depends on both the current coordinates and momenta; therefore, the same coordinates can lead to different futures when the momenta differ.

Most existing trajectory datasets do not provide momentum information; therefore, we conduct our experiments following common protocol and discuss this limitation. To assess the value of additional dynamical information, we implemented a momentum-augmented version of FlexMD and evaluated it on PET-MAD. This variant matches or outperforms baselines that also receive momenta. Providing momenta partially addresses the missing-state limitation, but does not eliminate variability due to stochastic dynamics or other unobserved variables.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Program Chairs, Senior Area Chairs, and Area Chairs,

We sincerely appreciate the committee’s time and effort in reviewing our work.

Our paper proposes FlexMD, a molecular dynamics prediction framework: an equivariant molecular representation with structured continuous-time latent dynamics for flexible-stride prediction. We thank Reviewers E4w7, KEVx, and b6EQ for their constructive feedback and their recognition of our experimental coverage, protein results, and constrained Koopman approach.

We summarize the main points of our rebuttal below.

Expanded Discussion and Clarifications

  • Novelty and related work: We clarified FlexMD’s relationship to ATOM, ITO, TITO, and prior structure-preserving Koopman methods. FlexMD combines molecular geometric representations with a quadratic-invariant latent generator and a matrix-exponential propagator for selectable prediction lags.
  • Model details and limitations: We clarified the equivariant decoder in detail. We also discussed the limits of deterministic, the coordinate-only prediction, and our modeling with momenta. A momentum-augmented FlexMD variant matched or outperformed baselines receiving momenta on PET-MAD. On MD-CATH, FlexMD achieved an MSM JSD of 0.17 ± 0.02, compared with 0.19 ± 0.00 for MARS-FM.
  • Evaluation protocols: We clarified baseline training and inference, the TICA metric, and how initial configurations were selected for repeated evaluations.

Additional Experiments and Results

  • Flexible-stride baselines: We added comparisons with ATOM, ITO, TITO, and Goyal et al. FlexMD achieved the lowest next-time-stride MSE at all evaluated AdK strides and lower error than ATOM at each larger stride tested on Aspirin.
  • Additional baselines: We compared FlexMD with PVB, EBA, BioMD, DynaMode, and Sengar et al. FlexMD achieved the lowest reported AdK error at each test stride and the best mean result on five of seven ATLAS metrics.
  • Ensemble and temporal correlations: We added ensemble metrics and autocorrelation comparisons. FlexMD achieved the lowest autocorrelation deviations at lags 80 and 120; these results provide a limited assessment of temporal correlations.
  • Runtime and visualizations: We added runtime information for major experiments, corrected the Figure 2a legend, and added a TICA visualization of reference and generated trajectories.

We hope these clarifications and additional results help the committee evaluate the contributions of our work.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.