acceptodds
Under review as a conference paper at ICLR 2027

Thermo-Symplectic Neural Langevin Flows

20 views

Abstract

Molecular dynamics (MD) underlies biomolecular simulation by modeling atomistic motion over physical time. In thermostatted settings, this evolution is commonly formulated as underdamped Langevin dynamics, whose sequential force-driven integration makes long horizon simulation computationally expensive. While learning based approaches use MD data to model ensembles or state transitions, their computational primitive is defined over empirical state transitions rather than the Langevin update itself. We propose a Thermo-Symplectic Neural Langevin Flow (TSL-Flow), a learnable framework that models underdamped Langevin dynamics through phase-space updates designed to preserve thermo-symplectic structure. Our approach lifts local stochastic updates into Lie algebraic summaries, allowing the learned dynamics to preserve the geometric structure of phase-space while exposing an associative composition across time. To enable accurate and efficient long horizon simulation, we use this compositional structure to define a prefix scan compatible product that propagates trajectory segments in parallel. Empirically, our method improves coordinate and physical fidelity across biomolecular benchmarks and achieves the lowest inference time. These results indicate that temporal efficiency in learned molecular simulation can be improved by representing the Langevin update as a composable stochastic physical primitive.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
thermo-symplectic symmetry, conformal symplectic geometry, Lie-Stratonovich stochastic flows, molecular dynamics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WmpHOpenReview ↗
Rating
4(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper reformulates step-by-step state prediction in molecular dynamics as a sequence of composable transformations with explicit physical and geometric structure, enabling parallel trajectory generation through a parallel prefix-scan. The method combines physically structured constraints with parallel computation, with comprehensive experiments reporting improvements in both accuracy and computational efficiency.

Critical strengths
  1. Physically structured constraints with a principled mathematical foundation. The method addresses Langevin dynamics with friction and thermal noise by constructing learnable updates using the conformal symplectic group and its Lie algebra, incorporating explicit physical and algebraic inductive biases.

  2. Efficient temporal computation. The method organizes local updates into composable algebraic representations and uses a parallel prefix-scan to reduce sequential depth along the temporal dimension, enabling more efficient trajectory generation.

  3. Comprehensive evaluation across multiple metrics. Beyond coordinate MSE and RMSD, the paper evaluates interatomic distance distributions, conformal symplectic residuals, and backbone dihedral angles, assessing the structural plausibility of generated trajectories from multiple perspectives.

Critical weaknesses
  1. Evidence for physical consistency is largely limited to geometric and conformational statistics. Conformal symplectic residuals, distance distributions, and dihedral angles support structural plausibility but do not establish that the model reproduces the correct stochastic dynamics. Important properties, including temperature and momentum distributions, equilibrium distributions, temporal autocorrelations, and state-transition kinetics, remain insufficiently evaluated.

  2. Runtime measurements and reporting require clarification. Some runtimes and speedup factors are inconsistent across the main tables, Figure 6, and the accompanying text. The scope of the reported training times is also unclear.

  3. The paper introduces numerous symbols and operations, making it difficult to follow for readers without the relevant background.

Questions for authors
  1. Beyond satisfying the imposed geometric constraint, what physical guarantees follow from restricting the updates to conformal symplectic affine transformations? For example, does this ensure the correct temperature distribution, consistency with the underlying potential energy, or other thermodynamic properties? The current evidence appears to rely primarily on geometric diagnostics and the quality of the generated conformations, rather than direct validation of these physical properties.

  2. The main text formulates learning as minimizing a path-space KL divergence between full trajectory distributions, whereas the implementation uses positional MSE. Could the authors explain the relationship between these objectives and clarify how the practical training procedure relates to the theoretical formulation?

  3. The discussion and evaluation of existing parallel trajectory-prediction methods appear incomplete, limiting the assessment of the claimed acceleration benefits. Could the authors include additional comparisons with relevant state-of-the-art parallel prediction baselines, such as EGNO [1], HEGNN [2], PAINET [3], and FlashMD [4], under a common evaluation protocol?

  4. Some reported results appear inconsistent with the stated speedups. The text on page 9 claims an approximately 13× speedup over GLDP, whereas the main tables indicate only approximately 1.1×. Moreover, Figure 6 labels the PVB runtime as approximately 1,150 seconds, which does not appear consistent with the stated 44× speedup. Could the authors reconcile these numbers and clarify the timing protocols?

[1] Equivariant graph neural operator for modeling 3d dynamics

[2] Are high-degree representations really unnecessary in equivariant graph neural networks?

[3] Physics-inspired all-pair interaction learning for 3d dynamics modeling

[4] FlashMD: long-stride, universal prediction of molecular dynamics

3 comments

Thank you for your careful reading and for raising important questions. We believe the responses below clarify our position and address the points raised in your review. We look forward to continuing the discussion and addressing any remaining questions.


W1/Q1. Evidence for physical consistency is largely limited to geometric...Beyond satisfying the imposed geometric constraint...

Faithful production of MD trajectories requires not only accurate molecular conformations but also consistency with the physical characteristics of the underlying dynamics. In this regard, CSP is included as a physically meaningful measure of consistency with the conformal symplectic behavior inherent in Langevin dynamics, rather than merely as a geometric diagnostic. Together with the established MSE, RMSD, and Pair JSD metrics, our evaluation therefore provides complementary evidence of trajectory accuracy and physical fidelity.

To further substantiate these findings, we conducted additional ensemble distribution evaluations using Root mean W2\mathcal W_2 and MD PCA W2\mathcal W_2 on ATLAS, as detailed in Table 1. TSL-Flow achieves the lowest discrepancies across both metrics, indicating improved agreement with the reference molecular distributions.

Table 1. Additional Conformational Distribution Evaluation on ATLAS.

MethodRoot mean W2\mathcal W_2 ↓MD PCA W2\mathcal W_2 ↓
MDGen4.04×1004.04 \times 10^05.49×10−15.49 \times 10^{-1}
GLDP8.23×10−18.23 \times 10^{-1}5.22×10−15.22 \times 10^{-1}
Ours (TSL-Flow)4.91×10−1\mathbf{4.91 \times 10^{-1}}1.02×10−1\mathbf{1.02 \times 10^{-1}}

We additionally performed a tICA-based analysis to assess the agreement of slow collective modes between generated and reference trajectories. As shown in Table 2, TSL-Flow achieves the lowest TICA-JSD and TICA-W2\mathcal W_2 across all compared methods, indicating improved fidelity in the conformational distributions associated with slow molecular motions. These findings are consistent with our main experimental results. The corresponding tICA visualizations are provided in Appendix A.3 of the revised manuscript.

Table 2. Comparison of slow collective dynamics on ATLAS.

MethodTICA-JSD ↓TICA-W2\mathcal W_2 ↓
MDGen8.90×10−18.90 \times 10^{-1}3.32×1003.32 \times 10^{0}
MarS-FM7.19×10−17.19 \times 10^{-1}2.44×1002.44 \times 10^{0}
UniSim8.12×10−18.12 \times 10^{-1}2.42×1002.42 \times 10^{0}
GLDP5.18×10−15.18 \times 10^{-1}7.89×10−17.89 \times 10^{-1}
TSL-Flow4.00×10−1\mathbf{4.00 \times 10^{-1}}6.23×10−1\mathbf{6.23 \times 10^{-1}}

W2/Q4. Runtime measurements and...Some reported results appear inconsistent with the stated speedups...

We thank the reviewer for identifying the error in Figure 6. The inference time was measured as the wall-clock time required for each model to generate a trajectory up to the specified number of timesteps. However, some values were inadvertently entered incorrectly in Figure 6, leading to discrepancies in the reported speedup factors. Based on the corrected results, TSL-Flow achieves an approximately 1.4×1.4\times inference speedup over GLDP, with a substantially larger advantage over PVB. Importantly, these corrections do not alter the overall conclusion of our ablation study regarding the computational advantage of TSL-Flow. We have updated Figure 6 and the corresponding statements accordingly, and appreciate the reviewer’s careful observation.

W3. The paper introduces numerous symbols

The mathematical notation and operations introduced in this work are necessary to rigorously formulate a framework that unifies stochastic Langevin dynamics, conformal symplectic geometry, and Lie-algebraic temporal composition. While we consider these mathematical components essential to the proposed formulation, we recognize that some of the notation may be less familiar to readers from different backgrounds. We will simplify some of the more notation-heavy parts to improve readability.

Q2. The main text formulates learning as minimizing a path-space KL divergence...

We believe this concern stems from an imprecise description in Table 5 and the Training Setup section. We mistakenly referred to the full phase-space state XtX_t as coordinates, which led to the incorrect description of the training loss and input representation. In fact, the input representation is X0X_0, and the training loss is computed between XtX_t and YtY_t, consistent with the phase-space formulation underlying our theoretical path-space KL objective. We apologize for the confusion and will correct these descriptions accordingly.

Q3. The discussion and evaluation of existing parallel trajectory-prediction methods appear incomplete, ...

The suggested methods address different aspects of learning-based 3D dynamics prediction. In particular, HEGNN primarily focuses on enhancing the expressivity of equivariant representations, while FlashMD accelerates molecular simulations through long-stride state prediction rather than temporal parallelization. In contrast, EGNO and PAINET are more closely aligned with our objective, as they enable parallel prediction of future molecular coordinates.

Accordingly, we conducted additional experiments with these two methods under the same evaluation protocol on ATLAS.

Table 3. Comparison with parallel trajectory-prediction methods on ATLAS.

MethodMSE ↓RMSD ↓CSP ↓Inference Time (s) ↓
EGNO1.06×10−21.06 \times 10^{-2}1.32×10−11.32 \times 10^{-1}2.39×10−12.39 \times 10^{-1}3.17×1013.17 \times 10^{1}
PAINET1.83×10−21.83 \times 10^{-2}1.58×10−11.58 \times 10^{-1}$9.66 \times10^{-1} $1.27×1011.27 \times 10^{1}
TSL-Flow1.48×10−3\mathbf{1.48 \times 10^{-3}}6.00×10−2\mathbf{6.00 \times 10^{-2}}1.19×10−2\mathbf{1.19 \times 10^{-2}}4.01×100\mathbf{4.01 \times 10^0}

As shown in Table 3, TSL-Flow consistently outperforms both parallel prediction baselines across all evaluated metrics, achieving higher fidelity and lower inference time. This is consistent with the findings of our main experiments.

Thank you for the additional clarification and experimental results, which have addressed most of my concerns. With these results incorporated into the final version, I believe the paper offers a solid theoretical foundation and strong empirical validation. I will therefore raise my score.

Reviewer_Axc3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a new model, TSL-Flow, for efficiently generating long-horizon simulations of biomolecules. This flow utilizes Lie algebra to better model the dynamics and also proposes a new technique to accelerate sampling. The experiments also suggest the effectiveness of the proposed model.

Critical strengths

This method sounds technical and interesting. Applying Lie algebra to model the MD simulation and using Rough Path Theory + Prefix Scan to accelerate trajectory sampling sounds novel and useful.

Critical weaknesses

I am not an expert in Lie algebra, so my concerns are mainly from the experimental analysis.

  1. This paper has a weak evaluation of experiments. Some important baselines are missing, like BioMD[1], EBA[2], ConfRover[3], etc.
  2. The metrics are limited in evaluating the generated trajectories. At the ensemble level, Alphaflow proposed a lot of metrics that can comprehensively evaluate the model, so it is better to evaluate the proposed model on those metrics, since they are commonly used metrics. Further, considering the model is a trajectory generation method, it would also mention how the model performs on kinetics. MSM is a good tool to do the analysis.
  3. The paper claims long-horizon generation, but didn’t specify how long the generated trajectory. From my understanding, a mu-second-level trajectory will be useful to explore the state transition and can be said to be relatively long-horizon. Hundreds of ps cannot be regarded as long-horizon.

[1]BioMD, ILCR2026 [2] Energy-based Alignment (ICML 2025) [3]Simultaneous modeling of protein conformation and dynamics via autoregression

Questions for authors

Is the proposed method generalizable to new systems? If so, the model's application scope will be limited, since it relies on transitional MD to generate the training dataset when it needs to be used for a new system.

4 comments

We appreciate your careful reading of our work and your thoughtful comments. Our responses below provide further explanation and clarify the reasoning behind our approach. Please let us know if there are any further points you would like us to address.


W1. This paper has a weak evaluation of experiments...

We believe our experimental evaluation is sufficiently comprehensive, as our study compares against six prior methods, exceeding the number of baselines evaluated in several closely related works. More importantly, BioMD has not released its implementation code, while ConfRover has not released its training code. It is therefore unclear how these methods could be expected to serve as fair baselines when the necessary implementations for reproducing and retraining them under identical experimental conditions are unavailable. We would therefore appreciate clarification on the criteria by which these particular methods are considered necessary baselines.

Additionally, among the suggested methods, EBA provides a publicly available implementation, and we have conducted an additional comparison. As shown in Table 1, TSL-Flow consistently achieves superior performance across all metrics, even with the inclusion of EBA as an additional baseline. These results further reinforce the findings of our evaluation.

Table 1. Additional baseline comparison on ATLAS.

MethodMean MSE ↓Mean RMSD ↓CSP ↓Inference Time ↓
GLDP4.36×10−34.36 \times 10^{-3}1.09×10−11.09 \times 10^{-1}7.48×10−27.48 \times 10^{-2}4.59×1004.59 \times 10^{0}
EBA8.92×10−38.92 \times 10^{-3}1.06×10−11.06\times 10^{-1}1.02×10−11.02 \times 10^{-1}1.75×1031.75\times 10^{3}
Ours (TSL-Flow)1.48×10−31.48 \times 10^{-3}6.00×10−26.00 \times 10^{-2}1.19×10−21.19 \times 10^{-2}4.01×1004.01 \times 10^{0}

W2. The metrics are limited in evaluating the generated trajectories...

Our evaluation already includes well-established metrics such as MSE and RMSD, alongside Pair JSD, providing a comprehensive assessment of coordinate accuracy and distributional fidelity. These complementary metrics offer a reliable basis for evaluating the quality of generated molecular structures.

To further validate the agreement in conformational distributions, we additionally evaluated TSL-Flow using the ensemble-level metrics adopted in AlphaFlow, as reported in Table 1. TSL-Flow consistently achieves the lowest errors across both metrics, further supporting the findings of our original evaluation.

Table 1. Additional Conformational Distribution Evaluation on ATLAS.

MethodRoot mean W2\mathcal W_2 ↓MD PCA W2\mathcal W_2 ↓
MDGen4.04×1004.04 \times 10^05.49×10−15.49 \times 10^{-1}
GLDP8.23×10−18.23 \times 10^{-1}5.22×10−15.22 \times 10^{-1}
Ours (TSL-Flow)4.91×10−1\mathbf{4.91 \times 10^{-1}}1.02×10−1\mathbf{1.02 \times 10^{-1}}

To complement the ensemble-level evaluation, we also employed time-lagged independent component analysis (TICA) to assess the recovery of slow collective dynamics. As shown in Table 2, TSL-Flow achieves the lowest TICA-JSD and TICA-W2\mathcal W_2, reducing errors by approximately 22.8% and 21.0%, respectively, compared with GLDP. These results further demonstrate that our method effectively captures the reference distributions along slow collective coordinates. The corresponding TICA visualizations are included in Appendix A.3 of the revised manuscript.

Table 2. Comparison of slow collective dynamics on ATLAS.

MethodTICA-JSD ↓TICA-W2\mathcal W_2 ↓
MDGen8.90×10−18.90 \times 10^{-1}3.32×1003.32 \times 10^{0}
MarS-FM7.19×10−17.19 \times 10^{-1}2.44×1002.44 \times 10^{0}
UniSim8.12×10−18.12 \times 10^{-1}2.42×1002.42 \times 10^{0}
GLDP5.18×10−15.18 \times 10^{-1}7.89×10−17.89 \times 10^{-1}
TSL-Flow4.00×10−1\mathbf{4.00 \times 10^{-1}}6.23×10−1\mathbf{6.23 \times 10^{-1}}

W3. The paper claims long-horizon generation...

In our experiments, we sample frames at intervals of 100 integration steps, with an integration timestep of 2 fs, resulting in trajectories of 10,240 frames. Thus, each generated trajectory spans approximately 2.05 ns of physical time. Moreover, our model predicts molecular states across 10,240 time points, representing a substantial prediction horizon that would incur considerable computational costs under conventional sequential generation. In this context, the term long-horizon generation refers to prediction over a large number of time steps and the associated temporal computational complexity.

Q1. Is the proposed method generalizable to new systems?

The reliance on MD-generated training data is not unique to our method, but is common across existing learned molecular dynamics approaches, including those pursuing broader transferability. This reflects the broader data-driven nature of learned molecular simulation. Achieving reliable long-horizon predictions across arbitrary, unseen molecular systems with a single trained model remains a fundamental and highly challenging open problem in learned molecular dynamics. To the best of our knowledge, no existing method has successfully demonstrated such generalization. Our contribution instead centers on preserving the conformal symplectic structure inherent in Langevin dynamics and exploiting its compositional properties for efficient long-horizon simulation. While cross-system generalization is distinct from the focus of the present work, it remains an important research direction in its own right.

Thanks for the response! I was wondering why you are using the partial ensemble metrics from Alphaflow/EBA. Could you report all of them on ATLAS?

I still think that 2.05ns and the only constraint on a single system are the main limitations of the paper. Like GPCR MD, several microseconds might be needed to observe the important state transition, and I think for single-system tasks, the user might hope the ML-based method can extend to very long-horizon generation. I tend to maintain my score.

Reviewer_fZ8jOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TSL-Flow, a learned surrogate for under damped Langevin equations for MD of protein structures. A network predicts a drift e(X_0, t) in the Lie algebra, with a diffusion Sigma(t). These drive a right-invariant Stratonovich SDE on the affine conformal-symplectic group, and the phase-space path is that group's action on X_0. In simple terms, the authors use a predictor to predict time indexed Lie algebra elements which can be cheaply combined. This is done through a prefix scan for the Lie algebra elements, which is a basic algorithmic speedup to compute a cumulative sum. The group choice is meant to impose "thermo-symplectic" structure (Def. 2.1): contraction of the symplectic form at rate γ. Per-segment log-signature increments are composed with BCH in a prefix scan, claimed to give O(log M) sequential depth. One model is trained per system, with position MSE, on the authors' own OpenMM trajectories (5 mdCATH systems and 5 ATLAS structures plus one MISATO complex). The paper reports about 3× lower MSE than GLDP (Tables 2–3), better protein–ligand metrics than UniSim and PVB (Tables 4, 6), and the lowest inference and training times.

Critical strengths

The symplectic claim is well developed and strongly motivated by theory. The generators satisfy XᵀJ + JX = λJ, so conformal symplecticity holds by construction and is not learned. This is a clean way to build a structured MD surrogate, and the paper gives explicit coordinate formulas for the truncated log-signature. It matters because it would give physically constrained paths without a penalty term, if the structure could be shown to help (see Weakness 1).

Critical weaknesses
  1. The model is a per-system, open-loop trajectory regressor, and the Langevin framing, the KL objective and the scan do not describe what is trained. So essentially it is a trajectory based finetuning method. The predictor e depends on (X_0, t) and not on the state Y_t, so the model is Y_t = A_t X₀ + a_t with no force or state feedback along the path. Increments for each time slice are predicted independently from X_0 and t and then composed in the Lie Algebra space. Apart from the coordinate MSE, nothing ties consecutive slices to a Markovian dynamics. Unfortunately due to the trajectory based finetuning, it does not transfer which is a difficult proposition in times of large time step based force field integrators or models like BioEmu and AlphaFlow. Ultimately, we want to sample the equilibrium distribution of such structures and MD is just a tool to get there. While compute is a good argument a lot of work has gone into making these newer trajectory cover larger time frames.
  2. The speed comes from non-autoregressive predictions. Any (X_0, t) to trajectory regressor has the same property. A temporal U-Net produces all steps in a single pass, but its cost still grows linearly with the horizon. The paper's own scan vs non-scan comparison (Fig. 8) shows both growing linearly, with a constant gain of about 1.6× and no log-depth behaviour. The scan is built on BCH/cumulative sum.
  3. Since the trajectory learner trains against a (seemingly) single trajectory it's also not entirely clear how much fidelity this setup can recover from the actual underlying dynamics. Similarly the long horizon is a bit shaky as 10,240 steps at 2 fs is about 20 ps which is decidely a short path.
Questions for authors
  1. Is my understanding of the core data and training setup correct? In particular about the use of a single trajectory and the linear Lie algebra elements being applied to the Lie group elements?
  2. Does this setup generalize such that a single TSL-FLOW model can predict arbitrary structures into the future at meaningful time horizons?
  3. Can you run the TSL-Flow model against many realizations and report whether such a model is able to recover for example the same TICA plots?
5 comments

Thank you for your detailed review and the time devoted to our work. Below, we respond to each point and explain the reasoning behind our position. We welcome further discussion of any remaining questions or concerns.


W1. The model is a per-system, open-loop trajectory regressor...

Underdamped Langevin dynamics provides a well-established physical foundation for thermostatted molecular simulation [1]. Importantly, its stochastic flow possesses an intrinsic conformal-symplectic structure that reflects the dissipative nature of the underlying dynamics [2]. Preserving this geometric structure is therefore a fundamental consideration for physically consistent Langevin simulation. Howerver, while existing learned molecular models have emphasized transferability and generative capabilities, the geometric consistency of the underlying stochastic dynamics has generally been addressed only implictly or partially in a probabilistic sense. In contrast, TSL-Flow explicitly incorporates the conformal-symplectic structure of underdamped Langevin dynamics into its stochastic update formulation, treating consistency as a fundamental design principle rather than an emergent consequence of trajectory fitting.

Additionally, our formulation induces a Markovian stochastic evolution. Specifically, the Lie-Stratonovich flow satisfies

Yt=ρGt(X0)=ρGtGs−1(Ys),0≤s<t≤TY_t = \rho_{G_t}(X_0) = \rho_{G_t G_s^{-1}}(Y_s), \quad 0\le s < t \le T

This relation naturally induces a conditional transition probability. Since GtGs−1G_tG_s^{-1}, conditioned on X0X_0, is driven by Brownian increments over [s,t][s,t] independently of the past, we obtain

Pθ(Yt∈B∣Fs,X0)=E[1B(ρGtGs−1(Ys))∣Ys,X0]=P(Yt∈B∣Ys,X0)\mathbb{P}_{\theta}(Y_t \in B \mid \mathcal{F}_s, X_0) = \mathbb{E}[\mathbf{1}_B(\rho_{G_t G_s^{-1}}(Y_s)) \mid Y_s, X_0] = \mathbb{P}(Y_t \in B \mid Y_s, X_0)

where B⊆MB\subseteq\mathcal M is a measurable subset of phase space, Fs\mathcal F_s denotes the information available up to time ss, and 1B\mathbf 1_B is the indicator function of BB. This establishes a time-inhomogeneous Markov evolution conditional on X0X_0. Thus, consecutive states are connected through stochastic phase-space transitions rather than independently regressed coordinates.

Viewing molecular dynamics solely as a tool for equilibrium sampling reflects a potentially misleading understanding of its scientific purpose. Understanding how molecular systems evolve is itself a central scientific objective, as demonstrated by studies of protein folding [3] and protein–ligand association and dissociation [4]. Equilibrium distributions alone cannot resolve the temporal sequence of molecular events, including ligand binding and unbinding pathways and the intermediate states encountered during these processes. Restricting the purpose of molecular dynamics to equilibrium sampling therefore overlooks its essential role in elucidating the mechanisms of molecular motion and interaction.

We hope the above clarifications help address the underlying concerns and provide a clearer perspective on our contributions. Given their relevance to the central points raised in the review, we would appreciate it if the reviewer could revisit the manuscript.


References

[1] Leimkuhler, B., & Matthews, C. (2015). Molecular dynamics: with deterministic and stochastic numerical methods (Vol. 39). Cham: Springer.

[2] Hong, J., Sun, L., & Wang, X. (2017). High order conformal symplectic and ergodic schemes for the stochastic Langevin equation via generating functions. SIAM Journal on Numerical Analysis, 55(6), 3006-3029.

[3] Lindorff-Larsen, K., Piana, S., Dror, R. O., & Shaw, D. E. (2011). How fast-folding proteins fold. Science, 334(6055), 517–520.

[4] Dror, R. O., Pan, A. C., Arlow, D. H., Borhani, D. W., Maragakis, P., Shan, Y., Xu, H., & Shaw, D. E. (2011). Pathway and mechanism of drug binding to G-protein-coupled receptors. Proceedings of the National Academy of Sciences, 108(32), 13118–13123.

W2. The speed comes from non-autoregressive predictions...

As demonstrated by the equations in our response to W1, TSL-Flow is fundamentally different from an arbitrary trajectory regressor. Equating these methods merely because they support non-autoregressive inference is a superficial comparison that overlooks the essential physical properties our formulation preserves.

Overall inference time and the sequential computational depth of composition are distinct measures of computational efficiency. The former includes the cost of predicting Lie-algebraic increments, which grows linearly with the prediction horizon as longer trajectories require more local increments. In contrast, our contribution lies in enabling these increments to be composed in parallel through a prefix-scan, reducing the sequential depth of composition from O(M)O(M) to O(log⁡M)O(\log M). Thus, the linear growth in overall inference time does not contradict the logarithmic sequential depth achieved by the proposed composition mechanism. Fig. 8 was specifically designed to isolate this computational benefit by comparing parallel scan-based composition with sequential composition. This scan-compatible formulation contributes to the faster inference achieved by our method relative to the evaluated baselines.

W3/Q1. Since the trajectory learner trains against...Is my understanding of the core data...

We would like to clarify that our model was not trained on a single trajectory. Instead, multiple trajectories were used for training and evaluation. Moreover, the 10,240 frames do not correspond to consecutive MD integration steps but were sampled every 100 steps from simulations with a 2 fs timestep. Thus, the actual prediction horizon is approximately 2 ns rather than 20 ps. We will clarify them in the revision.

Regarding the Lie-algebraic formulation, we use the inhomogeneous conformal-symplectic Lie algebra, incorporating both linear and translational components. These Lie-algebraic increments are mapped to Lie-group elements via the exponential map and subsequently composed through group multiplication.

Q2. Does this setup generalize such that a single...

Cross-system generalization is a separate research challenge from achieving long-horizon simulation for the molecular systems considered in this work. Our experiments demonstrate trajectory prediction extending to 10,240 steps, which already represents a substantial temporal horizon. To the best of our knowledge, no existing model can reliably predict such trajectories across arbitrary molecular systems using a single set of learned parameters. Achieving this level of generalization remains a fundamental open challenge in learned molecular dynamics, rather than an established capability of current approaches. Our work does not aim to develop a general-purpose molecular foundation model, but instead focuses on enforcing structural consistency with the underlying Langevin transition mechanism and enabling efficient long-horizon simulation. We regard broader generalization across molecular systems as an important direction for future research.

Q3. Can you run the TSL-Flow model against many...TICA plots?

We conducted additional experiments to evaluate whether TSL-Flow recovers the slow collective dynamics across multiple trajectory realizations using time-lagged independent component analysis (TICA). As shown in Table 1, TSL-Flow achieves the lowest TICA-JSD and TICA-W2\mathcal W_2 among all evaluated methods, outperforming the strongest baseline, GLDP, by 22.8% and 21.0%, respectively. These results indicate that TSL-Flow better captures the distribution of slow collective motions in reference MD trajectories, demonstrating fidelity beyond coordinate-level reconstruction. We have included the quantitative comparisons and corresponding TICA visualizations in Appendix A.3 of the revised manuscript.

Table 1. Comparison of slow collective dynamics on ATLAS.

MethodTICA-JSD ↓TICA-W2\mathcal W_2 ↓
MDGen8.90×10−18.90 \times 10^{-1}3.32×1003.32 \times 10^{0}
MarS-FM7.19×10−17.19 \times 10^{-1}2.44×1002.44 \times 10^{0}
UniSim8.12×10−18.12 \times 10^{-1}2.42×1002.42 \times 10^{0}
GLDP5.18×10−15.18 \times 10^{-1}7.89×10−17.89 \times 10^{-1}
TSL-Flow4.00×10−1\mathbf{4.00 \times 10^{-1}}6.23×10−1\mathbf{6.23 \times 10^{-1}}

Thank you for the clarifications. I take the corrections that frames are sampled every 100 steps (about 2 ns), that several trajectories are used, and that increments are composed by group multiplication.

My central concern is unchanged. The model receives only X₀ and t, as the response to Reviewer WmpH confirms ("the input representation is X₀"). It is therefore a trajectory prediction model, not a Langevin simulator, and the paper's claims should be judged as such.

  1. It is a trajectory regressor. Given X0, the drift e(X0, t) is fixed before the trajectory starts. Along the path, the model is a linear SDE, Y_t = A_t X0 + a_t plus noise, with no dependence on the current state or on ∇U. The Markov argument confirms this: the transition kernel depends on X₀, which the transition kernel of Langevin dynamics does not. The model cannot respond to where the protein actually is.
  2. This is what makes the scan possible, and it is also the limitation. A cumulative product of affine maps fixed in advance is Blelloch's first-order linear recurrence. Force-driven, state-dependent dynamics cannot be scanned without iterative linearisation (DEER, Lim et al., ICLR 2024; ELK, Gonzalez et al., NeurIPS 2024). The speed and the open loop are the same design choice, and Fig. 8 shows only a constant-factor gain.
  3. A trajectory regressor cannot transfer, and alternatives exist. ITO (Schreiner et al., NeurIPS 2023) uses the same per-system data but conditions on the current state, gains speed through large lags, and is evaluated on kinetics. It is missing as related work and as a baseline. Models that transfer across systems also exist, contrary to the rebuttal: MDGen (unseen tetrapeptides), PVB (zero-shot fast-folders) and Timewarp (unseen peptides). MDGen, MarS-FM and UniSim also report speedups over MD, contrary to Table 1.
  4. The metrics reward fitting a trajectory, and the results look like it.
    • Per-frame MSE against one stochastic realisation favours regression: the conditional mean scores \sigma^2, an exact sampler 2σ².
    • At 400 K, replica trajectories diverge by about 1 Å within picoseconds (Braxenthaler et al., Proteins 1997). Over 2 ns, staying within 0.60 Å (Tables 2–3) or 0.15 Å (Table 6) of a held-out realisation is not possible for a correct stochastic simulator. The near-zero error maps in Fig. 7 point the same way.
    • The TICA-JSD of 0.40 is hard to reconcile with these per-frame errors.
    • The conformal-symplectic constraint restricts A_t but says nothing about forces. The reported nonzero CSP, which equals 2n(lambda + gamma)^2, implies lambdda ≠ −gamma, so Def. 2.1 is not met.
  5. The comparison is not like-for-like. Baselines built for 80 ps to 50 ns strides, often pretrained on CHARMM data at 300 K, are scored at 200 fs on ff19SB data at 400 K. CSP is undefined for position-only baselines. Only GLDP's timing is comparable, with a margin of 1.14–1.17× (Tables 2–3), not 13 × (text) or 1.4× (rebuttal).

Questions that could change my assessment

  1. Are evaluation trajectories generated with seeds not used in training? Please report Eq. 7 for an independent MD replicate from the same X0 and for the static baseline Y_t = X0.
  2. Started from a frame of a held-out trajectory, a state never used as a training X0, does the model's accuracy hold?
  3. Can you report implied timescales or autocorrelation functions against the reference MD?

Thank you for doing the extra experiments, I'm raising my score by +1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.