acceptodds
Under review as a conference paper at ICLR 2027

PhysGNO: Physics-Guided Graph Neural Operators for Molecular Dynamics

Abstract

Molecular dynamics (MD) plays a central role in scientific applications such as drug discovery, materials design, and chemical reaction analysis. Recent advances in deep learning have further expanded the capabilities of MD modeling. In particular, graph neural network-based interatomic potentials have achieved remarkable success in modeling molecular interactions, enabling accurate prediction of molecular energies and forces while effectively preserving physical information. However, these methods still rely on a large number of iterative integration steps to simulate long-horizon trajectories. Neural operators provide a complementary paradigm by learning trajectory evolution without step-by-step integration, thereby offering significantly improved efficiency. Nevertheless, such data-driven approaches often lack sufficient physical priors, which limits their ability to produce accurate molecular trajectories. To address these challenges, we propose PhysGNO, a physics-guided neural operator for molecular trajectory learning. PhysGNO leverages physical information captured by potential energy networks to guide trajectory evolution, combining the efficiency of neural operators with physically meaningful inductive bias for accurate trajectory prediction. In addition, PhysGNO incorporates molecular flexibility and rich vector-valued features into its spatiotemporal evolution layers, making it better suited to molecular dynamics systems. Extensive experiments on MD17, rMD17, and MD22 demonstrate the effectiveness and efficiency of the proposed method.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular dynamics, AI for sicence

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rdE5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposed physics-guided graph neural operators for molecular dynamics based on pretrained potential energy networks. The author conducted experiments on MD17, rMD17, and MD22.

Critical strengths
  1. The motivation is good: balancing the efficiency of neural operators and the physical information.
  2. The idea is intuitive and simple.
Critical weaknesses
  1. In my opinion, the design has a critical weakness: it depends on a pretrained potential model. First, this adds training and deployment costs. Second, comparisons with trajectory-generation models may be unfair if these additional costs, data, or pretrained resources are not accounted for. Third, the contribution of the potential model is not fully evaluated, particularly its effect on generalization. When applying PhysGNO to a new molecular system, is a new potential model required?
  2. The paper does not discuss sufficient related work. FlashMD, MDGen, GGND, and other recent, closely related works should be discussed, and their differences from PhysGNO should be clarified.
  3. The main manuscript’s metrics, F-MSE and A-MSE, are insufficient for evaluating molecular dynamics. Molecular dynamics is stochastic, and reproducing one particular trajectory does not by itself establish that the generated trajectory is physically correct.
  4. Many important aspects of MD are not evaluated: performance in different ensembles, Markov states of the obtained trajectory, and the modeling incompleteness (without momenta) issue.
  5. The scales evaluated are limited. MD17 contains small molecules, while MD22 includes systems with up to 370 atoms. The temporal horizons also appear limited to the datasets’ timescales. Even though ab initio MD may be feasible at these scales for some systems, the experiments do not fully demonstrate the efficiency advantages of neural operators at larger spatial scales or over longer timescales.
Questions for authors
  1. What are the training and inference costs of the pretrained potential model, and are these included in the efficiency comparisons?

  2. Does PhysGNO need a separately trained potential model for each new molecule? How well does it generalize to unseen molecular systems?

  3. Why did the paper not discuss recent work such as FlashMD, MDGen, and GGND?

  4. Why did the paper put "matching reference coordinates" as the main metrics in the main manuscripts?

  5. Why did the paper not consider important evaluation aspects of MD simulations: evaluate in certain ensembles, evaluate the Markov states, or transition behavior?

  6. How does PhysGNO’s efficiency scale to larger systems and longer simulation times?

10 comments

Dear Reviewer rdE5,

Thank you for your insightful comments and valuable feedback.

W1

W1-1: Dependence on Pretrained Potential Model

  • First, we would like to emphasize that a central contribution of PhysGNO is the Physical Conditioning mechanism, which is designed to be largely independent of the specific pretrained potential architecture. The core motivation of PhysGNO is to effectively leverage the physical knowledge encoded in pretrained molecular potential models while retaining the efficiency of finite-time neural operators.

  • Although conventional ML-based potentials can accurately predict molecular energies and atomic forces, generating complete molecular trajectories still requires repeated force evaluations and sequential numerical integration at microscopic time steps. In contrast, PhysGNO incorporates this physical information into a neural operator through Physical Conditioning, thereby reducing the reliance on repeated numerical integration.

  • Moreover, as shown in Table 4, Physical Conditioning is compatible with widely used molecular potential models, including PaiNN, MACE, and NequIP, demonstrating its applicability across different pretrained potential architectures. The corresponding results are provided below.

PotentialAspirinEthanolToluene
None6.333.425.05
PaiNN5.512.222.32
MACE5.962.253.87
NequIP5.832.454.10

W1-2: Training and Deployment Cost

  • We acknowledge that pretrained molecular potential models introduce additional training cost and computational overhead. However, for long-horizon molecular trajectory generation, the major computational bottleneck of conventional potential-based methods arises from repeated force evaluations and sequential numerical integration during inference. In addition, some pretrained potential models have publicly available checkpoints, which alleviate the need for training from scratch and partially reduce the training cost.

  • As shown in Table 6, PhysGNO requires 12.10 ms per trajectory, compared with 199.46 ms for PaiNN rollout, corresponding to approximately 16.5× faster inference.

  • To further evaluate the inference efficiency of the Physical Conditioning mechanism, we additionally report the inference time in Table 6 using different potential networks. The detailed results are shown below. These results further demonstrate that the Physical Conditioning mechanism introduces only limited inference overhead, while PhysGNO remains substantially more efficient than conventional potential-based methods.

MethodTime
PaiNN Rollout199.46
PhysGNO w/o PC10.78
PhysGNO + PaiNN12.10
PhysGNO + MACE13.51
PhysGNO + NequIP12.44

W1-3: Comparison Fairness

  • To ensure a fair comparison and further verify the effectiveness of the PhysGNO backbone itself, we have already investigated this issue in Appendix C.3 (Table 10). Specifically, we apply the same pretrained potential network and Physical Conditioning mechanism to other trajectory-generation models, such as EGNO and PAINET, so that the additional pretraining cost and computational resources introduced by Physical Conditioning are kept consistent across methods.

  • We also list the results below for convenience. As shown, PhysGNO achieves better performance than other trajectory-generation models under the same Physical Conditioning setting, which further demonstrates the effectiveness of the proposed PhysGNO backbone itself.

BackboneSettingAspirinEthanolToluene
EGNOw/o PC9.644.5710.99
EGNO+ PC8.274.856.83
PAINETw/o PC8.844.5510.00
PAINET+ PC8.464.286.43
PhysGNOw/o PC6.333.425.05
PhysGNO+ PC5.512.222.32

W1-4: Generalization of the Potential Network

  • Following your insightful comment, we further evaluate the generalization capability of PhysGNO on new molecular systems. Specifically, we leverage a large pretrained potential model, MACE-OFF23 Large [1], which is primarily trained on the SPICE dataset. Such large-scale pretrained potential models can effectively capture interatomic interactions across diverse molecules and often exhibit strong transferability.

  • We directly use the official checkpoint of this pretrained model and apply it to the MD17 dataset without additional training. The detailed results on the S2T task (×10−2\times 10^{-2}) are shown below, further demonstrating the generalization ability of pretrained potentials.

MoleculeNo PotentialMACE-OFF23 Large
Aspirin6.336.22
Ethanol3.423.16
Toluene5.054.13
Naphthalene0.430.42

[1] MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules. JACS, 2025.

W2

Discussion with advanced FlashMD, MDGen, and GGND

  • We thank the reviewer for pointing out these relevant studies. We will include FlashMD, MDGen, and GGND in the Related Work section in the revised manuscript. Different from these advanced methods, the core motivation of PhysGNO is to combine the informative physical knowledge encoded in molecular potential models with the efficiency of neural operators, thereby enabling effective and efficient molecular trajectory prediction.

  • Specifically, FlashMD focuses on long-stride molecular dynamics prediction, incorporating Hamiltonian-inspired modeling principles and evaluating thermodynamic ensembles and long-term dynamical properties. MDGen studies generative modeling of molecular trajectory distributions, supporting forward simulation, transition path sampling, and trajectory upsampling. GGND proposes a plug-and-play module to alleviate error accumulation and improve the stability of molecular dynamics simulation. In contrast, PhysGNO focuses on incorporating physical knowledge from pretrained potential models into finite-time neural trajectory operators.

W3 & W4

W3: Clarification of F-MSE and A-MSE

  • First, we would like to clarify that the primary task of PhysGNO is finite-time molecular trajectory prediction from an observed initial molecular state. Therefore, we adopt F-MSE and A-MSE to directly evaluate the accuracy of predicted future molecular configurations. These metrics align directly with the learning objective of finite-time neural operators.

  • Moreover, F-MSE and A-MSE are established evaluation metrics used in widely adopted trajectory prediction methods, including EGNO, PaiNN, ATOM, SGDD, and GF-NODE.

  • Nevertheless, we agree that low coordinate prediction errors alone cannot fully characterize the physical fidelity of molecular dynamics. Therefore, in Appendix C.5, we additionally provide bond-length, bond-angle, dihedral-angle, and related structural metrics to further demonstrate the physical plausibility of the trajectories generated by our method.

W4-1: Evaluation of Markov State and Transition Behavior

  • We also thank the reviewer for the valuable suggestion regarding Markov states and transition behavior. Inspired by MDGen and established Markov State Model (MSM) frameworks [2] [3], we additionally adopt Transition Total Variation (Transition TV) to measure the discrepancy between the predicted and ground-truth transition probability distributions conditioned on the initial conformational state, as well as Joint-Pair Jensen-Shannon Divergence (Joint-Pair JSD) to compare the joint distribution of state pairs between the predicted and real trajectories. This metric captures both the frequency of different conformational states and how often transitions occur between them. The detailed experimental results are shown below and further support the effectiveness of our method.
MoleculeMethodTransition TV ↓Joint-Pair JSD ↓
AspirinEGNO0.150.06
AspirinPhysGNO0.120.03
EthanolEGNO0.180.08
EthanolPhysGNO0.160.05
MalonaldehydeEGNO0.460.41
MalonaldehydePhysGNO0.330.31

[2] PyEMMA 2: A Software Package for Estimation, Validation, and Analysis of Markov Models. Journal of Chemical Theory and Computation, 2015.
[3] Markov Models of Molecular Kinetics: Generation and Validation. The Journal of Chemical Physics, 2011.

W4-2: Modeling with Momenta

  • Regarding modeling with momenta, we would like to clarify that PhysGNO does not ignore the directional information associated with molecular motion. As discussed in Sections 3 and 4.3, molecular dynamics contains diverse directional signals, including velocities, forces, relative displacements, and local interaction directions. Accordingly, PhysGNO adopts a latent vector representation Q\mathbf{Q} to capture these directional signals in the molecular system, including motion-related information associated with momenta and velocities.

W5

Larger Molecular Systems and Longer Timescales

  • To evaluate effectiveness over longer timescales, in Appendix C.4 we further consider longer prediction horizons and a larger number of predicted states, in order to better demonstrate the performance of our method in longer-horizon trajectory prediction. These results further support the advantage of our method on longer-timescale dynamics.

  • Following the reviewer’s suggestion, we further evaluate PhysGNO on a larger biomolecular system consisting of 916 atoms, derived from the human FOS protein (PDB: 1A02, Chain F). Specifically, we extract a 53-residue protein fragment with ACE/NME terminal capping groups, comprising 447 heavy atoms and 469 hydrogen atoms. We construct the reference molecular dynamics trajectory using classical molecular mechanics simulations with the AMBER14 force field and the GBn2 implicit-solvent model, implemented in OpenMM under a 300 K Langevin NVT ensemble. Based on this trajectory, we evaluate PhysGNO and EGNO over a 30 ps sparse prediction horizon with 64 queried future configurations, providing an extended evaluation beyond conventional small-molecule benchmarks.

  • The detailed results are presented below. PhysGNO preserves the computational efficiency characteristic of neural operators, similar to EGNO, while achieving better performance. These results further demonstrate the effectiveness of our method on large molecular systems.

916-atom ExperimentA-MSETime (ms)
PaiNN rollout0.441458
EGNO0.49235
PhysGNO w/o PC0.42285
PhysGNO0.38315

Q1 & Q2

Additional cost and generalization of introducing pretrained potentials

Please refer to our response to W1, where we discuss the inference cost and generalization issues introduced by pretrained potentials.

Q3

We thank the reviewer for the suggestion. We will expand the Related Work section to discuss FlashMD, MDGen, and GGND and clarify their differences from PhysGNO. Please refer to W2.

Q4 & Q5

Please refer to W3 & W4 for our discussion of F-MSE/A-MSE, molecular geometry metrics, Markov states, transition behavior, and momenta.

Q6

Please refer to W5 for our response regarding longer prediction horizons, larger molecular systems, and computational efficiency.


We hope our replies address your concerns. If you have any further questions, please do not hesitate to let us know.

Reviewer_867dOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PhysGNO, a neural operator for molecular dynamics that uses forces from pretrained molecular potentials to guide trajectory prediction. The authors also introduce equivariant spatial layers and flexibility-aware temporal attention to better capture molecular dynamics. They evaluate the method on MD17, rMD17 and MD22, demonstrating strong improvements in trajectory prediction accuracy over existing methods.

Critical strengths

Although the methodological contribution is somewhat incremental, I still see this as a solid contribution to the growing field of MD neural operators. In particular, leveraging pretrained molecular potentials is a practical development that should improve reproducibility and make the method easier to apply to real-world molecular systems.

Critical weaknesses

The authors themselves acknowledge that rMD17 provides more reliable trajectories than the original MD17. Given the known issues with MD17, I would prefer to see rMD17 used as the primary benchmark rather than relegated to the appendix. The rMD17 evaluation also excludes several important baselines, including ATOM, making the results less convincing than those presented for MD17.

Although the physical conditioning is motivated by Newtonian dynamics, it ultimately amounts to incorporating pretrained forces through learned gates. Unlike Lagrangian neural networks, which derive their dynamics from the Euler–Lagrange equations, PhysGNO does not enforce any underlying equations of motion or conservation laws. The theoretical analysis only establishes equivariance. I would like to see more discussion of what physical consistency the proposed conditioning actually provides beyond simply incorporating force information.

Questions for authors

Could the authors provide a more comprehensive evaluation on rMD17, including the stronger baselines used in the main experiments?

Since PhysGNO benefits from additional energy and force supervision through its pretrained potentials, how much of its improvement over existing neural operators comes from this additional information rather than the proposed architecture?

It would be interesting to see PhysGNO trained on the much larger TG80 dataset introduced in ATOM, to better assess its generalisation against existing methods. However, I appreciate that training at this scale may be too computationally expensive for the rebuttal period.

Given that the method explicitly incorporates pretrained forces, could the authors evaluate physical fidelity beyond geometric errors, particularly energy conservation and long-term dynamical stability?

Could the authors ablate SE(3) equivariance? There is mixed evidence in the literature regarding its benefits, particularly when comparing equivariant methods such as EGNO with non-equivariant approaches such as ATOM. It would be interesting to see how much equivariance contributes to PhysGNO's performance.

3 comments

Dear Reviewer 867d,

Thank you for your insightful comments and valuable feedback.

W1

More Comprehensive Evaluation on rMD17.

  • Following your insightful comment, we will make the rMD17 results more prominent in the revised manuscript.
  • We also provide additional results for the S2T task (×10−2\times 10^{-2}) on rMD17, including advanced baselines such as ATOM and PAINET, as shown below. To ensure a fair comparison, we use three different random seeds and adopt the official codebases released by ATOM and PAINET. The proposed PhysGNO achieves competitive performance compared with these strong baselines.
MethodAspirinEthanolMalonaldehydeNaphthaleneParacetamol
EGNO9.234.5712.810.3912.35
PAINET9.244.1415.830.3612.28
ATOM8.344.4312.420.3910.41
PhysGNO5.433.3812.340.388.18

W2

Physical Consistency of Physical Conditioning

  • We would like to clarify that PhysGNO is designed to incorporate physical priors into efficient finite-time trajectory prediction, rather than to model the dynamics through an explicit Lagrangian formulation as in Lagrangian Neural Networks.
  • In addition, physical conditioning goes beyond simply concatenating force features. First, the forces are derived from the negative gradient of a pretrained potential, providing a well-defined energy-force relationship. Second, the gating mechanism, inspired by Newtonian dynamics, adaptively incorporates this physical information into future-state prediction. Moreover, the conditioning mechanism preserves the rotational equivariance of the force information, ensuring geometric consistency across coordinate systems.
  • Therefore, PhysGNO provides physical guidance through energy-derived forces and geometric symmetry, rather than explicitly enforcing conservation laws.
  • We provide additional experimental results in our response to Q4 to further demonstrate energy conservation and long-term dynamical stability after incorporating pretrained force information.

Q1

More Comprehensive rMD17 Evaluation.

We thank the reviewer for this valuable suggestion. Please also refer to our response to W1. We will move the rMD17 experiments to the main text to provide a clearer comparison among different methods.

Q2

More Clarification of Architectural Contribution.

  • To better demonstrate the effectiveness of the PhysGNO architecture itself, in Appendix C.3 (Table 10) we apply the same physical conditioning mechanism (PC) to different backbones, such as EGNO and PAINET.
  • As shown in Table 10, PhysGNO still achieves better prediction accuracy on Aspirin, Ethanol, and Toluene. These results not only demonstrate the effectiveness of physical conditioning, but also verify the effectiveness of the PhysGNO backbone itself.
BackboneSettingAspirinEthanolToluene
EGNOw/o PC9.644.5710.99
EGNO+ PC8.274.856.83
PAINETw/o PC8.844.5510.00
PAINET+ PC8.464.286.43
PhysGNOw/o PC6.333.425.05
PhysGNO+ PC5.512.222.32

Q3

Pretraining on the TG80 Dataset.

  • We thank the reviewer for this valuable suggestion. We agree that TG80 is a valuable benchmark for evaluating cross-molecule generalization. To further investigate broader generalization, we use TG80 as the pretraining dataset for PhysGNO and follow the same evaluation protocol and cluster assignment as ATOM to assess the generalization capability of our method.
  • As shown in the table below, PhysGNO exhibits competitive performance compared with the advanced ATOM baseline, further demonstrating the generalization capability and effectiveness of our method.

Multitask S2T MSE (×10−2\times 10^{-2}) on TG80.

TG80 OODPhysGNOATOMEGNO
Cluster 131.5135.0545.95
Cluster 295.10106.99115.43
Cluster 356.5160.95151.74
Cluster 470.4266.68163.90
Cluster 543.1147.49113.68

Q4

Energy Conservation and Structural Stability.

  • Beyond coordinate errors, we provide additional physical and structural metrics in Appendix C.5, including bond length, bond angle, dihedral angle, and dRMSD, to evaluate the geometric plausibility of predicted molecular conformations. For example, on Aspirin, PhysGNO achieves a bond-length MAE of 0.069 Å, compared with 0.089 Å for EGNO. PhysGNO also achieves a dRMSD of 0.198 Å, compared with 0.241 Å for EGNO. These results suggest that PhysGNO better preserves molecular structural integrity over the evaluated prediction horizons.
  • Following the reviewer’s insightful suggestion, we further evaluate the energy consistency of trajectories predicted by PhysGNO. Specifically, we use a shared frozen MACE potential to evaluate the potential energy of predicted configurations and calculate the total energy by combining potential and kinetic energies. We then measure the maximum total-energy drift relative to the initial state over a 15 ps prediction horizon. The results show that PhysGNO maintains relatively small energy drift, with 0.120 meV/atom for Benzene and 0.330 meV/atom for Ethanol.
MoleculeEnergy Drift (meV/atom) ↓
Benzene0.120 ± 0.024
Ethanol0.330 ± 0.053

Q5

Contribution of SE(3) Equivariance.

  • Following your insightful comment, to further isolate the contribution of equivariance, we additionally consider several non-equivariant spatial baselines, including a standard MPNN without equivariant constraints and GraphSAGE, while keeping the physical conditioning (PC), flexible temporal evolution (FTE), and all other training settings unchanged. We also include standard EGNO as an additional equivariant baseline for comparison.
  • The results of the S2T task (×10−2\times 10^{-2}) on the MD17 datasets are shown below. From the table, we observe that EGNO achieves certain improvements over the non-equivariant MPNN and GraphSAGE baselines, suggesting that equivariance provides some benefit.
  • As shown in the table, the benefits of equivariance vary across molecular systems. EGNO outperforms both non-equivariant baselines on Naphthalene, while the differences are less consistent on Aspirin and Malonaldehyde. These results suggest that equivariance provides a useful geometric inductive bias, but does not solely explain the overall performance of PhysGNO. The additional improvement is also associated with the proposed spatial interaction architecture.
MoleculePhysGNOEGNO + PC/FTEMPNN + PC/FTEGraphSAGE + PC/FTE
Aspirin5.518.279.057.84
Malonaldehyde10.4310.5610.4611.52
Naphthalene0.280.320.410.33

We hope our replies address all your concerns. If you have any further questions, please do not hesitate to let us know.

Dear Reviewer 867d,

Thank you for your time and effort in reviewing our paper. Your comments have been very valuable in improving our work. As the author–reviewer discussion period will end in about two days, we would be grateful if you could check our response and let us know whether your concerns have been addressed. Thank you again for your time and help.

Best regards,

The Authors

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

This paper uses a physics-guided graph neural operator for molecular trajectory prediction. The method improves the efficiency and informativeness of molecular dynamics simulation by providing molecular status at multiple time horizons, instead of through numerical integration.

PhyscGNO includes physical information from a pre-trained potential model through a conditioning mechanism. It introduces SE(3) equivalent evolution to capture molecular interactions. The flexibility is constructed using normalized variations in pairwise atomic distances estimated from the molecular configuration.

Experiments on MD17, rMD17, and MD22 demonstrate good predictive performance against existing molecular dynamics predictive models. Also, component ablations, comparisons across pre-trained potential models, and computational efficiency are provided.

Critical strengths

Good empirical performance and comprehensive experiments. The proposed method achieves substantial improvements over competitive baselines across multiple molecular systems. The experimental evaluation is well designed, covering multiple datasets, established baselines, component ablation studies, and inference-time efficiency comparisons.

Well-motivated incorporation of physical priors. I appreciate the motivation behind leveraging physical priors to improve molecular dynamics modeling. In particular, the use of invariant scalar features and equivariant vector representations is well motivated for capturing molecular interactions in three-dimensional space.

Efficient multi-horizon trajectory prediction. The idea of directly predicting molecular states at multiple future horizons is reasonable and practically useful, as it avoids repeated numerical integration steps required by conventional molecular dynamics approaches.

Critical weaknesses

In Section 4.4, the proposed flexibility prior is constructed from the empirical mean and variance of pairwise atomic distances. While this measure is intuitive, its estimation may be sensitive to temporal dependence among observations and the coverage of molecular configurations in the training set. It would be useful to investigate whether the model's performance is sensitive to how this statistic is estimated or how the training configurations are sampled.

Related to this, the ablation study only evaluates the complete flexibility-aware temporal evolution module, without isolating the contribution of the proposed flexibility statistic. It is therefore unclear how much this prior contributes to the overall performance improvement. Additional comparisons using constant or randomly permuted flexibility params, with the current temporal architecture unchanged, would help establish the effectiveness of the proposed prior.

Questions for authors

1- How sensitive are the proposed flexibility statistics to the sampling of molecular data? 2- Contribution of the flexibility prior. Can the authors provide additional ablation experiments to isolate the contribution of the flexibility statistic from the temporal attention and gating mechanism? 3- Is there empirical evidence that the proposed statistic captures meaningful molecular flexibility?

2 comments

Dear Reviewer owJf,

Thank you for your insightful comments and valuable feedback.

W1

Sensitivity to Flexibility-Statistic Estimation.

  • We agree that the estimation of flexibility statistics may depend on the number and sampling of training configurations. To further investigate the robustness of our method to flexibility estimation, we conduct additional experiments.

  • Specifically, we compare variants using different numbers of sampled training configurations for flexibility estimation: (1) Full, which uses all original training configurations; (2) 50%, which uses half of the configurations; (3) 25%, which uses one quarter of the configurations; and (4) w/o FlexOOD, which removes flexibility-conditioned OOD exploration.

  • As shown below, we report the S2T A-MSE results (×10−2\times 10^{-2}) on MD17. The results show that prediction performance remains relatively stable across different sampling ratios, while all evaluated sampling strategies outperform the w/o FlexOOD setting, further supporting the robustness of the proposed flexibility mechanism.

    MoleculeFull50%25%w/o FlexOOD
    Aspirin5.515.475.946.26
    Ethanol2.222.252.483.42
    Toluene2.322.462.374.54

W2

Effectiveness of the Proposed Flexibility Statistic.

  • In Table 3, we primarily evaluate the effectiveness of the complete flexibility-aware temporal evolution module. To further isolate the contribution of the proposed flexibility statistic, we conduct a more detailed ablation following your suggestion. Specifically, we consider the following variants: (1) replacing the flexibility statistic with a constant value; (2) replacing it with randomly generated values; and (3) the original mechanism.

  • The S2T results on MD17 are summarized below. As shown in the table, the original mechanism, based on the proposed flexibility statistic, consistently outperforms both constant and random alternatives, further demonstrating the effectiveness of the proposed flexibility statistic.

    MoleculeConstantRandomOriginal
    Aspirin6.266.385.51
    Ethanol2.754.642.22
    Malonaldehyde11.5212.4710.43
    Naphthalene0.420.390.28
    Toluene2.984.792.32

Q1

Sensitivity to Data Sampling.

We thank the reviewer for this suggestion. We provide additional experiments using different training-configuration sampling ratios. Please refer to our response to W1.

Q2

Contribution of the Flexibility Statistic.

Please refer to our response to W2, where we further compare constant, random, and original flexibility statistics.

Q3

Physical Meaning of Flexibility.

  • Our flexibility statistic characterizes local molecular structural variations by measuring normalized changes in pairwise atomic distances across training configurations. The underlying motivation is that relatively rigid molecular regions generally exhibit smaller local structural variations, whereas flexible regions may undergo larger conformational changes.

  • To further examine the physical relevance of the proposed flexibility statistic, we evaluate its correlation with atomic RMSF (root mean square fluctuation), a standard molecular dynamics metric that quantifies the extent to which each atom fluctuates around its mean position over a trajectory. The results on MD17 are summarized below. As shown in the table, the proposed flexibility statistic exhibits positive correlations with RMSF, supporting its ability to capture physically meaningful molecular flexibility.

    MetricFlexibility Statistic
    RMSF Pearson0.678
    RMSF Spearman0.716

We hope our responses address all your concerns. If you have any further questions, please do not hesitate to let us know.

Reviewer_owJfOpenReview ↗

Response

I would like to thank the authors for the detailed response and additional experiments.

Q1 and W1: The subsampling approach reduces concerns about sensitivity to the training configurations Q2 and W2: The new ablation study targets the mission isolation of the proposed flexibility statistics concern Q3: RMSF correlations provide additional support, meaning that the proposed flexibility statistics capture meaningful information about atomic fluctuations

Overall, I find that the additional results have substantially addressed my main concerns and strengthened the empirical justification of the proposed method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.