- 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
- The motivation is good: balancing the efficiency of neural operators and the physical information.
- The idea is intuitive and simple.
Critical weaknesses
- 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?
- 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.
- 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.
- 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.
- 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
-
What are the training and inference costs of the pretrained potential model, and are these included in the efficiency comparisons?
-
Does PhysGNO need a separately trained potential model for each new molecule? How well does it generalize to unseen molecular systems?
-
Why did the paper not discuss recent work such as FlashMD, MDGen, and GGND?
-
Why did the paper put "matching reference coordinates" as the main metrics in the main manuscripts?
-
Why did the paper not consider important evaluation aspects of MD simulations: evaluate in certain ensembles, evaluate the Markov states, or transition behavior?
-
How does PhysGNO’s efficiency scale to larger systems and longer simulation times?
10 comments
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 () are shown below, further demonstrating the generalization ability of pretrained potentials.
| Molecule | No Potential | MACE-OFF23 Large |
|---|---|---|
| Aspirin | 6.33 | 6.22 |
| Ethanol | 3.42 | 3.16 |
| Toluene | 5.05 | 4.13 |
| Naphthalene | 0.43 | 0.42 |
[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.
| Molecule | Method | Transition TV ↓ | Joint-Pair JSD ↓ |
|---|---|---|---|
| Aspirin | EGNO | 0.15 | 0.06 |
| Aspirin | PhysGNO | 0.12 | 0.03 |
| Ethanol | EGNO | 0.18 | 0.08 |
| Ethanol | PhysGNO | 0.16 | 0.05 |
| Malonaldehyde | EGNO | 0.46 | 0.41 |
| Malonaldehyde | PhysGNO | 0.33 | 0.31 |
[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 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 Experiment | A-MSE | Time (ms) |
|---|---|---|
| PaiNN rollout | 0.44 | 1458 |
| EGNO | 0.49 | 235 |
| PhysGNO w/o PC | 0.42 | 285 |
| PhysGNO | 0.38 | 315 |
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.
Dear Reviewer rdE5,
Thank you for your insightful comments and valuable feedback.
W1
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.
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.
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.