acceptodds
Under review as a conference paper at ICLR 2027

Learning composable time-unconditional molecular energy landscapes for zero-shot environmental transfer

Abstract

Diffusion and flow-matching models are now the standard generative approaches for molecular conformational ensembles, but their learned vector fields are non-conservative and yield no scalar potential. Energy-based formulations restore conservativity by taking forces as the gradient of a learned scalar, yet the landscapes they recover remain noise-conditional or are defined only up to an arbitrary scale. In neither case can they be added to a classical force field in physical units or composed with external environmental potentials inside a molecular dynamics engine. Classical potentials are useful precisely because environmental interactions enter additively: a solute potential can be moved between vacuum, solvent, and a protein binding pocket by summing it with a potential for the new environment. We introduce LEVI, a time-unconditional energy-based generative model that is trained by both score matching and an energy-regression term that aligns it to the physical energies carried by the data. We also introduce LEViMD, which composes the resulting potential with external environmental terms at every step of a molecular dynamics integrator, so that a single solute model transfers to environments it never saw during training. We validate this in two systems: a two-dimensional system where data lies on a known one-dimensional manifold, and on alanine dipeptide, where a learned potential trained on vacuum structures alone can be composed with explicit water at inference, recapitulating known conformational shifts between vacuum and water.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion, generative, energy-based models, transferable, drug discovery, molecular dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_idK7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents LEVI, a diffusion-based learned energy model calibrated using force labels from clean structures. Through this calibration, LEVI learns not only the energy landscape but also the correct magnitude, allowing it to be additively combined with other force fields at inference time in a molecular dynamics engine. The method is evaluated on a tilted double-well potential and the alanine dipeptide system, demonstrating its ability to be incorporated into MD simulations in unseen environments.

Critical strengths

None

Critical weaknesses
  • LEVI involves a relatively complex training procedure, yet its ultimate goal appears to be replacing the computation of classical force fields during MD simulations. I find the significance of this direction somewhat unclear. On the one hand, classical force fields are generally more computationally efficient than deep learning-based force fields, while the paper does not provide any comparison of the computational efficiency between MD simulations using LEVI and those using conventional force fields. On the other hand, both experiments in the main text are trained and evaluated on a single system, without demonstrating whether LEVI has the ability to generalize across different systems. Therefore, the practical utility of the proposed method should be questioned.

  • LEVI obtains a conservative force field by taking the gradient of a predicted scalar quantity. Compared with machine learning force fields (MLFFs) that directly predict an unconstrained vector field, LEVI involves a more complicated training procedure, yet the experiments do not clearly demonstrate its advantages over existing MLFFs. For example, the Introduction mentions that MLFFs may lead to simulation instability. In this case, it would be more convincing to include an MLFF baseline and run MD simulations on the same systems for comparison. Otherwise, given that MLFFs are already quite mature and have achieved near ab initio level accuracy on small systems, the novelty and necessity of the proposed method are not sufficiently clear.

Questions for authors
  1. As mentioned in the Weakness section, after undergoing a relatively complex training procedure, LEVI appears to merely replace part of the role of classical force fields during MD simulations. Since the main text does not provide a specific explanation of this point, I do not see a clear necessity for introducing LEVI in this setting. Could the authors provide additional experiments comparing LEVI-MD with conventional force-field MD in terms of computational efficiency, cross-system generalization, and related aspects, so as to better demonstrate the advantages of LEVI?
  2. I believe that the "zero-shot environment transfer" claim made in the title and throughout the main text is somewhat overstated. Due to the additive decomposition of the potential energy shown in Eq. (1), when LEVI is used to replace UsolU^{\mathrm{sol}}, the term that actually describes the interaction between the target system and the environment (e.g., water molecules), UintU^{\mathrm{int}}, does not involve the model at all. Therefore, changing the environment does not directly expose the model to out-of-distribution data. Could the authors provide further theoretical or experimental justification for referring to this setting as zero-shot environment transfer?
  3. LEVI adopts a conservative force-field formulation, which appears to be the main source of its increased training complexity. However, the experiments do not demonstrate clear advantages of LEVI over MLFF approaches that directly fit the force field using an unconstrained vector field. In particular, MLFFs do not necessarily lead to simulation instability. For example, AI2BMD [1] fits protein force fields with near-ab initio accuracy and is able to run MD simulations on the scale of hundreds of nanoseconds. It would therefore be preferable for the authors to include an appropriate MLFF baseline and perform MD simulations on the same systems for comparison, in order to justify the use of a conservative force-field formulation.
  4. As shown in Fig. 2(a), LEVI-MD appears to overestimate the populations of C7axC_{\textrm{7ax}} and αL\alpha_{\textrm{L}}, resulting in a generated distribution that is inconsistent with the true Boltzmann distribution. I believe this observation challenges the claim that LEVI can perform test-time transfer. The authors should provide further clarification on this issue.
  5. Experiments on a small system such as alanine dipeptide may not be sufficient to fully demonstrate the effectiveness of the model. Could the authors include additional experiments involving training and MD simulations on a larger and more representative system, such as Chignolin?
2 comments
AuthorsOpenReview ↗

Response to Reviewer idK7

We thank the reviewer for the review. In response, we clarified the purpose of LEVI and its relation to classical force fields and MLFFs
(including AI2BMD), added a cost comparison with classical MD, clarified the zero-shot claim, and traced the cause of the αL / C7ax overpopulation. Our point-by-point responses are listed below.

W1: LEVI involves a relatively complex training procedure, yet its ultimate goal appears to be replacing the computation of classical force fields during MD simulations. I find the significance of this direction somewhat unclear. On the one hand, classical force fields are generally more computationally efficient than deep learning-based force fields, while the paper does not provide any comparison of the computational efficiency between MD simulations using LEVI and those using conventional force fields. On the other hand, both experiments in the main text are trained and evaluated on a single system, without demonstrating whether LEVI has the ability to generalize across different systems. Therefore, the practical utility of the proposed method should be questioned.

A1: LEVI is not meant to replace classical force fields for simulating a known molecule (see Common Response 1). Its goal is to generate a molecule inside an environment it never saw in training, with the environment responding, which neither classical nor ML force fields can do. We give the cost comparison in Q1 below and in Common Response 4: classical MD is far cheaper for a known molecule, and we state this openly (Section 5). Generalization across systems is addressed in Q5 and in Common Response 4.

W2: LEVI obtains a conservative force field by taking the gradient of a predicted scalar quantity. Compared with machine learning force fields (MLFFs) that directly predict an unconstrained vector field, LEVI involves a more complicated training procedure, yet the experiments do not clearly demonstrate its advantages over existing MLFFs. For example, the Introduction mentions that MLFFs may lead to simulation instability. In this case, it would be more convincing to include an MLFF baseline and run MD simulations on the same systems for comparison. Otherwise, given that MLFFs are already quite mature and have achieved near ab initio level accuracy on small systems, the novelty and necessity of the proposed method are not sufficiently clear.

A2: We clarify two points. For simulation instability, we believe there is a misunderstanding on this point, and we would like to clarify it for the reviewer.

  • Our Introduction does not claim that MLFFs lead to unstable simulations. The instability statement refers to deep generative models (diffusion and flow matching), whose outputs "parameterize an unconstrained vector field rather than the gradient of a scalar. Deployed as surrogate thermodynamic forces, such fields are non-conservative" (Introduction). About MLFFs, the Introduction says only that they "depend on ab initio energy labels for non-equilibrium trajectories".
  • Most MLFFs do not predict an unconstrained vector field. This includes AI2BMD (see Q3). Like LEVI, they take the force as the negative gradient of a predicted energy. Any models that are trained to match physical forces would be likely to "learn" how to make these forces conservative, even if they are not strictly conservative.

We agree with the reviewer that well-trained MLFFs run stable MD. We have revised the sentence so that this distinction cannot be misread: it now reads "The outputs of these generative models, however, cannot be substituted into F = −∇U" (Section 1).

LEVI is conservative because composition requires it. The solute term must be an energy so that it can be added to U^env + U^int, and so that the energy-regression term can fix its physical scale. The training cost of the conservative form is the same as for any MLFF trained on forces (a gradient of the network output inside the loss). The extra cost specific to LEVI is one forward pass at the clean structure for the energy term (Sec. 5).

The novelty is not higher accuracy than an MLFF on a known molecule. It is a learned energy that can (i) build the molecule from randomly initialized atoms and (ii) be added in physical units to an environment it never saw. For the MLFF baseline, please see Common Response 1.

(continued..)

AuthorsOpenReview ↗

Response to Reviewer idK7 (continued)

A1: LEVI is not intended to replace a classical force field for simulating a known molecule, and it is not faster (Sections 1 and 5). Measured on our cluster, classical explicit-water MD of alanine dipeptide runs at about 870 ns/day on one GPU, against ~4.5 ns/day for LEVI-MD (Common Response 4).

LEVI's purpose is generative: assembling a molecule from scattered atoms inside an environment it never saw in training, which classical MD cannot do. To illustrate this, please see the video LEVIMD_B2AR.mp4 in the supplementary material (Common Response 1). Cross-system generalization is addressed in Common Response 4.

A2: The reviewer is right that U^int does not involve the model. However, the environment changes which conformations the model must describe, and those conformations can be outside its training data. The learned U^sol must therefore be somewhat accurate in conformations it rarely or never saw. In this sense the environment is never seen in training, and the model is evaluated out of distribution. We have made this precise in the revision (Section 1) and use "zero-shot" only in that sense.

A3: We agree that AI2BMD shows MLFFs can run long, stable protein MD. AI2BMD in fact uses both of the design choices the reviewer questions:

  • It is conservative. In its Methods, "ViSNet was trained as an energy-conserving potential model; that is, the predicted atomic forces were derived from the negative gradients of the potential energy". This is the same formulation as LEVI.
  • It composes a learned solute potential with a classical environment. The protein is described by the ML potential and the explicit solvent by the polarizable AMOEBA force field. This is the additive composition LEVI-MD uses.

AI2BMD solves a different problem from LEVI. It simulates a given protein from given starting structures (folded, unfolded and intermediate conformations from replica-exchange MD). It cannot generate a structure. LEVI's solute term is generative: it builds the molecule inside the environment from randomly initialized atoms, and it is trained from equilibrium structures and their energies. We have added AI2BMD and ML/MM to the Related Work (Section 2, Machine-learned force fields and ML/MM). Please see also Common Response 1.

A4: We agree the overpopulation is real, and we have traced its cause: it is set while the molecule assembles from random positions in water, not by the learned energy. Please see Common Response 2. We have stated this limitation of the assembly protocol in the revision (Sections 4.3 and 5, Appendix D).

A5: We could not complete a Chignolin model in the rebuttal period. Our intended scale-up is a small molecule in a large environment, such as a ligand in a binding pocket (the β2-adrenergic receptor demonstration in Common Response 1), rather than a larger solute. We have listed larger solutes as a limitation and future work (Section 5; Common Response 4).

Reviewer_V8ykOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary: The paper introduces LEVI, a time-unconditional energy-based generative model that aligns its estimated energy to the physical energies from the data. In this way, LEVI can then be used to run molecular dynamics (LEVI-MD) in environments different from the one it was trained on since it can be additively used with traditional potentials.

Contributions:

  1. An energy-based generative model that is also aligned to the physical units of the data using an extra regression loss term. This model can be used in different environments for MD later since it allows additivity.
  2. They show that they can train the model on vacuum and then run MD on both vacuum and explicit water and recover the qualitative conformational shift (in a zero-shot manner).

Validations:

  1. Langevin System: one-dimensional toy system. LEVI recovers the Boltzmann distribution of the tilted double well when setting the weight between the two losses at or above 0.3.
  2. Alanine Dipeptide: dipeptide with known conformational changes between vacuum and water. Trained on vacuum frames only, LEVI-MD on vacuum approximately reproduces the basin populations and when moving to explicit water simulations, samples frames move towards the reference.
Critical strengths
  1. The idea is clear and addresses a well-motivated gap. The regression term in the loss fixes the scale issue of energy-based models and enables the authors to sum the learned potential with classical force fields. LEVI-MD would enable transferability of a learned potential from one environment to another.
  2. Results on alanine-dipeptide (a standard benchmark to validate energy-based models for MD) are promising and illustrate the transferability of the proposed model.
Critical weaknesses

My main concern is on the validation part. While it is clearly shown that the model presents some transferability across conditions, the agreement with the solvated reference is modest (for example, PPII 37.1% vs 58.4% in Fig. 2a) and there is no quantitative comparison to any baseline, even on the vacuum task. Also, alanine dipeptide is a small peptide, ideal for experimenting with new ideas but it is unclear whether this approach scales to longer, more complex systems. Without baselines or a harder system, the paper can’t show that the method is competitive with existing approaches or that the zero-shot claim holds beyond this simple case.

Questions for authors

Questions for the authors:

  1. What limits scaling to larger systems? As I understand, you need vacuum MD data for training the model but getting MD data for each new system is computationally expensive since the model is system-specific.
  2. Can you also do a comparison of the two scenarios: 1) generate MD for vacuum, train the model on vacuum and do inference on water vs 2) generate MD for vacuum and water, to further illustrate the time you save by having a transferable model across different conditions.
  3. Can you compare with other baselines from the area to show that they fail to transfer to new conditions (like ScoreMD that you mention in Related Work)?
1 comment
AuthorsOpenReview ↗

Response to Reviewer V8yk

We thank the reviewer for the constructive feedback. In response, we added quantitative comparisons using ScoreMD's metrics, explained why ScoreMD-type models cannot be used in the environmental transfer setting, and clarified the cost and scaling of LEVI. Our point-by-point responses are listed below.

W1: My main concern is on the validation part. While it is clearly shown that the model presents some transferability across conditions, the agreement with the solvated reference is modest (for example, PPII 37.1% vs 58.4% in Fig. 2a) and there is no quantitative comparison to any baseline, even on the vacuum task. Also, alanine dipeptide is a small peptide, ideal for experimenting with new ideas but it is unclear whether this approach scales to longer, more complex systems. Without baselines or a harder system, the paper can't show that the method is competitive with existing approaches or that the zero-shot claim holds beyond this simple case.

A1: Thanks for the feedback. We now report ScoreMD's two metrics (PMF error and JS divergence), each compared with the best score achievable and with a vacuum-only model at the same sample size (Section 4.3, Appendix E, Table 7); please see Common Response 3. The remaining gap in water is the population error analysed in Common Response 2 (Appendix D). For scaling and harder systems, please see Q1 below and Common Response 4.

Q1: What limits scaling to larger systems? As I understand, you need vacuum MD data for training the model but getting MD data for each new system is computationally expensive since the model is system-specific.

A: The model is not system-specific by design: atoms are described only by force-field attributes (partial charge, Lennard-Jones ε and σ) and the bond graph, not by element identity or a molecule index, so one network can be trained on datasets that include different molecules (such as GEOM-Drugs) and generate de novo molecules during inference. The current limitation is the cost of LEVI-MD: each run assembles one molecule, which takes ~1 h per sample for alanine dipeptide (Section 5, Appendix A.3). Please see Common Response 4 for details.

Q2: Can you also do a comparison of the two scenarios: 1) generate MD for vacuum, train the model on vacuum and do inference on water vs 2) generate MD for vacuum and water, to further illustrate the time you save by having a transferable model across different conditions.

A: For alanine dipeptide, transfer does not save time, and we state this in the revision (Section 5). Explicit-water MD of this system runs at about 870 ns/day on one GPU, so a 100 ns reference takes about 3 hours. The value of transfer is not speed for a known molecule: it is the ability to assemble a molecule from scratch inside an environment, such as a ligand being generated in a binding pocket (see Common Response 1).

Q3: Can you compare with other baselines from the area to show that they fail to transfer to new conditions (like ScoreMD that you mention in Related Work)?

A: Composition with an environment requires the learned energy to be in physical units, so that it can be summed with U^env + U^int. Energy-based diffusion models trained by score matching alone, such as ScoreMD and Two-for-One, determine the energy only up to an arbitrary scale. ScoreMD also keeps the diffusion time as an input, so it learns noise-conditional energy surfaces rather than one static energy (Section 2).

To the best of our knowledge, none of these models, including ScoreMD, can be run in the environmental transfer setting. For example, ScoreMD is trained on samples of the target ensemble (in their paper, alanine dipeptide in implicit solvent) and has no term through which an environment can enter at inference. Trained on our vacuum data, it could at best reproduce the vacuum ensemble. That is the vacuum-only model in Common Response 3 (Table 7): in water, PMF error 23.8 and JS 0.41, against 7.02 and 0.189 for LEVI-MD.

Reviewer_2a8WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper learns an energy function for a small molecule from vacuum simulations. It uses score matching plus an extra term that ties the output, through a learned scale and offset, to the CHARMM36m energies of the training frames, so the result is in physical units. At simulation time this learned energy is summed with a standard water model and classical solute–water interaction terms, so the molecule can be simulated in water without any water training data. The approach is tested on 1D/2D toy problems and on alanine dipeptide. There, success is measured by how 1,000 independent simulations distribute over the backbone (φ,ψ) basins, compared with 400 ns of explicit-water reference MD.

Critical strengths

The idea is simple and sensible: fix the missing energy scale of score-matched models with an energy-regression term. Training uses vacuum data only, and the water MD is used only for evaluation, which is a good zero-shot setup. The model gets the direction of the solvent effect right. In the model, PPII rises from 6% to 37% and C5 falls from 65% to 16%. In the reference MD, PPII goes from 10% to 58% and C5 from 63% to 14% (Fig. 2a).

Critical weaknesses
  1. The remaining errors in water are large, and the authors' explanation for them is untested. PPII reaches 37% against 58% in the reference. Two rarely visited states (αL and C7ax) are over-populated: about 14% combined, against about 2%. With about 1.5 points of sampling error from 1,000 runs, these gaps are far outside noise. The authors blame the compressed energy scale (slope 0.668) and state that the composition framework is not what limits the results (Sec. 5). Neither is tested. PPII is already about 40% too low in vacuum (6% vs 10%), so part of the water gap may be the same model error. The large over-population of αL and C7ax appears only in water (in vacuum C7ax is 0.9% vs 0.3%). Each water structure is also a single frame taken 20 ps after the learned energy reaches full strength, after a ramp from random positions. That ramp is shared with vacuum, but in water it adds a connectivity restraint and, during the ramp only, a solute–solute repulsion. The paper doesn't show that 20 ps is enough to equilibrate in water. Two easy checks would separate these effects: the same protocol with the true force field, and with the energy scale corrected.
  2. The evidence is one small molecule in one solvent, with no baseline. The abstract talks about transfer to "new environments", but only one is tested. The method needs energy labels anyway, so the natural comparison is a standard ML force field trained on the same energies and added to water the same way. That comparison is missing. The "first framework" claim is scoped to learned generative potentials, which is fair. But the paper never cites ML/MM work, where a learned intramolecular potential is already summed with classical environment terms, and it should position itself against that.

Overall: The two controls in Q1, plus either an ML force-field baseline composed the same way or a second solute, could move me to 3.

Questions for authors
  1. What does the same water protocol give with the true CHARMM36m solute energy, and with the energy scale rescaled? These are the results that would most change my assessment.
  2. Are the water results stable with a longer final simulation stage?
  3. Were the restraint and schedule settings, or the checkpoint, chosen using the water results?
  4. What is the energy error (RMSE) on held-out vacuum frames?
  5. Can you add a second molecule or a second solvent?
  6. Minor: Table 1's γ_E list omits values used in Fig. 1 (e.g. 0.3, 0.4), while the text says 0.01–10; please state the full grid. Line 327 refers to Fig. 1c where Fig. 1d is meant.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 2a8W

We thank the reviewer for the careful and constructive review. In response, we added the rescaled-energy control, a direct test that starts LEVI-MD from equilibrated structures, a longer final stage, the held-out energy error and the full γ_E grid, and we clarified LEVI's purpose and its relation to ML force fields and ML/MM. Our point-by-point responses are listed below.

W1: The remaining errors in water are large, and the authors' explanation for them is untested. PPII reaches 37% against 58% in the reference. Two rarely visited states (αL and C7ax) are over-populated: about 14% combined, against about 2%. With about 1.5 points of sampling error from 1,000 runs, these gaps are far outside noise. The authors blame the compressed energy scale (slope 0.668) and state that the composition framework is not what limits the results (Sec. 5). Neither is tested. PPII is already about 40% too low in vacuum (6% vs 10%), so part of the water gap may be the same model error. The large over-population of αL and C7ax appears only in water (in vacuum C7ax is 0.9% vs 0.3%). Each water structure is also a single frame taken 20 ps after the learned energy reaches full strength, after a ramp from random positions. That ramp is shared with vacuum, but in water it adds a connectivity restraint and, during the ramp only, a solute–solute repulsion. The paper doesn't show that 20 ps is enough to equilibrate in water. Two easy checks would separate these effects: the same protocol with the true force field, and with the energy scale corrected.

A1: We thank the reviewer for pushing us to test this. We have now separated the effects, and the explanation in Sec. 5 was incomplete; we have revised it (Sections 4.3 and 5) and added the new controls in Appendix D. We obtained two additional insights: (i) the energy scale is not the cause, since rescaling the learned energy by 1/0.668 worsens agreement in water; and (ii) the αL / C7ax excess is set while the molecule assembles in water, not by the learned energy. A direct test confirms this: LEVI-MD started from equilibrated structures keeps φ > 0 at the reference level (4.0% vs 3.8%). Please see Common Response 2 for the full analysis. The longer final stage is addressed in Q2, and the true force field in Q1.

W2: The evidence is one small molecule in one solvent, with no baseline. The abstract talks about transfer to "new environments", but only one is tested. The method needs energy labels anyway, so the natural comparison is a standard ML force field trained on the same energies and added to water the same way. That comparison is missing. The "first framework" claim is scoped to learned generative potentials, which is fair. But the paper never cites ML/MM work, where a learned intramolecular potential is already summed with classical environment terms, and it should position itself against that. Overall: The two controls in Q1, plus either an ML force-field baseline composed the same way or a second solute, could move me to 3.

A2: We thank the reviewer and clarify the goal of this work. LEVI is a generative model. Its ultimate purpose is to generate de novo molecules inside environments it never saw during training, such as a solvent or a protein binding pocket, while the environment atoms respond to the growing molecule. Alanine dipeptide is the first validation step: it tests whether a learned generative energy can be added to a classical environment in physical units, on a system where the correct answer is known.

ML force-field baseline and ML/MM. A standard ML force field describes a molecule whose identity and bonding are already given and needs an already assembled starting structure, so composed the same way it would address a different task (conformational sampling of a known structure). If an ML/MM or MLFF model was used on a randomly initialized conformation, the simulation would likely explode immediately. The task of smoothly guiding molecules from random initial positions to accurate structures, while preserving population differences in the generated ensemble, is not addressed by these force fields. We agree, however, that some similarities of our approach with ML/MM models bear mention, and we have added them to the Related Work (Section 2, Machine-learned force fields and ML/MM).

New environments. We provide a video (supplementary material, LEVIMD_B2AR.mp4) of LEVI-MD generating a molecule de novo inside the binding pocket of the β2-adrenergic receptor, embedded in a lipid membrane with water; please see Common Response 1. The second solute is addressed in Q5.

(continued...)

AuthorsOpenReview ↗

Response to Reviewer 2a8w (continued)

Q1: What does the same water protocol give with the true CHARMM36m solute energy, and with the energy scale rescaled? These are the results that would most change my assessment.

A: Rescaled energy. We repeated the water LEVI-MD simulations with the learned solute energy × 1.5 (≈ 1/0.668). Rescaling makes agreement worse (PPII 18% against 37%) (Appendix D); please see Common Response 2 (i) and rescaled_grid_water.pdf in the supplementary material.

True CHARMM36m solute energy. CHARMM36m is not a generative model, so it cannot replace LEVI as U^sol in this protocol. LEVI-MD was designed to work with forces from the smoothed energy surface of a generative model: the solute starts from atoms placed at random positions and is assembled by the learned force. CHARMM36m's bonded terms are stiff harmonic springs, V = K_b(b − b_0)², parameterised only for small deviations around the equilibrium length b_0 (par_all36m_prot.prm, BONDS section):

Bond in alanine dipeptideK_b (kcal/mol/Ų)b_0 (Å)
CA–CB222.51.538
CA–C, CH3–C2501.490
N–CA, N–CH33201.430
C–H (methyl / Cα)322 / 3301.111 / 1.080
peptide C–N3701.345
N–H4400.997
C=O6201.230

The resulting force is 2K_b(b − b_0). For a C=O pair starting 9 Å apart, this is about 9,700 kcal/mol/Å at full strength. Starting from random positions, these terms would make the simulation unstable.

Starting instead from an already assembled alanine dipeptide, LEVI-MD with CHARMM36m as U^sol reduces to standard explicit-solvent CHARMM36m MD. That is the reference simulation we compare against. CHARMM36m cannot replace the LEVI energy because it is not a generative model. It is designed to describe a molecule whose structure and bonding are already given, and it only models small fluctuations around that structure. It cannot generate the structure itself from atoms at random positions or identities, which is the task LEVI performs in LEVI-MD. We have added this point to Appendix D (Classical force field as the solute energy).

Q2: Are the water results stable with a longer final simulation stage?

A: Yes, at the level of the major basins. We extended the final stage of LEVI-MD water runs from 20 ps (0.02 ns) to 2 ns (Figure 7, Appendix D). Every run stays in the major basin it occupies at 20 ps, either the extended β/PPII basin (ψ ≈ 150°) or the helical α basin (ψ ≈ 0°), and none crosses into the other or to φ > 0. Within each basin, a run keeps exchanging between neighbouring cells (C5 ↔ PPII and αR ↔ C7eq) throughout the extended stage, as expected for thermal fluctuations within one basin.

Q3: Were the restraint and schedule settings, or the checkpoint, chosen using the water results?

A: None of these choices were made by comparing LEVI-MD water populations with the reference water MD.

  • Schedule. The staged ramp exists so that the molecule assembles well, not to tune populations. At first the LEVI force is zero, so the solvent alone moves the particles and sets their initial arrangement. As the LEVI contribution grows, it gradually takes over and brings the particles into a well-assembled molecule
  • Restraints. The connectivity and solute intermolecular forces used in water only help the solute assemble against the solvent. The connectivity restraint keeps the solute particles from dispersing before they assemble; it exerts no force once the molecule is a single connected group (gaps ≤ 2.5 Å). The solute–solute nonbonded term is scaled by 1 − λ, so it vanishes at full coupling. Every reported frame comes from the λ = 1 stage and is a correctly bonded molecule (all frames pass the structure filter), so neither term acts on the sampled structures.
  • Checkpoint. It was chosen from the vacuum training run alone, after the vacuum validation loss had converged.

Q4: What is the energy error (RMSE) on held-out vacuum frames?

A: On the 50,000 held-out vacuum frames (a replica never used in training), the RMSE between the LEVI energy and the CHARMM36m energy is 5.25 kJ/mol per molecule (Section 4.3).

Q5: Can you add a second molecule or a second solvent?

A: We could not train and validate a second solute within the rebuttal period. The β2-adrenergic receptor demonstration adds a second and much more complex environment: protein, lipid membrane and water. It can only be qualitative, because no reference distribution exists for de novo generation in a pocket. Please see Common Responses 1 and 4.

Q6: Minor: Table 1's γ_E list omits values used in Fig. 1 (e.g. 0.3, 0.4), while the text says 0.01–10; please state the full grid. Line 327 refers to Fig. 1c where Fig. 1d is meant.

A: Thank you for your careful reading of the paper. Both have been corrected (Appendix A.1, Table 1; Section 4.1). The full grid was γ_E ∈ {0, 0.01, 0.02, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 1, 2, 5, 10} (1D system) and {0.02, 0.05, 0.1, 0.2, 0.5, 1, 2} (2D system).

Comments on the paper

5 comments
AuthorsOpenReview ↗

Author response

Response to all reviewers

We sincerely thank the reviewers for their constructive feedback and the AC for coordinating the process. Their feedback has helped us test our explanation of the water results, add quantitative comparisons, and state the purpose of LEVI more clearly. We have uploaded a revised manuscript; new and changed text is marked in blue.

We are encouraged that the reviews recognize strong points of the manuscript, such as:

  • A simple, well-motivated idea. An energy-regression term fixes the missing energy scale of score-matched models, so the learned potential can be summed with classical force fields (Reviewers 2a8W and V8yk).
  • A clean zero-shot setup. Training uses vacuum data only, and the water MD is used only for evaluation (Reviewer 2a8W).
  • The direction of the solvent effect on alanine dipeptide is reproduced: PPII rises and C5 falls (Reviewers 2a8W and V8yk).

We have categorized the reviewers' questions into four main areas. Our major improvements include:

  1. Purpose of LEVI and its relation to force fields, MLFFs and ML/MM (Raised by: idK7, 2a8W). To address questions about what LEVI is for, we: clarify how it differs from classical force fields, MLFFs (including AI2BMD) and ML/MM (Section 1); add the ML/MM literature to the Related Work (Section 2); and provide a video of LEVI-MD generating a ligand de novo in a protein binding pocket (supplementary material, LEVIMD_B2AR.mp4). (Common Response 1)

  2. New controls on the remaining errors in water (Raised by: 2a8W, idK7, V8yk). To address concerns about the PPII deficit a
    nd the αL / C7ax excess in water, we ran LEVI-MD with the rescaled energy, ran LEVI-MD started from equilibrated reference structures, a
    nd analysed at which coupling stage each run settles on its φ side (Appendix D, Figure 6 and Table 6). We have revised Sec
    tions 4.3 and 5
    accordingly. (Common Response 2)

  3. Quantitative comparison with baselines (Raised by: V8yk, 2a8W, idK7). To address the lack of quantitative comparisons, we report ScoreMD's metrics (PMF error and JS divergence) together with the best achievable score and a vacuum-only model at the same sample size (Section 4.3, Appendix E and Table 7), and discuss why ScoreMD-type models cannot be used in the environmental transfer setting (Section 2). (Common Response 3)

  4. Cost, scaling and systems beyond alanine dipeptide (Raised by: V8yk, idK7, 2a8W). We report the cost of LEVI-MD against classical MD, explain why the model is not system-specific by design, and state the current limitations and future work in the revision (Section 5). (Common Response 4)

  5. Others:

  • Energy error (RMSE) on held-out vacuum frames (Section 4.3) (Reviewer 2a8W, Q4).
  • Longer final simulation stage (Appendix D, Figure 7) (Reviewer 2a8W, Q2).
  • Full γ_E grid (Appendix A.1, Table 1) and corrected figure reference (Section 4.1) (Reviewer 2a8W, Q6).
  • Clarification of the instability statement (Section 1) (Reviewer idK7, W2) and of the zero-shot claim (Section 1) (Reviewer idK7, Q2).

Below, we provide detailed common responses for these areas and individual responses to each reviewer. We welcome further discussion and would be happy to provide additional clarification or results.

AuthorsOpenReview ↗

Common Response 1: Purpose of LEVI and its relation to force fields, MLFFs and ML/MM

To address questions about the purpose of LEVI and how it relates to classical force fields, MLFFs and ML/MM (Reviewers idK7, 2a8W), we clarify the following. In the revision, we have made these distinctions explicit in the Introduction (Section 1) and in a new Related Work paragraph, Machine-learned force fields and ML/MM (Section 2), and added citations to the relevant ML/MM literature.

  • LEVI is a generative model. It learns one time-independent energy whose gradient builds the molecule from atoms initialized at random positions. Simple extensions to the model also calculate gradients for atomic attributes, which allows for de novo molecular generation. Because the learned energy is in physical units, it can be added to a classical potential for an environment never seen in training, and the environment atoms respond while the molecule forms. Its ultimate purpose is to generate de novo molecules inside environments it never saw during training, such as a solvent or a protein binding pocket. Alanine dipeptide is the first validation step: it tests whether a learned generative energy can be added to a classical environment in physical units, on a system where the correct answer is known.

  • Classical force fields and MLFFs, including ML/MM schemes, describe a molecule that is already built and whose topology is fixed. A standard MM force field, like CHARMM36m itself, describes a molecule whose identity and bonding are already given. It needs an already assembled starting structure and transitions between conformational states and orientations is generally slow. These force fields cannot build a molecule from atoms at random positions (or atomic identity), as LEVI-MD does. Composed the same way, an ML force field would therefore address a different task (conformational sampling of a known structure). AI2BMD and ML/MM both use the same conservative energy and additive composition as LEVI-MD, but they simulate a given structure rather than generating one from noisy initial conditions.

    • AI2BMD [1]. This is an AI-driven molecular dynamics framework that simulates full-atom large proteins with ab initio accuracy. The application is mainly protein folding and unfolding and explores conformational spaces. Forces are the negative gradient of the learned energy, and the solvent is the classical AMOEBA force field. However, the simulations start from given protein structures.
    • ML/MM [2, 3]. These models describe the ligand with an ML potential and the protein and water with a classical force field; the two parts interact through Coulomb and Lennard-Jones terms. Simulations start from crystallographic protein–ligand complexes.
  • To show the intended use, we provide a video of LEVI-MD generating a molecule de novo inside the binding pocket of the β2-adrenergic receptor (B2AR), embedded in a lipid membrane with water (supplementary material, LEVIMD_B2AR.mp4). The model was trained on the vacuum-optimized GEOM-Drugs dataset, and the Boltzmann weight attached to each conformer of each molecule was used to calculate its relative energy label. During inference, the ligand starts as randomly initialized particles in terms of both identities and coordinates, and its atom positions and attributes are generated together while the pocket residues and waters remain dynamic. This environment, a protein, membrane and solvent, was never seen in training. We present it as a qualitative demonstration; improving metrics such as chemical validity and drug-likeness of de novo samples is ongoing work.

[1] T. Wang, X. He, M. Li, Y. Li, R. Bi, Y. Wang, C. Cheng, X. Shen, J. Meng, H. Zhang, H. Liu, Z. Wang, S. Li, B. Shao, T.-Y. Liu. Ab initio characterization of protein molecular dynamics with AI2BMD. Nature 635, 1019–1027 (2024). https://doi.org/10.1038/s41586-024-08127-z

[2] S.-L. J. Lahey, C. N. Rowley. Simulating protein–ligand binding with neural network potentials. Chem. Sci. 11, 2362–2368 (2020). https://doi.org/10.1039/c9sc06017k

[3] R. Galvelis, A. Varela-Rial, S. Doerr, R. Fino, P. Eastman, T. E. Markland, J. D. Chodera, G. De Fabritiis. NNP/MM: Accelerating molecular dynamics simulations with machine learning potentials and molecular mechanics. arXiv:2201.08110 (2023).

AuthorsOpenReview ↗

Common Response 2: Origin of the remaining errors in water

To address concerns about the remaining errors in water (Reviewers 2a8W, idK7, V8yk), we ran new controls that separate the possible causes. The explanation in Sec. 5 was incomplete, and we have revised it (Sections 4.3 and 5), with the new controls in Appendix D. We obtain two additional insights into these residual population errors.

(i) The energy scale is not the cause. We repeated the water LEVI-MD simulations with the learned solute energy × 1.5 (≈ 1/0.668), with the protocol and checkpoint otherwise identical. Rescaling worsens agreement in water: PPII 18% against 37% (Appendix D). This indicates that the slope is not responsible for the energy differences between basins.

NPPIIC5C7eqαRαLC7axφ > 0
Water, MD reference200,00058.413.59.211.01.70.13.8
Water, LEVI-MD ×1.01,00037.115.612.413.07.86.616.9
Water, LEVI-MD ×1.51,00018.038.117.96.81.114.216.4

The full comparison is provided in the supplementary material as rescaled_grid_water.pdf: explicit water, reference MD (left), LEVI-MD with the learned energy ×1.0 (middle) and ×1.5 (right); top, (φ, ψ) density; bottom, basin populations (%).

(ii) The αL / C7ax / αD (φ > 0) excess does not come from the learned energy; it is set while the molecule assembles in water. The excess is naturally present at early stages of the coupling ramp (Figure 6, Appendix D). Before the LEVI force acts, about 26% of runs sit at φ > 0, as determined by the random particle positions. In vacuum, the ramp removes almost all of the φ > 0 structures (0.9% at λ = 1). In water, their proportion decreases, but only to 16.5%, and 95% of the runs that end at φ > 0 are already there by λ = 0.2. Crossing between φ < 0 and φ > 0 is rare even in the reference MD: 0.64% of 20 ps windows, about once per 3 ns. A 20 ps final stage therefore cannot correct this population. It is a sampling limitation of the assembly protocol in solvent, not of the composed energy.

(iii) Direct test: LEVI-MD started from equilibrated structures. We ran 100 LEVI-MD simulations in water that start from conformations of the reference CHARMM36m water MD, with the full LEVI energy for all 160 ps (λ = 1 in all stages), so there is no assembly step. Populations (%), final frame of each run (160 ps) (Table 6, Appendix D):

PPIIC5C7eqαRφ > 0 (mostly αL, C7ax, αD)
Reference MD58.413.59.211.03.8
LEVI-MD from reference structures (N = 100)57.017.011.08.04.0
LEVI-MD (random start, N = 1,000)37.115.612.413.016.9

The φ > 0 population stays at the reference level (4.0% vs 3.8%) instead of rising to 16.9%. Only 4 of 100 runs cross between φ < 0 and φ > 0. The excess is therefore set during assembly, not by the learned energy. The results are also stable with a longer final stage (Appendix D, Figure 7; Reviewer 2a8W, Q2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.