acceptodds
Under review as a conference paper at ICLR 2027

ABLE: All-Atom Biomolecular Trajectory Generation with Active-Ligand Enrichment

Abstract

Molecular dynamics simulations provide an atomistic bridge between in silico modeling and experiments, but their high computational cost limits applications in high-throughput drug discovery. Recent generative models accelerate molecular dynamics by modeling large-lag transitions, yet models trained primarily on active protein–ligand complexes may inherit inductive biases that weaken active–decoy discrimination, limiting their ability to serve as general substitutes for physics-based simulations. To address this limitation, we introduce ABLE, an all-atom biomolecular trajectory generation framework designed to extend trajectory modeling toward virtual screening. At the post-training stage, ABLE incorporates an on-policy reinforcement learning procedure centered on a newly defined protein–ligand interfacial expansion rate. By quantifying the instantaneous entropy growth of a local interfacial distribution, this quantity provides the optimization signal that drives distinct dynamical behaviors for active ligands and decoys using only static cross-docked complex structures. Experiments demonstrate that ABLE achieves competitive trajectory generation performance on both proteins and protein–ligand complexes and, after reinforcement learning, delivers state-of-the-art early enrichment in virtual screening among the evaluated baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
biomolecular trajectory generation, molecular dynamics, protein--ligand complexes, virtual screening, active-ligand enrichment, reinforcement learning, generative modeling

Reviews

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

This paper proposes ABLE, an all-atom biomolecular trajectory generation framework for virtual screening, aiming to address the limited ability of existing models to distinguish active ligands from decoys due to their reliance on MD trajectories of active protein–ligand complexes for training. ABLE learns molecular dynamics through a variational bridge matching framework and introduces the Interfacial Expansion Rate as a reward signal for on-policy reinforcement learning (RL). Experiments on the ATLAS, MISATO, and DUDE-Z datasets evaluate the model’s trajectory generation and virtual screening capabilities. Notably, RL improves the EF 1% on DUDE-Z from 0.524 to 2.807, demonstrating the method’s effectiveness in enhancing early enrichment of active ligands.

Critical strengths
  1. This paper extends generative molecular dynamics beyond conventional trajectory modeling to virtual screening based on dynamical behavior. It addresses the limitation that existing training datasets predominantly consist of active protein–ligand complexes and lack decoy trajectories, presenting a clear research objective and a corresponding solution.

  2. This paper introduces the Interfacial Expansion Rate, defined based on the entropy growth of local tracer particles, and leverages the divergence of the velocity field to construct an optimizable reinforcement learning signal. The proposed method enables post-training using only static cross-docked complexes, without requiring additional MD trajectories of decoy complexes, demonstrating a degree of novelty in both methodological design and data requirements.

  3. This paper conducts experiments on the ATLAS, MISATO, and DUDE-Z datasets to evaluate protein trajectory generation, protein–ligand trajectory generation, and virtual screening performance, respectively. Notably, reinforcement learning improves the EF 1% on DUDE-Z from 0.524 to 2.807, providing direct empirical evidence supporting the effectiveness of the proposed post-training approach.

Critical weaknesses
  1. The proposed Interfacial Expansion Rate is central to the reinforcement learning approach, but its effectiveness is primarily validated indirectly through downstream virtual screening performance. The paper lacks direct quantitative evidence establishing its relationship with actual ligand dissociation dynamics. Therefore, the reliability of this metric as a physically meaningful descriptor of molecular dynamics remains insufficiently validated.

  2. Although the paper evaluates trajectory generation quality before reinforcement learning, it does not provide an equally systematic assessment of the physical plausibility and dynamical fidelity of the generated trajectories after RL fine-tuning. Given that the authors have observed abnormally large ligand RMSD values for certain targets, further analysis is needed to determine whether reinforcement learning introduces non-physical motions and whether KL regularization effectively constrains such deviations.

Questions for authors

None

4 comments
AuthorsOpenReview ↗

Rebuttal 1

We are grateful to the reviewer for the insightful comments and helpful suggestions. In the following, we respond to each point individually and clarify the corresponding revisions.

  • W1: The proposed Interfacial Expansion Rate is central to the reinforcement learning approach, but its effectiveness is primarily validated indirectly through downstream virtual screening performance. The paper lacks direct quantitative evidence establishing its relationship with actual ligand dissociation dynamics. Therefore, the reliability of this metric as a physically meaningful descriptor of molecular dynamics remains insufficiently validated.

Thank you for raising this important concern. We agree that the proposed interfacial expansion rate does not directly represent real ligand dissociation kinetics. Its design is motivated by the empirical observation that active ligands generally maintain more stable binding poses and protein-ligand contacts during MD simulations, whereas decoys tend to exhibit less stable interactions [1]. Accordingly, we measure the instantaneous entropy growth of an auxiliary tracer distribution near the protein-ligand interface and use it as a physically motivated surrogate reward to encourage interfacial expansion for decoys while preserving the reference dynamics for active ligands.

Importantly, our primary objective is not to faithfully reproduce ligand dissociation pathways, which would likely require additional MD trajectories of protein-decoy complexes for training. Instead, we aim to introduce active-decoy discrimination into trajectory generation models using only readily available static cross-docked structures. Beyond the downstream screening metrics, Appendix D.2 (Figure 5) also provides qualitative evidence that RL finetuning induces the intended differences in generated trajectories: active ligands generally exhibit lower RMSD and retain more initial protein-ligand contacts than decoys. These findings support the use of the proposed interfacial expansion rate as an empirically effective optimization signal for trajectory-based virtual screening, without requiring faithful reproduction of real ligand dissociation dynamics.

  • W2: Although the paper evaluates trajectory generation quality before reinforcement learning, it does not provide an equally systematic assessment of the physical plausibility and dynamical fidelity of the generated trajectories after RL fine-tuning. Given that the authors have observed abnormally large ligand RMSD values for certain targets, further analysis is needed to determine whether reinforcement learning introduces non-physical motions and whether KL regularization effectively constrains such deviations.

Thank you for this important suggestion. We agree that reinforcement learning may introduce non-physical motions, as improved active-decoy discrimination does not necessarily guarantee the physical fidelity of generated trajectories. To address this concern, we have added a systematic physical fidelity assessment in Appendix D.4 (Table 8), covering ligand geometric validity, intra- and intermolecular steric clashes, and protein Cα validity on the DUDE-Z test set. For convenience, we also provide the results in Table A below. To avoid confounding effects from pre-existing structural violations, we restrict the evaluation to trajectories whose initial structures pass all physical plausibility checks.

The results reveal the importance of path-space KL regularization in preserving physical fidelity after RL. Without KL regularization ($\beta=0), RL substantially degrades ligand bond-length, bond-angle, and protein validity, while increasing intramolecular steric clashes from 0.072 in the reference model to 0.168. Introducing KL regularization progressively mitigates these distortions as β\beta increases from 0 to 0.4. Specifically, protein validity improves from 0.812 to 0.991, while intramolecular clashes decrease from 0.168 to 0.131. Moreover, At β=0.4\beta=0.4, most physical validity metrics reach levels comparable to or better than those of the reference model, providing practical guidance for hyperparameter selection. Overall, these findings demonstrate that appropriate path-space KL regularization can effectively mitigate RL-induced structural distortions, allowing ABLE to substantially improve virtual screening performance while largely preserving the physical plausibility of generated trajectories.

(to be continued)

AuthorsOpenReview ↗

Rebuttal 2

Table A: Physical fidelity of generated trajectories on the DUDE-Z test set.

ModelsLigand Validity: Bond Length (↑)Ligand Validity: Bond Angle (↑)Ligand Validity: Aromatic (↑)Ligand Validity: Double Bond (↑)Steric Clashes: intra (↓)Steric Clashes: inter (↓)Protein Validity (↑)
NeuralMD-SDE1.000/0.0001.000/0.0001.000/0.0001.000/0.0000.000/0.0000.000/0.000-
PVB0.978/0.0100.991/0.0040.957/0.0280.980/0.0220.026/0.0100.000/0.0000.724/0.154
ABLE (ref)0.926/0.0420.949/0.0260.890/0.0380.978/0.0220.072/0.0390.000/0.0000.990/0.013
ABLE (rl, β=0\beta=0)0.838/0.1250.863/0.1280.908/0.0800.981/0.0200.168/0.1310.000/0.0000.812/0.192
ABLE (rl, β=0.2\beta=0.2)0.906/0.0700.939/0.0670.949/0.0410.981/0.0200.146/0.1070.000/0.0000.979/0.056
ABLE (rl, β=0.4\beta=0.4)0.925/0.0420.960/0.0240.960/0.0150.982/0.0180.131/0.0650.000/0.0000.991/0.025

References

[1] Guterres, H., & Im, W. (2020). Improving protein-ligand docking results with high-throughput molecular dynamics simulations. Journal of chemical information and modeling, 60(4), 2189.

Reviewer_GXvQOpenReview ↗

Thanks for the Rebuttal

Thank you for the detailed rebuttal and additional experiments. Given that my initial rating was already favorable and that this topic is outside my primary area of expertise, I will maintain my rating and increase my confidence score from 1 to 2.

We sincerely thank the reviewer for their positive assessment of our work and for carefully considering our responses and additional experiments. We greatly appreciate the time and effort devoted to reviewing our manuscript, as well as the constructive feedback that has helped us improve its quality. We wish the reviewer all the best in their research and future endeavors.

Reviewer_N82KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors train a generative model for protein–ligand MD trajectories and propose using RL to fine-tune the model based on active and inactive complexes from the DUD-E set. The motivation is to improve the model’s performance for virtual screening, given that essentially all available dynamics data are for active ligands. The efficacy of the approach is supported by improved enrichment factors and AUC on ligand-dissimilar DUD-E test cases. The pretrained model also performs well compared with similar trajectory models on the distribution-matching metrics presented.

Critical strengths

The paper addresses a substantial issue in terms of adding negative data to ML models for molecular dynamics. Since it is debatable what a comparable trajectory for an inactive molecule should look like (i.e., it would likely never enter the binding site at all), it is convenient to use static data for negatives. Generally, the availability of static data is much greater than that of dynamic data, so approaches such as the one presented here, which can integrate information from both, are likely to be a necessary way forward.

Secondly, while I still have doubts that the trajectory-matching metrics used in this paper and in ours really ensure the physical plausibility or practical utility of the generated trajectories, the developed model does appear to perform well compared with the evaluated baselines.

Critical weaknesses

My biggest issue is conceptual: much of the actual implementation, the choice of an escape direction, the notation for the ligand interfacial points, and the way the RL is implemented is very idiosyncratic. That is, there is no physical basis for the behaviour of the inactive trajectories that is learned under the RL scheme. The connection between even a trajectory solved with a physics-based oracle and affinity is complex and depends critically on energetics (RBFE, ABFE), not only on geometric quantities.

We do not know that the ligand should leave via a specific vector based on the initial structure. In fact, it is likely that the inactive ligand-bound state is simply too low in probability to be sampled. While I understand that there is a pragmatic need to do something with the data, the enrichment factors and AUROC reported in Table 3 are relatively weak compared with those of orthogonal approaches to VS, such as docking and dedicated affinity models. For example, ProMol_Func, BIND, and HelixVS report enrichment factors greater than 20 on DUD-E, and docking protocols often show AUC values of 60–80%. I do not find it surprising that raw trajectory RMSD is not as efficient a classifier. It is worth noting that the cited work by Guterres and Im uses MD in order to refine docking hits and reports an AUROC of 0.683 for docking alone. The unfortunate outcome of this is that I do not find that the proposed approach is compared with adequate baselines to sufficiently motivate it as a credible alternative to existing VS methods.

Questions for authors

Virtual-screening performance needs to be compared with approaches for virtual screening, such as docking, the many static-pose ML models trained for this task, and the Boltz-2 affinity heads. At the very least, an ablation study predicting active versus inactive complexes using the static input structure needs to be performed to establish whether trajectory modelling actually adds value for the VS use case.

What is the sensitivity of the RL-tuned model to the initial pose?

In Figure 4, it appears that there are often much larger deviations in interaction RMSF for high atom indices that are not found in the MD simulations. These deviations seem to occur at the end rather than at both the beginning and the end, as would be expected for the termini. Why is this?

6 comments
AuthorsOpenReview ↗

Rebuttal 1

We appreciate the reviewer's careful evaluation and valuable feedback. We provide point-by-point responses to the questions and concerns raised below.

  • W1: There is no physical basis for the behaviour of the inactive trajectories that is learned under the RL scheme. The connection between even a trajectory solved with a physics-based oracle and affinity is complex and depends critically on energetics (RBFE, ABFE), not only on geometric quantities.

Thank you for this important comment. We agree that protein-ligand binding affinity is governed by complex thermodynamic factors, including energetic and entropic contributions, and cannot be fully characterized by geometric stability alone. Importantly, our objective is not to directly estimate binding free energies or faithfully reproduce ligand dissociation pathways, but to introduce discriminative dynamical behaviors into trajectory generation for virtual screening. This approach is motivated by previous studies showing that active ligands tend to maintain more stable binding poses and protein-ligand contacts than decoys [1], rather than by assuming a direct correspondence between geometric stability and binding affinity. Building on this observation, we propose the interfacial expansion rate as a physically motivated surrogate reward to promote ligand separation for decoys while preserving the reference dynamics for active ligands.

Although we cannot theoretically guarantee that the RL-optimized model reproduces physically accurate ligand dissociation trajectories, we can empirically assess the extent to which the model after reinforcement learning preserves physical plausibility. To this end, we have added a comprehensive evaluation in Appendix D.4 (Table 8), covering ligand geometric validity, steric clashes, and protein structural validity. For convenience, we also provide the results in Table A below. The results highlight the importance of path-space KL regularization in preserving structural plausibility during RL finetuning. Without KL regularization, RL substantially degrades ligand geometric validity and protein validity, while increasing the intramolecular clash rate from 0.072 to 0.168 compared with the reference model. Introducing KL regularization progressively mitigates these distortions, reducing the clash rate to 0.146 at β=0.2\beta=0.2 and further to 0.131 at β=0.4\beta=0.4. Moreover, at β=0.4\beta=0.4, most validity metrics reach levels comparable to or better than those of the reference model, providing a practical guidance for hyperparameter selection. Overall, these findings suggest that appropriate KL regularization enables RL to improve virtual screening performance while largely preserving the physical plausibility of generated trajectories.

Table A: Physical fidelity of generated trajectories on the DUDE-Z test set.

ModelsLigand Validity: Bond Length (↑)Ligand Validity: Bond Angle (↑)Ligand Validity: Aromatic (↑)Ligand Validity: Double Bond (↑)Steric Clashes: intra (↓)Steric Clashes: inter (↓)Protein Validity (↑)
NeuralMD-SDE1.000/0.0001.000/0.0001.000/0.0001.000/0.0000.000/0.0000.000/0.000-
PVB0.978/0.0100.991/0.0040.957/0.0280.980/0.0220.026/0.0100.000/0.0000.724/0.154
ABLE (ref)0.926/0.0420.949/0.0260.890/0.0380.978/0.0220.072/0.0390.000/0.0000.990/0.013
ABLE (rl, β=0\beta=0)0.838/0.1250.863/0.1280.908/0.0800.981/0.0200.168/0.1310.000/0.0000.812/0.192
ABLE (rl, β=0.2\beta=0.2)0.906/0.0700.939/0.0670.949/0.0410.981/0.0200.146/0.1070.000/0.0000.979/0.056
ABLE (rl, β=0.4\beta=0.4)0.925/0.0420.960/0.0240.960/0.0150.982/0.0180.131/0.0650.000/0.0000.991/0.025
  • W2 && Q1: Virtual-screening performance needs to be compared with approaches for virtual screening, such as docking, the many static-pose ML models trained for this task, and the Boltz-2 affinity heads. At the very least, an ablation study predicting active versus inactive complexes using the static input structure needs to be performed to establish whether trajectory modelling actually adds value for the VS use case.

(to be continued)

AuthorsOpenReview ↗

Rebuttal 2

Thank you for this valuable suggestion. We agree that evaluating ABLE for virtual screening requires comparisons with established methods designed specifically for this task. Accordingly, we have expanded Section 4.4 (Table 3) to include AutoDock Vina, Boltz-2, EquiScore, and CORDIAL. Please also refer to Table B below for convenience. Specifically, AutoDock Vina scores the provided poses without additional docking, while Boltz-2 and EquiScore are evaluated using their publicly released pretrained weights. For a more controlled comparison, we additionally train CORDIAL, a static structure-based virtual screening model, from scratch using the same DUDE-Z data split and active-decoy labels as ABLE, with implementation details provided in Appendix C.2.

On DUDE-Z, RL-finetuned ABLE achieves an EF@1% of 2.807 and a mean AUROC of 0.542, compared with 0.760 and 0.527, respectively, for CORDIAL. Moreover, the MISATO-finetuned ABLE reference model achieves an EF@1% of only 0.524, highlighting the substantial improvement introduced by our RL procedure. These results suggest that RL-optimized trajectory generation can extract more informative structural signals from limited cross-docked data than the evaluated static prediction baseline, particularly for early enrichment.

We acknowledge that ABLE still underperforms established virtual screening methods such as Boltz-2 and EquiScore, which may benefit from substantially larger and more diverse training datasets, with potential training-test overlap that cannot be ruled out. Nevertheless, our primary goal is to demonstrate the feasibility of enabling trajectory generation models to perform virtual screening using only static cross-docked structures, rather than directly competing with established and extensively benchmarked virtual screening methods. Achieving a more controlled comparison with specialized virtual screening methods may require further large-scale training, which is beyond the limited rebuttal period and remains an important direction for future work.

Table B: Virtual screening results on the DUDE-Z test set. ∗^* indicates models trained from scratch using the same data split.

MethodsEF@1% (↑)EF@2% (↑)EF@5% (↑)EF@10% (↑)AUROC median (↑)AUROC mean (↑)AUROC std
Virtual Screening Method
AutoDock Vina3.591/6.3184.257/6.6303.547/5.0612.548/3.0330.6030.5910.216
Boltz-218.75/11.0816.08/8.16412.30/5.8787.184/3.1030.9930.9090.144
EquiScore4.395/6.7633.953/5.3523.238/4.1582.494/2.9630.6250.6200.185
CORDIAL*0.760/2.0050.974/2.0881.195/2.0891.120/1.2450.5440.5270.107
Trajectory Generation Model
NeuralMD-SDE1.214/3.6601.422/3.2290.796/1.6720.831/1.2400.4950.4780.168
PVB0.430/2.1340.488/1.9110.862/1.6760.980/1.1830.5350.5290.126
ABLE (ref)0.524/1.9600.551/1.8190.965/1.7341.034/1.2070.5230.4990.158
ABLE (rl)2.807/4.7602.009/3.4651.754/2.1791.593/1.7440.5530.5420.185
  • Q2: What is the sensitivity of the RL-tuned model to the initial pose?

Thank you for this important question. To evaluate the sensitivity of ABLE to initial ligand poses, we have added a dedicated experiment in Appendix D.3.3 (Figure 6). Specifically, we use AutoDock Vina to generate up to five alternative initial poses for each protein-ligand complex in the DUDE-Z test set and evaluate the resulting trajectories under identical inference settings. We then assess virtual screening performance across five independent runs, each using a different initial pose for every complex.

The results show that mean AUROC remains relatively stable across runs, ranging from 0.542 to 0.559, whereas EF@1% exhibits greater variability, ranging from 1.64 to 2.30, suggesting that early enrichment is more sensitive to initial poses than overall ranking performance. This sensitivity may be attributed to two factors. First, even conventional MD simulations can exhibit dependence on initial ligand poses over finite simulation times, as different starting configurations may lead to distinct ligand rearrangements, protein-ligand contacts, and sampled conformational regions [2]. Second, although our data augmentation strategy described in Appendix B.1 is designed to reduce dependence on initial ligand configurations, the pose variations explored during training may not fully cover those introduced by independent docking. Expanding the diversity of initial poses during augmentation therefore represents a promising direction for improving robustness. Nevertheless, all five screening runs achieve EF@1% values above 1, indicating that ABLE retains meaningful early enrichment despite its sensitivity to initial poses.

AuthorsOpenReview ↗

Rebuttal 3

  • Q3: In Figure 4, it appears that there are often much larger deviations in interaction RMSF for high atom indices that are not found in the MD simulations. These deviations seem to occur at the end rather than at both the beginning and the end, as would be expected for the termini. Why is this?

Thank you for this observation. We would like to clarify that the atom indices in Figure 4 refer to the heavy atoms of the protein binding pocket and ligand, rather than the sequential residue positions of the full protein. In our representation, ligand atoms are placed after the protein pocket atoms, so the larger deviations at high indices primarily reflect differences in ligand fluctuations rather than increased mobility at the protein termini. These discrepancies may arise because the generative model overestimates ligand mobility relative to the binding pocket in certain systems, leading to larger RMSF values than those observed in reference MD trajectories. We have clarified the atom ordering in the figure to avoid this ambiguity.

References

[1] Guterres, H., & Im, W. (2020). Improving protein-ligand docking results with high-throughput molecular dynamics simulations. Journal of chemical information and modeling, 60(4), 2189.

[2] Liu, K., & Kokubo, H. (2017). Exploring the stability of ligand binding modes to proteins by molecular dynamics simulations: a cross-docking study. Journal of chemical information and modeling, 57(10), 2514-2522.

Reviewer_UK7cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing trajectory models mostly learn from active complexes and struggle to distinguish active ligands from decoys. ABLE addresses this through post-training on labeled static docking poses. Its main reward uses the protein–ligand interface’s “expansion rate” to encourage decoys to leave the pocket while keeping active ligands stable.

Critical strengths

The training data strategy is practical. Static active/decoy structures and labels avoid the need to collect costly decoy MD trajectories. The interface expansion rate provides a computable, mathematically defined signal for local optimization.

Critical weaknesses

The screening comparison lacks strong baselines with equivalent supervision. Table 3 mainly compares trajectory models that do not use screening labels, whereas ABLE’s post-training adds active/decoy supervision. Missing baselines include the original DOCK score and static structure or molecular fingerprint classifiers trained on the same labels. It remains unclear how much of the gain comes from trajectory modeling rather than added supervision, or whether generating 100 frames is necessary.

Discriminative trajectories need not reflect realistic dynamics. Using labels to push decoys out of pockets can improve RMSD-based ranking without making dissociation more physically accurate. The mathematical derivation of interface expansion does not establish a correspondence between the vector field in generative time and real dynamics. Physical fidelity after RL is also evaluated less thoroughly than for the base model. Appendix D.2 acknowledges excessive displacement for targets such as ROCK1.

Generalization and consistency remain limited. Training and testing cover the same 43 targets, with chemical-similarity splits only within each target; testing retains just 100 decoys per target. Mean AUROC of 0.542 is close to chance, and the across-target standard deviation of EF@1% is 4.760, indicating highly uneven gains. Validation on new targets, larger candidate libraries, and repeated sampling is needed. This standard deviation measures variation across targets and is not itself evidence of statistical nonsignificance.

The ablations do not establish whether the proposed mechanism is necessary. Experiments vary β, bandwidth, and the number of key points, but do not compare against simple displacement rewards, contact-reduction rewards, random key points, or alternative sampling strategies during training. The contribution of the interface expansion mechanism therefore remains unclear.

Questions for authors

None

5 comments
AuthorsOpenReview ↗

Rebuttal 1

We sincerely thank the reviewer for the thoughtful comments and constructive suggestions. Below, we address each concern in detail.

  • W1: The screening comparison lacks strong baselines with equivalent supervision.

Thank you for this valuable suggestion. We have expanded the virtual screening evaluation in Section 4.4 (Table 3) to include four additional baselines: AutoDock Vina, Boltz-2, EquiScore, and CORDIAL. The updated table is also provided below as Table A for reference. Specifically, Boltz-2 and EquiScore are evaluated using their publicly released pretrained weights, while AutoDock Vina scores the provided poses in score_only mode. For a more controlled comparison, we additionally train CORDIAL, a static structure-based virtual screening model, from scratch using the same DUDE-Z data split and active-decoy labels as ABLE.

On DUDE-Z, RL-finetuned ABLE achieves an EF@1% of 2.807 and a mean AUROC of 0.542, compared with 0.760 and 0.527, respectively, for CORDIAL. The substantially higher early enrichment suggests that our trajectory-based RL approach can extract more informative structural signals from limited cross-docked data than the evaluated static prediction baseline. ABLE also narrows the gap with AutoDock Vina in early enrichment, achieving an EF@1% of 2.807 compared with 3.591 for Vina.

We acknowledge that ABLE still underperforms pretrained virtual screening methods, particularly Boltz-2, which achieves substantially stronger overall screening performance. However, these methods may benefit from larger and more diverse training datasets, and potential overlap between their pretraining data and our test set cannot be ruled out, making direct comparisons less controlled. A more controlled comparison may require training on similarly large-scale datasets, which would extend beyond the primary scope of this work. Instead, our goal is to demonstrate that trajectory generation models can acquire meaningful virtual screening capabilities through lightweight RL finetuning using readily available static cross-docked complex structures, without requiring additional MD trajectories. We believe this represents a promising step toward extending trajectory generation models to practical virtual screening applications.

Table A: Virtual screening results on the DUDE-Z test set. ∗^* indicates models trained from scratch using the same data split.

MethodsEF@1% (↑)EF@2% (↑)EF@5% (↑)EF@10% (↑)AUROC median (↑)AUROC mean (↑)AUROC std
Virtual Screening Method
AutoDock Vina3.591/6.3184.257/6.6303.547/5.0612.548/3.0330.6030.5910.216
Boltz-218.75/11.0816.08/8.16412.30/5.8787.184/3.1030.9930.9090.144
EquiScore4.395/6.7633.953/5.3523.238/4.1582.494/2.9630.6250.6200.185
CORDIAL*0.760/2.0050.974/2.0881.195/2.0891.120/1.2450.5440.5270.107
Trajectory Generation Model
NeuralMD-SDE1.214/3.6601.422/3.2290.796/1.6720.831/1.2400.4950.4780.168
PVB0.430/2.1340.488/1.9110.862/1.6760.980/1.1830.5350.5290.126
ABLE (ref)0.524/1.9600.551/1.8190.965/1.7341.034/1.2070.5230.4990.158
ABLE (rl)2.807/4.7602.009/3.4651.754/2.1791.593/1.7440.5530.5420.185
  • W2: Whether generating 100 frames is necessary is unknown.

Thank you for raising this point. We adopt 100 frames as a practical evaluation horizon rather than a strictly necessary trajectory length. Specifically, ABLE uses a physical time lag of τ=80\tau=80 ps for protein-ligand complexes, so a 100-frame rollout corresponds to approximately 8 ns. This is comparable to the timescale of short MD simulations used in prior studies to assess ligand binding stability for virtual screening [1]. Moreover, as illustrated in Figure 5, the ligand RMSD curves for several representative targets exhibit a slower rate of change toward the end of the rollout, suggesting that the generated trajectories can capture informative differences in binding-pose stability within this horizon.

AuthorsOpenReview ↗

Rebuttal 2

  • W3: Discriminative trajectories need not reflect realistic dynamics. Physical fidelity after RL is also evaluated less thoroughly than for the base model.

Thank you for raising this important concern. We agree that improved active-decoy discrimination does not necessarily imply physically accurate ligand dissociation dynamics. Our primary goal is to introduce such discriminative capabilities into trajectory generation models using readily available static cross-docked structures, rather than faithfully reproducing ligand dissociation pathways, which, to the best of our knowledge, would inevitably require MD trajectories of actual protein-decoy complexes for training.

Although our RL procedure does not guarantee faithful reproduction of real molecular dynamics, we can still assess the physical plausibility of generated trajectories to gain insight into their physical fidelity. To this end, we have added a comprehensive physical fidelity evaluation in Appendix D.4 (Table 8), covering ligand geometric validity, intra- and intermolecular steric clashes, and protein CA validity on DUDE-Z. For convenience, we also provide the results in Table B below. Compared with the reference checkpoint, RL finetuning with β=0.2\beta=0.2 largely preserves geometric validity and improves aromatic-ring and double-bond planarity, although intramolecular clash rates increase from 0.072 to 0.146. Increasing the KL regularization weight to β=0.4\beta=0.4 reduces this rate to 0.131, while restoring bond-length, bond-angle, and protein validity to levels comparable to the reference model. Notably, no intermolecular steric clashes are observed across these configurations. These results suggest that RL can improve screening performance while largely preserving structural plausibility, with stronger regularization helping mitigate physical distortions.

Table B: Physical fidelity of generated trajectories on the DUDE-Z test set.

ModelsLigand Validity: Bond Length (↑)Ligand Validity: Bond Angle (↑)Ligand Validity: Aromatic (↑)Ligand Validity: Double Bond (↑)Steric Clashes: intra (↓)Steric Clashes: inter (↓)Protein Validity (↑)
NeuralMD-SDE1.000/0.0001.000/0.0001.000/0.0001.000/0.0000.000/0.0000.000/0.000-
PVB0.978/0.0100.991/0.0040.957/0.0280.980/0.0220.026/0.0100.000/0.0000.724/0.154
ABLE (ref)0.926/0.0420.949/0.0260.890/0.0380.978/0.0220.072/0.0390.000/0.0000.990/0.013
ABLE (rl, β=0\beta=0)0.838/0.1250.863/0.1280.908/0.0800.981/0.0200.168/0.1310.000/0.0000.812/0.192
ABLE (rl, β=0.2\beta=0.2)0.906/0.0700.939/0.0670.949/0.0410.981/0.0200.146/0.1070.000/0.0000.979/0.056
ABLE (rl, β=0.4\beta=0.4)0.925/0.0420.960/0.0240.960/0.0150.982/0.0180.131/0.0650.000/0.0000.991/0.025
  • W4: Generalization and consistency remain limited. Validation on new targets, larger candidate libraries, and repeated sampling is needed.

Thank you for this important suggestion. We agree that the original DUDE-Z evaluation primarily assesses generalization to chemically dissimilar ligands within the same protein targets, leaving cross-target generalization insufficiently explored. To address this concern, we have added a new experiment in Section 4.5 (Table 4) using seven previously unseen protein targets from DUD-E: DPP4, DRD3, BRAF, CDK2, FAK1, MK10, and MK14. For convenience, we also provide the results in Table C below. For each target, we sample 1,000 compounds while approximately preserving the original active-to-decoy ratio exclusively for evaluation. Importantly, the initial complex structures are obtained from an independent GNINA docking dataset, introducing a different docking protocol from that used in DUDE-Z.

Despite being RL-finetuned only on the 43 DUDE-Z targets, ABLE demonstrates promising cross-target transferability, achieving an EF@1% of 2.754, slightly exceeding AutoDock Vina at 2.654 and approaching EquiScore at 3.321. Its mean AUROC of 0.538 also remains comparable to the in-distribution result of 0.542, but falls below the pretrained AutoDock Vina and EquiScore. This gap may stem from insufficient coverage of diverse decoy configurations during training, limiting the model's ability to reliably assess their binding-pose stability. In contrast, CORDIAL, trained from scratch on the same DUDE-Z split, achieves an AUROC of 0.487 and no enrichment at the top 1%. These results suggest that our trajectory-based RL approach can learn structural signals that remain informative on unseen protein targets, achieving stronger cross-target generalization than the evaluated static prediction baseline even under a different docking protocol.

AuthorsOpenReview ↗

Rebuttal 3

Table C: Virtual screening results on 7 unseen protein targets from DUD-E. ∗^* indicates models trained from scratch using the same data split.

MethodsEF@1% mean (↑)EF@1% stdAUROC median (↑)AUROC mean (↑)AUROC std
AutoDock Vina2.6543.3460.7410.7140.071
EquiScore3.3214.4710.6020.6400.125
CORDIAL*0.0000.0000.4880.4870.072
ABLE (rl)2.7543.4550.5460.5380.063
  • W5: The ablations do not establish whether the proposed mechanism is necessary. Experiments do not compare against simple displacement rewards, contact-reduction rewards, random key points, or alternative sampling strategies during training.

Thank you for this valuable suggestion. To better assess the necessity of the proposed RL mechanism, we have added a dedicated ablation study in Appendix D.3.1 (Table 6), comparing our approach with two simpler reward functions based on ligand displacement and protein-ligand contact retention, as well as a variant that replaces the proposed keypoint scoring strategy with uniform random selection of interfacial keypoints. For convenience, we also provide the results in Table D below.

The results show that the simple displacement and contact-based rewards achieve limited screening performance, with EF@1% values of 1.098 and 1.054, respectively, both close to random ranking. In contrast, our full method achieves an EF@1% of 2.807 and a mean AUROC of 0.542, substantially outperforming these alternatives. Interestingly, random keypoint selection retains some discriminative capability, achieving an EF@1% of 1.363 and a mean AUROC of 0.535. This suggests that the interfacial expansion objective can provide useful optimization signals even without the proposed keypoint scoring strategy, while selecting more informative interfacial locations further improves early enrichment. Overall, these results demonstrate the effectiveness of combining the interfacial expansion reward with the proposed keypoint selection strategy, beyond what can be achieved with the simpler alternatives evaluated here.

Table D: Ablation study of the reinforcement learning mechanism on the DUDE-Z test set.

MethodsEF 1% (↑)EF 2% (↑)EF 5% (↑)EF 10% (↑)AUROC median (↑)AUROC mean (↑)AUROC std
ABLE (full)2.807/4.7602.009/3.4651.754/2.1791.593/1.7440.5530.5420.185
  w/ displacement1.098/3.3831.117/2.8461.291/2.3090.842/1.3430.4880.4820.165
  w/ contact1.054/3.9920.866/2.8261.585/3.1621.127/1.8800.4930.5030.191
  w/o keypoint scoring1.363/3.7031.273/2.8701.366/2.4001.284/1.4890.5280.5350.178

References

[1] Guterres, H., & Im, W. (2020). Improving protein-ligand docking results with high-throughput molecular dynamics simulations. Journal of chemical information and modeling, 60(4), 2189.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their time, careful evaluation, and constructive feedback. In response to their comments, we have conducted additional experiments and substantially revised the manuscript, with the updated text highlighted in red. Below, we summarize the major revisions addressing several concerns shared across the reviews.

  1. Comparison with established virtual screening methods (Section 4.4, Table 3). We have expanded our evaluation to include AutoDock Vina, Boltz-2, EquiScore, and CORDIAL. In particular, CORDIAL is trained from scratch using the same DUDE-Z data split as ABLE to provide a more controlled comparison with static structure-based prediction. The results show that ABLE underperforms specialized virtual screening methods such as Boltz-2 and EquiScore, potentially due to their substantially larger and more diverse training datasets, suggesting that further large-scale training is needed to close this gap. However, under a more controlled comparison, RL-finetuned ABLE achieves substantially higher early enrichment than CORDIAL trained on the same DUDE-Z data split, with an EF@1% of 2.807 compared with 0.760, highlighting the potential of trajectory-based RL to extract more informative structural signals from limited cross-docked data than prediction baselines.
  2. Physical fidelity after RL finetuning (Appendix D.4, Table 8). We have added a systematic evaluation of ligand geometric validity, steric clashes, and protein structural validity on DUDE-Z. The results indicate that RL finetuning largely preserves the structural plausibility of generated trajectories. Although intramolecular clashes increase after RL, stronger path-space KL regularization mitigates this degradation while maintaining other physical validity metrics at levels comparable to the reference model. These findings provide additional evidence that screening performance can be improved without substantially compromising structural plausibility.
  3. Cross-target generalization (Section 4.5, Table 4). To evaluate generalization beyond the 43 training targets, we have constructed an external test set comprising 7,000 protein-ligand pairs across seven unseen DUD-E targets, using docking poses obtained from a different protocol. Despite being RL-finetuned only on DUDE-Z, ABLE achieves an EF@1% of 2.754, slightly exceeding AutoDock Vina at 2.654 and approaching EquiScore at 3.321. Its mean AUROC of 0.538 also remains comparable to the in-distribution result of 0.542. In contrast, CORDIAL, trained from scratch on the same DUDE-Z split, exhibits a decline in mean AUROC from 0.527 to 0.487, while its EF@1% drops from 0.760 to 0.000. These results suggest that our trajectory-based RL approach exhibits stronger cross-target transferability and reduced susceptibility to target-specific overfitting compared with the evaluated static prediction baseline.
  4. Additional ablation studies (Appendix D.3). We have expanded the ablation analyses to investigate the effectiveness of the proposed RL mechanism (Appendix D.3.1), sensitivity to key hyperparameters including path-space KL regularization (Appendix D.3.2), and robustness to alternative initial docking poses (Appendix D.3.3). These experiments aim to clarify the contributions of individual design choices and assess the sensitivity of the proposed approach to training configurations and initial structures.

We hope these additional experiments and revisions address the reviewers' main concerns and provide a more comprehensive evaluation of ABLE. Detailed responses to the remaining questions and suggestions are provided in our individual replies to each reviewer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.