acceptodds
Under review as a conference paper at ICLR 2027

Making the Most of Molecular Dynamics Simulations for Binding Affinity Prediction

Abstract

Molecular dynamics (MD) simulations provide rich conformational information for protein–ligand binding affinity prediction, yet existing approaches typically treat simulated conformation samples independently or compress trajectory-level dynamics into static representations, leaving fine-grained interaction dynamics underexploited. In this work, we introduce DrugDyn, a unified framework that makes more effective use of molecular dynamics simulations for protein–ligand binding affinity prediction. DrugDyn integrates Geometry-Aware Spatiotemporal Attention (GASA) to capture geometry-dependent interactions within individual conformations and their dependencies across conformations, together with Multi-Cardinality Contrastive Learning (MCCL) to learn coherent and discriminative complex-level representations from different subsets of the conformational ensemble. Experiments across five backbone architectures under both standard PDBbind core-set evaluation and cross-protein LBA splits demonstrate consistent gains in both predictive accuracy and ranking consistency. On the challenging LBA-30 split, DrugDyn reduces RMSE by 13.8% and improves Kendall's by 9.3% on average over the baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
binding affinity prediction, molecular dynamics, spatiotemporal modeling, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_AzLMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors point out that molecular dynamics (MD) simulations provide rich conformational information for protein–ligand binding affinity prediction, yet existing methods make limited use of it. Most approaches describe a complex using a single static conformation, which captures only a snapshot of an inherently dynamic binding process. Methods that do incorporate MD either treat sampled conformations as independent training instances (data augmentation) or compress trajectory-level dynamics into static graph edges, leaving fine-grained interaction dynamics underexploited. To address this, the authors propose DrugDyn, a unified framework that makes more effective use of MD simulations. It consists of two components. Geometry-Aware Spatiotemporal Attention (GASA) captures geometry-dependent protein–ligand interactions within each conformation and models how each atom’s interaction environment varies across conformations, followed by an adaptive aggregation that weights conformations by their importance. Multi-Cardinality Contrastive Learning (MCCL) treats subsets containing different numbers of conformations as different views of the same complex, aligning views from the same complex while separating those from different complexes, thereby yielding coherent and discriminative complex-level representations. DrugDyn is evaluated on five representative structure-based models used as backbones, and compared against the static baseline as well as two MD-based strategies: MD-based data augmentation (DA) and motion-derived edge augmentation (MDEA). On both the standard PDBbind core-set benchmarks and the cross-protein LBA splits, DrugDyn outperforms these alternatives across all backbones.

Critical strengths
  1. The experiments are extensive, covering five representative structure-based models (EGNN, GIGN, DTIGN, DeepRLI, and BA-Pred) as backbones, which supports the claim that the framework is backbone-agnostic.
  2. The evaluation uses a comprehensive set of metrics, covering both prediction error and the linear and ranking consistency between predicted and experimental binding affinities.
  3. Motivated by the observation that existing MD-based strategies map complexes with very different binding affinities to similar latent representations, the authors design MCCL, which treats conformational subsets of different sizes as views of the same complex, to make the representations of different complexes more distinguishable.
Critical weaknesses
  1. Conformation selection. The authors select 10 representative conformations from the 100 provided by MISATO, but the selection criterion is not described. Since the method relies on these conformations to represent the ensemble, the selection strategy may affect the results and should be specified.
  2. Inconsistency in the MCCL design. MCCL constructs each view by simply averaging conformation embeddings, and the view with k = 1 reduces to a single conformation. This seems to contradict the authors’ own motivation in Section 3.2 that “directly treating individual conformations as separate views may be suboptimal.”
  3. Limited ablation. The ablation study is conducted only on the v2016 core set with three of the five backbones. It does not cover LBA-30, which is the most informative benchmark for assessing cross-protein generalization.
Questions for authors

I have some questions about the data splits used in the experiments, as the paper does not describe them in detail.

  1. Overlap between the training set and the core sets. The paper states that models are trained on the PDBbind v2019 general set and evaluated on the v2013 and v2016 core sets. Since PDBbind versions are cumulative and the core set is a subset of the refined set, which is in turn a subset of the general set, the v2019 general set contains the complexes in both the v2013 and v2016 core sets. Could the authors clarify whether these core-set complexes were removed from the training data? In addition, does “general set” here refer to the full general set (including the refined set), or to the general set excluding the refined set as distributed in the PDBbind download?
  2. Training data for the LBA experiments. The LBA-30 and LBA-60 splits are derived from the PDBbind v2019 refined set, which is also contained in the v2019 general set. Could the authors clarify which data the models were trained on in the LBA experiments? Specifically, were the models trained on the training portion of each LBA split, or were the models trained on the v2019 general set reused for evaluation on the LBA test sets? I would also appreciate it if the authors could release the code or the exact lists of training, validation, and test complexes, which would make these settings easier to verify.
5 comments
AuthorsOpenReview ↗

Response to Reviewer AzLM (Part 1)

We sincerely thank the reviewer for the thoughtful and insightful feedback. We have carefully considered the concerns raised and provide detailed point-by-point responses below.

In response to Weakness 1

Following MDDTA [1], we adopt a clustering-based sampling strategy to select 10 representative conformations from the 100 MD trajectory frames provided by MISATO. Specifically, the 100 frames are clustered based on their three-dimensional atomic coordinates into 10 clusters, and one representative conformation is selected from each cluster. This strategy aims to capture diverse structural states rather than selecting frames randomly or at fixed temporal intervals.

To further evaluate the impact of different conformation selection strategies on DrugDyn's performance, we conducted additional experiments on the PDBbind v2016 core set using GIGN and DTIGN as backbones. We evaluated random frame sampling (RF), contiguous temporal windows (CTW), and uniform temporal sampling with a fixed interval of 10 frames. We maintained 10 frames, consistent with our default setting. The results are summarized below.

BackboneSampling StrategyRMSE↓MAE↓RR↑ρ\rho↑τ\tau↑
GIGNBaseline1.7601.3760.6460.6470.464
Ours (RF)1.4751.1340.7370.7280.540
Ours (CTW)1.4751.1400.7360.7250.534
Ours (Fixed Interval 10)1.4511.1250.7470.7440.554
Ours1.3301.0400.7980.8030.604
DTIGNBaseline1.5231.2310.7230.7050.515
Ours (RF)1.4981.1560.7340.7340.543
Ours (CTW)1.4721.1670.7370.7260.535
Ours (Fixed Interval 10)1.4511.1250.7470.7440.554
Ours1.3301.0700.8020.7900.597

As shown above, DrugDyn consistently outperforms the corresponding baselines under all evaluated sampling strategies, while the clustering-based selection achieves the best performance on both backbones, suggesting that capturing representative and diverse structural states is beneficial for modeling protein–ligand interactions. We will clarify the conformation selection procedure in the revised manuscript and incorporate these additional experimental results into the manuscript or supplementary materials, as appropriate.

[1] Zhao, Long, et al. "MDDTA: A Drug Target Binding Affinity Prediction Method Based on Molecular Dynamics Simulation Data Enhancement." IEEE Journal of Biomedical and Health Informatics (2026).


In response to Weakness 2

We would like to clarify that the averaging operation in MCCL serves to construct multi-conformation views and does not need to model cross-conformation interactions. This is because the conformation embeddings have already undergone spatiotemporal modeling through GASA, which explicitly captures atom-level dependencies across conformations. MCCL therefore only needs to average these representations over different conformational subsets to construct views of varying cardinalities.

Furthermore, our motivation is that relying solely on individual conformations as contrastive views may be suboptimal, rather than excluding single-conformation views entirely. We recognize that our original statement, "directly treating individual conformations as separate views may be suboptimal," could be ambiguous. To improve clarity, we will revise it to "Relying solely on individual conformations as contrastive views may be suboptimal." In addition to k=1k=1, MCCL incorporates k=3,5,7k=3,5,7 to capture different observations of the underlying conformational ensemble, thereby encouraging consistent complex-level representations across different view cardinalities while reducing the influence of individual conformational fluctuations and thermal noise.

AuthorsOpenReview ↗

Response to Reviewer AzLM (Part 2)

In response to Weakness 3

To further validate the effectiveness of our proposed components under cross-protein generalization, we extended our ablation study to the LBA-30 benchmark using GIGN and DTIGN as backbones. The additional results are summarized below.

BackboneMethodRMSE↓MAE↓RR↑ρ\rho↑τ\tau↑
EGNNBaseline2.0841.6220.4400.4460.307
GASA1.7531.3960.4540.4530.312
SCCL1.8361.4540.4140.4220.289
MCCL1.7791.4020.4490.4540.313
Full1.7351.3920.4580.4620.318
GIGNBaseline1.9971.6170.4860.4660.319
GASA1.6211.2790.5210.4840.339
SCCL1.8131.4000.4360.4370.306
MCCL1.6641.3560.4640.4550.313
Full1.4981.1980.5260.5000.350
DTIGNBaseline1.9671.5810.4120.4030.276
GASA1.8291.4360.4690.4700.328
SCCL1.9001.5560.4340.4140.283
MCCL1.6751.3350.4590.4480.308
Full1.5661.2810.5040.4800.331

As shown above, both GASA and MCCL contribute to improving prediction performance, while their combination consistently achieves the best results across all five metrics on both backbones. These findings further demonstrate the complementary benefits of the two components under cross-protein generalization.


In response to Question 1

We clarify that the PDBbind v2019 general set refers to the full general set, including the refined set. In addition, we explicitly excluded all complexes belonging to the PDBbind v2013 and v2016 core sets from the training data to prevent direct overlap between the training and test sets. Therefore, neither core set contains complexes used during training.


In response to Question 2

We clarify that for the LBA-30 and LBA-60 experiments, all models were trained independently on the training portion of their respective splits, rather than reusing models trained on the PDBbind v2019 general set. Specifically, following the established ATOM3D protocol [1], also adopted by CheapNet [2], we use the PDBbind v2019 refined set with protein sequence identity thresholds of 30% and 60% to construct the LBA-30 and LBA-60 benchmarks, respectively. For each benchmark, we use its corresponding training set for model training, validation set for model selection, and held-out test set for evaluation.

The exact training, validation, and test split indices are publicly available through the official ATOM3D dataset release (https://zenodo.org/records/4914718), including separate split files for LBA-30 and LBA-60. We strictly follow these official splits in our experiments.

[1] Townshend, Raphael JL, et al. "Atom3d: Tasks on molecules in three dimensions." arXiv preprint arXiv:2012.04035 (2020).

[2] Lim, Hyukjun, Sun Kim, and Sangseon Lee. "Cheapnet: Cross-attention on hierarchical representations for efficient protein-ligand binding affinity prediction." International Conference on Learning Representations. Vol. 2025. 2025.

Thank you to the authors for the detailed and constructive rebuttal. I appreciate the additional experiments on conformation sampling strategies and the extended ablation study on LBA-30, which address two of my major concerns.

However, I noticed that the results for GIGN and DTIGN under the "Fixed Interval 10" sampling strategy are exactly identical across all five metrics (RMSE: 1.451, MAE: 1.125, R: 0.747, ρ: 0.744, τ: 0.554). Given that these are two different backbone models, this coincidence seems unusual. Could the authors please verify whether these results are correct or whether there might be a reporting error in the table?

The clarification regarding dataset splits is also helpful and largely resolves my concerns about potential training–test overlap.

Overall, the rebuttal has addressed most of my concerns and improved my assessment of the paper. I am inclined to reconsider my initial rating, pending clarification of the identical results noted above.

Reviewer_t9wEOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

DrugDyn predicts protein–ligand binding affinity from several conformations in a molecular dynamics (MD) trajectory. Geometry-Aware Spatiotemporal Attention (GASA) applies sparse protein–ligand cross-attention with distance-based radial-basis biases in each conformation, then uses atom-wise self-attention across conformations before adaptive aggregation. Multi-Cardinality Contrastive Learning (MCCL) averages conformation-level representations over subsets of cardinality 1, 3, 5, and 7, treats those subsets as views of one complex, and contrasts them with views from other complexes. The authors add the two modules to five structural backbones and report results on PDBbind v2013/v2016 core sets and Atom3D LBA-30/LBA-60 split.

Critical strengths
  • The paper models cross-conformation dependencies at the atom level. Many MD-based affinity models treat sampled conformations as independent augmented examples or reduce the trajectory to static graph edges.
  • The GASA and MCCL design choices are sensible.
  • The evaluation covers five backbones, which checks whether the framework extracts information that transfers across architectures.
Critical weaknesses
  • The paper argues that DrugDyn uses MD samples more fully, but the result depends on how frames are selected. It does not discuss the effect of frame count, inter-frame similarity, or time span. Without that, it is hard to tell whether the gain holds for arbitrary MD sampling and frame selection or only for one set of continuous frames.
  • Efficiency and accuracy are both important for affinity scoring, especially when screening large compound libraries. DrugDyn processes 10 conformations and adds cross-molecular attention, temporal attention, and contrastive subsets, but the paper does not report training time, GPU memory, or inference latency relative to the baseline, DA, and MDEA.
  • The experimental details and uncertainty analysis are thin. Appendix B is only a short implementation pointer. The paper omits DA/MDEA settings, number of repeats, model-selection criteria, and full backbone hyperparameters. It also does not assess random seeds, conformer selection, or MD sampling time; small ablation gaps, such as GIGN + MCCL lowering RMSE from 1.464 to 1.330 and DTIGN + MCCL from 1.400 to 1.330 on v2016, cannot be interpreted without that analysis.
  • The learned components are not analyzed biologically. The paper does not show whether adaptive aggregation favors low-energy or interaction-rich conformations, or whether MCCL embeddings separate complexes in a way that tracks affinity or known interaction structure. Without that, the mechanism stays architectural.
Questions for authors
  • How does performance depend on conformer selection? Does using more conformations improve results? Were the 10 of 100 conformations chosen randomly, uniformly in time, by clustering, or by energy, and does the advantage hold under a different selection rule?
  • The paper stresses structural effects, but LBA-30/60 are split by protein sequence identity. Do the splits still differ in structural similarity, and does the observed advantage remain meaningful under that stratification?
  • What do the temporal attention and adaptive aggregation weights show? Are they stable across complexes, and do they correlate with MM/GBSA energies or interaction frequencies?
7 comments
AuthorsOpenReview ↗

Response to Reviewer t9wE (Part 1)

We sincerely thank the reviewer for the positive assessment of our work and recognition of its strengths. We appreciate the constructive feedback and address the concerns in detail below.

In response to Weakness 1

We first clarify that our default setting does not use consecutive or uniformly sampled frames. Instead, following MDDTA [1], as stated in Section 4.1 (Experimental Setup), we adopt a clustering-based sampling strategy that groups 100 MD trajectory frames into 10 clusters and selects one representative conformation from each cluster. This strategy aims to capture substantial structural variations.

To systematically investigate the effects of frame selection, we conducted additional experiments on the PDBbind v2016 core set using GIGN and DTIGN as backbones. Specifically, we evaluated random frame sampling (RF), contiguous temporal windows (CTW), different numbers of sampled frames (3 and 5), and uniform temporal sampling with a fixed interval of 10 frames. For RF, CTW, and fixed-interval sampling, we maintained 10 frames, consistent with our default setting. The results are presented below.

BackboneSampling StrategyRMSE↓MAE↓RR↑ρ\rho↑τ\tau↑
GIGNBaseline1.7601.3760.6460.6470.464
Ours (RF)1.4751.1340.7370.7280.540
Ours (CTW)1.4751.1400.7360.7250.534
Ours (3 frames)1.4501.1500.7500.7400.547
Ours (5 frames)1.4321.0920.7540.7490.557
Ours (Fixed Interval 10)1.4511.1250.7470.7440.554
Ours (10 frames)1.3301.0400.7980.8030.604
DTIGNBaseline1.5231.2310.7230.7050.515
Ours (RF)1.4981.1560.7340.7340.543
Ours (CTW)1.4721.1670.7370.7260.535
Ours (3 frames)1.4441.1650.7490.7410.548
Ours (5 frames)1.4401.1230.7510.7570.566
Ours (Fixed Interval 10)1.4751.1770.7430.7350.541
Ours (10 frames)1.3301.0700.8020.7900.597

First, regarding frame count, increasing the number of sampled conformations from 3 to 5 and subsequently to 10 consistently improves performance across all five metrics for both backbones. This suggests that incorporating more representative conformations enables DrugDyn to capture richer structural variations and protein–ligand interactions.

Second, we observe that random sampling, contiguous temporal windows, and fixed-interval sampling all achieve improvements over their respective baselines, although their performance is lower than that of our default setting. We note that these alternative strategies may select structurally similar conformations, introducing redundant information and limiting conformational diversity. In contrast, the clustering-based strategy adopted in our study selects representative conformations from different structural clusters, allowing DrugDyn to better characterize variations in protein–ligand interactions throughout the MD trajectory.

Importantly, DrugDyn consistently outperforms the corresponding baselines under all evaluated sampling strategies, demonstrating that its effectiveness is not restricted to a specific set of selected frames. Nevertheless, the performance differences also indicate that conformational diversity and frame selection are important factors in fully exploiting MD information. We will incorporate these additional results and discussions into the revised manuscript.

[1] Zhao, Long, et al. "MDDTA: A Drug Target Binding Affinity Prediction Method Based on Molecular Dynamics Simulation Data Enhancement." IEEE Journal of Biomedical and Health Informatics (2026).

AuthorsOpenReview ↗

Response to Reviewer t9wE (Part 2)

In response to Weakness 2

We conducted additional experiments using GIGN as the backbone to compare training time, peak GPU memory consumption, and average inference latency across the baseline, DA, MDEA, and DrugDyn under the same experimental setting. The results are summarized below.

MethodTraining Time (s/epoch)Peak GPU Memory (GiB)Inference Latency (ms/complex)
Baseline6.095.053.41
DA54.145.053.87
MDEA16.069.333.89
DrugDyn (Ours)57.3662.3025.69

As shown above, although DrugDyn introduces additional computational overhead, it still achieves millisecond-level inference latency per complex, making it feasible for large-scale compound screening. Moreover, experimentally measured protein-ligand complexes are relatively small compared with large-scale graph data, such as social networks. Together with ongoing advances in GPU hardware, these computational costs are increasingly manageable, particularly for pharmaceutical companies with sufficient computational resources.


In response to Weakness 3

For DA, the ten selected MD conformations of each complex are treated as independently labeled training graphs. For MDEA, displacement vectors are calculated between pairs of conformations. For each protein–ligand atom pair, we compute the cosine similarity between their displacement vectors and average the similarities across conformation pairs. A bidirectional edge is added when the mean similarity exceeds 0.8. For all four methods (baseline, DA, MDEA, and DrugDyn), models are trained for 50 epochs using AdamW with a learning rate of 1×10−41 \times 10^{-4}. The batch size is set to 128 for EGNN, GIGN, and DTIGN, and reduced to 32 for DeepRLI and BA-Pred due to GPU memory constraints. For each training run, we split the training set into training and validation subsets at a 9:1 ratio for the standard PDBbind benchmarks, while using the predefined validation split for LBA. We save the checkpoint with the lowest validation RMSE and use it for final evaluation.

The full hyperparameter settings for each backbone are provided below:

EGNN & GIGN & DTIGN. We use three graph message-passing layers for both backbones, following the architectures described in their respective original papers, with a hidden dimension of 256 and an MLP regression head with LeakyReLU activations and batch normalization.

DeepRLI. We use ten graph transformer layers with eight attention heads per layer, 33 Gaussian radial basis functions, a distance cutoff of 6.5 Å, a feed-forward expansion ratio of 2, and a two-layer regression head.

BA-Pred. We use six update iterations, a 20-dimensional random-walk positional encoding, and a two-layer regression head.

Regarding the effects of conformer selection and MD sampling time, please refer to our detailed response to Weakness 1. Following the reviewer's suggestion, we conducted additional ablation experiments with five random seeds on the PDBbind v2016 core set using GIGN and DTIGN as backbones. We further assessed statistical significance using the Welch's t-test. The results are summarized below.

BackboneMethodRMSE (Mean ± Std)p-value
GIGNBaseline1.575 ± 0.017-
+ MCCL1.458 ± 0.0285.65×10−5\mathrm{5.65 \times 10^{-5}}
Full (DrugDyn)1.369 ± 0.0220.0003
DTIGNBaseline1.534 ± 0.020-
+ MCCL1.428 ± 0.0343.64×10−4\mathrm{3.64 \times 10^{-4}}
Full (DrugDyn)1.365 ± 0.0270.0063

The p-values compare +MCCL against the baseline and the full model against +MCCL, respectively.

As shown above, MCCL significantly reduces RMSE compared with the corresponding baselines on both backbones (p=5.65×10−5p = 5.65 \times 10^{-5} and 3.64×10−43.64 \times 10^{-4}). Furthermore, the full model achieves additional statistically significant improvements over MCCL alone (p=0.0003p = 0.0003 and 0.00630.0063). All p-values are below 0.05, demonstrating that the improvements remain statistically significant across multiple random seeds.

AuthorsOpenReview ↗

Response to Reviewer t9wE (Part 3)

In response to Weakness 4

To address the reviewer's concern, we first analyzed the learned aggregation weights of ten MD conformations alongside their corresponding MM/GBSA interaction energies. Specifically, we performed inference using GIGN on the PDBbind v2013 and v2016 core sets to obtain the aggregation weights of individual MD conformations for each protein–ligand complex. We then normalized the interaction energies to the range [−1, 0] within each complex and combined the results across all complexes for statistical analysis. We observed a statistically significant association between lower interaction energies and higher aggregation weights on both v2013 (p=4.34×10−13p = 4.34 \times 10^{-13}) and v2016 (p=5.54×10−12p = 5.54 \times 10^{-12}). We then show two representative complexes, 1OWH and 2YGE, where more negative interaction energies indicate more favorable protein–ligand interactions.

PDB IDFrameInteraction EnergyWeight
1OWH5−22.020.025
22−26.210.088
27−33.640.198
40−22.280.068
44−30.550.055
50−26.570.084
66−33.450.163
84−31.210.115
89−36.660.137
100−26.430.067
2YGE2−25.780.023
6−34.200.266
15−27.520.072
23−23.860.023
34−28.760.040
43−20.500.050
53−24.330.025
80−31.740.262
90−26.150.033
96−28.850.206

For 1OWH, conformations with interaction energies below −30 receive an average aggregation weight of 0.134, approximately twice that of the remaining conformations (0.066). Similarly, for 2YGE, the two conformations with interaction energies below -30 receive substantially higher weights of 0.266 and 0.262.

Furthermore, using the same inference setup, we computed the pairwise differences in experimental binding affinities and the cosine similarities between the corresponding complex embeddings. We observed a statistically significant association between smaller affinity differences and higher embedding similarities on both v2013 (p=0.0072p = 0.0072) and v2016 (p=0.0026p = 0.0026). We then selected three representative protein-ligand complexes to illustrate this relationship.

Reference (pK)Similar-pK ComplexCos. Sim.Different-pK ComplexCos. Sim.
2J7H (7.19)2VVN (7.30)0.8603O9I (11.82)−0.363
1YC1 (6.17)2FXS (6.06)0.8484QAC (9.40)−0.102
3E5A (8.23)4E6Q (8.36)0.8302W66 (4.05)−0.090

As shown above, complexes with similar binding affinities exhibit high embedding similarities, whereas those with substantially different affinities exhibit much lower similarities. These results suggest that MCCL embeddings can capture affinity-related differences between protein–ligand complexes. We will incorporate these additional results and discussions into the revised manuscript or supplementary materials, as appropriate.


In response to Question 1

As detailed in our response to Weakness 1, following MDDTA [1], we adopt a clustering-based sampling strategy that groups 100 MD trajectory frames into 10 clusters and selects one representative conformation from each cluster, rather than using random, uniform temporal, or energy-based selection. We further conducted additional experiments with different frame counts (3, 5, and 10) and alternative sampling strategies, including random sampling, contiguous temporal windows, and fixed-interval sampling. The results show that increasing the number of conformations consistently improves performance, highlighting the need to balance predictive performance against computational cost when selecting the number of conformations. Importantly, DrugDyn outperforms the corresponding baselines under all evaluated sampling strategies, demonstrating that its effectiveness is not limited to the default selection rule. Please refer to our response to Weakness 1 for the detailed experimental results and analysis.

Reviewer_o727OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DrugDyn, a framework for using molecular-dynamics conformations in protein–ligand binding affinity prediction. Geometry-Aware Spatiotemporal Attention (GASA) models geometry-dependent intermolecular interactions within each conformation and dependencies across conformations. Multi-Cardinality Contrastive Learning (MCCL) treats subsets with different numbers of conformations as different views of the same complex. Experiments use five backbone architectures on PDBbind core sets and LBA-30/LBA-60 protein-level splits, comparing DrugDyn with static baselines, data augmentation, and motion-derived edge augmentation.

Critical strengths
  1. The evaluation across five backbone architectures and both standard and protein-level generalization splits provides evidence that the approach is not tied to one encoder.\n\n2. The multi-cardinality contrastive objective is a reasonable attempt to model that an MD trajectory is only a partial observation of a conformational ensemble. The ablations of GASA, MCCL, and single-conformation contrastive learning are useful.
Critical weaknesses
  1. The central motivation is well established, and the methodological contribution appears to combine spatial attention, temporal attention, and contrastive learning. The paper should better articulate what is fundamentally different from existing temporal, multi-view, and trajectory-level representation learning.\n\n2. MCCL assumes that different subsets of conformations from the same complex should be close in representation space. This may be problematic when different conformations correspond to distinct binding states whose differences are informative. The results also depend on selecting ten representative conformations, but sensitivity to frame selection, temporal order, frame count, and sampling interval is not sufficiently established.
Questions for authors
  1. What happens when the temporal order of the sampled conformations is randomly permuted?\n\n2. How do results change with random frames, contiguous temporal windows, different numbers of frames, and different sampling intervals?\n\n3. Can the authors compare MCCL against a standard multi-view contrastive objective and a temporal transformer under matched parameter and compute budgets?\n\n4. Are protein sequences, ligand scaffolds, or near-duplicate complexes shared across the reported train and test splits?
6 comments
AuthorsOpenReview ↗

Response to Reviewer o727 (Part 1)

We sincerely thank the reviewer for the constructive feedback. We appreciate the recognition of DrugDyn's generalizability across different backbone architectures and evaluation settings, as well as the motivation behind our approach and the supporting ablation studies. Below, we carefully address the reviewer's concerns and provide further clarifications and additional evidence to strengthen our work.

In response to Weaknesses 1

We understand that spatial attention, temporal attention, and contrastive learning have been explored in previous studies within the broader fields of molecular representation learning and drug discovery. However, DrugDyn distinguishes itself through fine-grained modeling of geometry-aware protein–ligand interactions at the atomic level, together with robust representation learning of conformational ensembles. We further clarify the differences from existing methods below.

1. Distinction from existing spatial and temporal learning.

The most closely related work is TimeNucy and VideoNucy [1], which also investigate spatiotemporal learning for protein–ligand binding affinity prediction. Specifically, they first employ 3D convolutional neural networks (CNNs) to extract spatial features from MD conformations and then use LSTM or ConvLSTM architectures to capture temporal information across conformations. However, the authors explicitly report that these spatiotemporal models did not consistently outperform simpler MD-based data augmentation approaches. In contrast, our GASA differs in both spatial and temporal modeling. For spatial modeling, we introduce geometry-aware cross-molecular attention that explicitly emphasizes protein–ligand interactions and incorporates interatomic distances between them as bias terms in the attention weight computation. For temporal modeling, rather than relying on recurrent modeling of aggregated conformation-level features, we perform atom-wise attention across conformations to explicitly model variations in the local interaction environment of each atom.

We further investigate representative temporal transformer architectures in video understanding, including ViViT [2] and TimeSformer [3]. Specifically, ViViT employs a factorized encoder that first extracts frame-level representations through a spatial Transformer and then applies temporal self-attention to model dependencies across frames. TimeSformer adopts divided space-time attention, where each image patch attends to patches at the same spatial location across frames to capture temporal variations. This is conceptually related to the standard multi-head attention mechanism considered in our GASA w/MHA ablation.

However, the same spatial location across video frames does not necessarily correspond to the same physical object, particularly when objects move. In contrast, molecular graphs provide consistent atom identities across conformations, allowing our GASA to explicitly track each atom and model its variations through atom-wise attention. Furthermore, unlike video understanding, protein–ligand binding affinity prediction requires modeling interactions between two distinct molecular entities. To address this task-specific characteristic, we introduce sparse geometry-aware cross-molecular attention that explicitly focuses on spatially neighboring ligand–protein atom pairs and incorporates their interatomic distances as attention biases. These designs enable more targeted modeling of binding-relevant interactions, rather than directly applying generic spatiotemporal attention architectures.

[1] Libouban, Pierre-Yves, et al. "Spatio-temporal learning from molecular dynamics simulations for protein–ligand binding affinity prediction." Bioinformatics 41.8 (2025): btaf429.

[2] Arnab, Anurag, et al. "Vivit: A video vision transformer." 2021 IEEE/CVF international conference on computer vision (ICCV). IEEE, 2021.

[3] Bertasius, Gedas, Heng Wang, and Lorenzo Torresani. "Is Space-Time Attention All You Need for Video Understanding?" International Conference on Machine Learning (ICML), 2021.

AuthorsOpenReview ↗

Response to Reviewer o727 (Part 2)

In response to Weaknesses 1 (Continued)

2. Distinction from existing multi-view and contrastive representation learning.

We acknowledge that multi-view contrastive learning has also been widely studied. For example, GraphCL [1] constructs different graph augmentations as contrastive views through node dropping, edge perturbation and attribute masking, while GraphMVP [2] aligns 2D molecular graphs with their 3D geometric representations. Although TS2Vec [3] also describes itself as a multi-view contrastive learning framework, it constructs multiple views by progressively downsampling CNN-extracted temporal features and performing contrastive learning at different resolutions.

Our MCCL differs from these methods, as it is motivated by the physical nature of protein–ligand conformational ensembles. We aim for any finite subset of MD conformations, despite providing only a partial observation, to faithfully reflect the underlying conformational ensemble. Accordingly, rather than treating individual conformations or conventional graph augmentations as views, MCCL constructs subsets containing different numbers of conformations as multiple observations of the same complex. It explicitly encourages consistency across these different sampling scales while maintaining discriminability between distinct protein–ligand complexes.

We hope these clarifications help highlight the key methodological distinctions between DrugDyn and existing work.

[1] You, Yuning, et al. "Graph contrastive learning with augmentations." Advances in neural information processing systems 33 (2020): 5812-5823.

[2] Liu, Shengchao, et al. "Pre-Training Molecular Graph Representation with 3D Geometry." International Conference on Learning Representations (ICLR) (2022).

[3] Yue, Zhihan, et al. "Ts2vec: Towards universal representation of time series." Proceedings of the AAAI conference on artificial intelligence. Vol. 36. No. 8. 2022.


In response to Weaknesses 2

We agree that different conformations may exhibit distinct binding states, which is precisely why we introduce MCCL rather than treating individual conformations as separate contrastive views. Specifically, MCCL does not directly enforce consistency between distinct individual conformations. Instead, it constructs contrastive views by aggregating different subsets of conformations and encourages consistency across subset-level representations of the same underlying ensemble while maintaining discriminability between different complexes. Our ablation results in Table 3 of the manuscript further support this design, where MCCL consistently outperforms Single-Conformation Contrastive Learning (SCCL) across all evaluated backbones.

To address the reviewer's concern regarding the sensitivity of DrugDyn to frame selection, temporal ordering, frame count, and sampling intervals, we conducted comprehensive additional experiments. Please refer to our detailed responses to Questions 1 and 2 below.

AuthorsOpenReview ↗

Response to Reviewer o727 (Part 3)

In response to Question 1

To investigate the sensitivity of DrugDyn to temporal ordering, we conducted additional experiments by randomly permuting the order of sampled conformations while keeping the other experimental settings unchanged. The results are summarized below.

BackboneDatasetMethodRMSE↓MAE↓RR↑ρ\rho↑τ\tau↑
GIGNv2013Baseline1.7601.3760.6460.6470.464
DrugDyn1.4581.1330.7770.8050.605
DrugDyn (Shuffled)1.3851.0880.7850.7830.590
GIGNv2016Baseline1.5721.2090.6990.6860.506
DrugDyn1.3301.0400.7980.8030.604
DrugDyn (Shuffled)1.3671.0720.7860.7820.589
DTIGNv2013Baseline1.6261.3480.7060.6930.507
DrugDyn1.5351.2230.7510.7530.559
DrugDyn (Shuffled)1.5061.1540.7390.7460.565
DTIGNv2016Baseline1.5231.2310.7230.7050.515
DrugDyn1.3301.0700.8020.7900.597
DrugDyn (Shuffled)1.3441.0420.7870.7950.600

As shown above, randomly permuting the sampled conformations yields performance comparable to the original DrugDyn across both backbones and datasets, with minor fluctuations across evaluation metrics. These results suggest that strict temporal ordering is not essential to the effectiveness of DrugDyn.

Following MDDTA [1], as explicitly stated in Section 4.1 (Experimental Setup), we adopt its conformation selection strategy. Specifically, MDDTA clusters MD frames based on their three-dimensional atomic coordinates and selects a representative conformation from each cluster to capture substantial structural variations rather than uniformly sampling frames along the MD trajectory. Consequently, DrugDyn primarily focuses on learning ensemble-level structural representations and geometric relationships across conformations, making the model less sensitive to the chronological order of sampled conformations. We will incorporate these additional results and corresponding discussions into the revised manuscript.

[1] Zhao, Long, et al. "MDDTA: A Drug Target Binding Affinity Prediction Method Based on Molecular Dynamics Simulation Data Enhancement." IEEE Journal of Biomedical and Health Informatics (2026).

Comments on the paper

2 comments
AuthorsOpenReview ↗

A Gentle Follow-Up

Dear Reviewers,

We sincerely appreciate your careful evaluation of our work and valuable suggestions. Guided by your feedback, we have conducted additional experiments (presented in our individual replies), provided further clarifications, and revised the manuscript accordingly.

We hope that these additional results and revisions have addressed your concerns. We would be grateful if you could reconsider your ratings in light of our responses and the updated manuscript. We also warmly welcome any further questions or discussions and would be happy to provide additional clarifications. Thank you again for your time and thoughtful consideration.

AuthorsOpenReview ↗

Summary of Discussion

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair for handling our paper and all three reviewers for their thoughtful assessments and constructive feedback. We appreciate their positive recognition of the clear presentation, extensive evaluation, atom-level modeling of cross-conformation interactions, and the motivation behind Multi-Cardinality Contrastive Learning (MCCL).

We are grateful that Reviewer o727 gave our work an initial score of 3, Reviewer t9wE increased the score from 3 to 4 after our rebuttal, and Reviewer AzLM indicated that the score may be raised to 3 following our clarification.

We have carefully addressed the reviewers' concerns through detailed responses, additional experiments, and manuscript revisions. In particular, we strengthened the evaluation of conformational sampling, extended the ablation studies, provided biological interpretations, and clarified important methodological and implementation details.

We believe these additional results and clarifications substantially strengthen our work and provide a more comprehensive understanding of DrugDyn's effectiveness, robustness, and biological relevance. We sincerely appreciate the reviewers' valuable feedback in improving the quality and completeness of our paper.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.