acceptodds
Under review as a conference paper at ICLR 2027

Dynamics-Driven Protein–Ligand Binding Affinity Prediction via Conformational Flexibility Modeling

Abstract

Protein-ligand binding affinity (PLBA) prediction has benefited substantially from structural information, yet available structures typically provide only static snapshots of the binding state, limiting the characterization of conformational and energetic evolution during molecular recognition. This limitation becomes particularly pronounced when binding spans multiple conformational states that reshape local interactions and energetics, leading to state-dependent variations in binding strength. To address this, we propose D-PLBA, a dynamics-driven framework that integrates dynamic pocket-ligand trajectory prediction into PLBA prediction. D-PLBA predicts motion trajectories by modeling pocket motion, ligand rigid-body displacement, and torsional deformation, while estimating the corresponding binding interaction energies. The resulting conformations and energies are then incorporated into the affinity predictor to jointly characterize the spatiotemporal evolution of binding structures and interactions. Extensive experiments demonstrate that D-PLBA substantially outperforms state-of-the-art methods across most evaluated metrics on all benchmarks, while reducing dynamic multi-conformation inference time to approximately 1% of that required by existing approaches. Our code and datasets can be found at: https://anonymous.4open.science/r/anonymous-703E.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein-ligand binding affinity prediction, Conformational flexibility modeling, Drug discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jBsPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes D-PLBA, a two-stage framework for protein-ligand binding affinity prediction. A trajectory model pretrained on MISATO autoregressively predicts pocket and ligand conformations together with interaction energies. The frozen model then supplies multiple conformations to an affinity predictor with energy-conditioned aggregation. The paper reports improvements on PDBbind, CSAR-HiQ, and OpenBind, together with efficient trajectory inference. The application is well motivated, but the evaluation contains an unresolved metric inconsistency, and the evidence does not adequately establish the claimed contribution of physically meaningful dynamics. Overall I recommend rejection because the reliability of the comparative evaluation and the claimed contribution of physical dynamics are not adequately established. The SD inconsistency directly affects confidence in the benchmark reporting; the RMSF interpretation, incomplete ablation controls, and qualitative energy analysis leave the main scientific claims insufficiently supported. Addressing these concerns requires reconciliation of the evaluation and focused controlled validation.

Critical strengths
  1. Incorporating local pocket-ligand flexibility into affinity prediction is a relevant direction. Pretraining a reusable trajectory model and freezing it for downstream learning is a practical design.
  2. The reported affinity improvements span several datasets, and the manuscript includes both component ablations and a frame-count sensitivity study. These experiments provide useful initial evidence for incorporating multiple conformations and energy information.
Critical weaknesses
  1. The reported SD values are inconsistent with the stated metric under its standard interpretation Equation (11) defines SD using residuals after fitting a regression line. If the intercept and slope are fitted by least squares to the evaluated predictions, the identity line is an admissible fit, giving SD <= sqrt(n/(n - 1)) * RMSE. For the stated 285-complex PDBbind test set, SIGN's RMSE of 1.306 therefore implies SD <= 1.3083, whereas Table 1 reports 1.382. The same problem occurs for IGN (RMSE 1.372; maximum SD 1.3744; reported SD 1.417) and MBP (RMSE 1.231; maximum SD 1.2332; reported SD 1.244). These discrepancies exceed rounding error and would persist when averaging matched runs. If a different fitting protocol or evaluation set was used, it needs to be specified and reconciled with Equation (11). This inconsistency concerns the reliability and comparability of the central benchmark results. The authors should provide the metric implementation and fitting protocol and recheck the table using consistently evaluated predictions.
  2. Larger RMSF does not establish better trajectory prediction Table 2 labels RMSF with an upward arrow and ranks larger values as better, while Section 4.2 uses high RMSF and low clash scores to support physically meaningful flexibility. However, Equation (12) measures fluctuations around the predicted trajectory's own mean; it does not measure agreement with reference fluctuations. Larger RMSF can reflect realistic flexibility, excessive motion, or accumulated drift. Low clash scores do not distinguish these possibilities. Although reference-based RMSD is also reported, it does not justify treating RMSF as a quantity to maximize. The relevant evaluation is agreement with reference MD RMSF profiles under matched alignment, atom selection, and sampling conditions, including fluctuation-amplitude errors. An initial-structure persistence baseline would also show how much the reported coordinate accuracy improves over retaining the starting structure. Without these comparisons, the RMSF ranking and the interpretation of increasing RMSF during rollout are not convincing evidence of accurate flexibility modeling.
  3. The ablations do not adequately isolate the source of the gains Figure 6(a-c) compares SS, LM, DS, SS+LM, and DS+LM, and explicitly labels DS+LM as the authors' method. However, Section 3.2 and Figure 2 do not specify an LM branch or its integration. It is therefore unclear whether the headline results in Tables 1 and 4 correspond to DS or DS+LM, and what is held fixed across the comparisons. An LM comparison is reasonable, but this undocumented component obscures the attribution of improvements to the proposed dynamics module. Figure 6(d,e) already varies the number of conformations from 1 to 9. At the fixed 1-ns sampling interval, however, this simultaneously changes the number of input frames, the rollout duration, and accumulated prediction error. It does not separate the benefit of additional conformations from that of temporal information. The key controls should hold the LM features, trajectory pretraining, downstream architecture, and training protocol fixed while comparing repeated static inputs with predicted DS inputs. Varying the number of retained DS frames over a fixed generated horizon, and shuffling the order of paired conformation-energy frames, would further test the claimed temporal contribution. Table 4 supports the usefulness of the added components, but does not establish that their gains require faithful temporal dynamics.
  4. Figure 5 and the trajectory analysis do not substantiate the claimed physical dynamics Figure 5 plots predictions and MD references against separate vertical scales within each panel, without labeling the energy quantity or units on those axes. Visual overlap on independently scaled axes cannot establish agreement in energy magnitude or fluctuation amplitude. Even for the narrower claim of similar temporal trends, the manuscript should quantify the claimed association under an explicit temporal sampling and normalization protocol. Appendix C acknowledges differences between the AMBER and GROMACS pipelines, but does not report a numerical value for the cited RMSD*_IE statistic. The energy transformation is also unclear: Appendix B describes a log-transformed evaluation scale while defining only the linear transformation IE = -IE_AMBER/10, and Appendix C again refers to a linear transformation. A shared calibrated scale, or a specified common normalization with quantitative trend statistics, is needed to interpret this experiment. The manuscript does assign 1 ns to each autoregressive step. Nevertheless, assigning a physical lag does not establish that the learned transitions reproduce time-dependent molecular behavior. Structural snapshots, coordinate RMSD, clash scores, and qualitative energy curves do not demonstrate correct temporal correlations or transition behavior. Validation of a relevant time-dependent observable, such as pocket-ligand contact or energy autocorrelation at the stated lag, would directly address this claim. As presented, the evidence for physically meaningful dynamics remains insufficient for the interpretation on which the paper's central contribution rests.
Questions for authors

. How were the regression parameters in Equation (11) fitted, and how do the authors reconcile the IGN, SIGN, and MBP SD values with their RMSE values? Were all metrics computed on the same predictions and test examples? 2. How closely do predicted RMSF profiles and amplitudes match reference MD under the same alignment and sampling protocol? How does trajectory RMSD compare with an initial-structure persistence baseline? 3. Do Tables 1 and 4 use DS or DS+LM, and what LM and fusion mechanism are used? With that configuration fixed, do the gains persist against repeated-static inputs and shuffled temporal order while keeping each conformation paired with its energy, and when frame count is separated from rollout duration? 4. What quantities, units, and transformations do Figure 5's two axes represent? Please report the definition and value of RMSD*_IE, quantify temporal-trend agreement under an explicit sampling and normalization protocol, and provide evidence that the generated sequence reproduces a relevant time-dependent MD

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

This paper presents D-PLBA, a protein–ligand binding affinity prediction method that incorporates predicted conformational dynamics. The model first predicts pocket and ligand motions together with interaction energies, then uses the resulting conformations and energy features for affinity prediction. The authors evaluate the method on PDBbind, CSAR-HiQ, OpenBind, and MISATO, reporting improvements in both affinity and trajectory prediction.

Critical strengths
  1. Incorporating conformational flexibility into affinity prediction is well motivated. I like the idea of focusing on local pocket–ligand motions rather than simulating the entire protein. The combination of trajectory prediction and energy-aware feature aggregation is also reasonable.

  2. D-PLBA improves RMSE from 1.218 to 1.151 on PDBbind and from 1.657 to 1.601 on CSAR-HiQ compared with strong baselines (Table 1). The trajectory prediction results are also impressive, particularly the reduction in inference time from 6.76 to 0.06 hours compared with ProAR (Table 2). The ablations suggest that both dynamic conformations and interaction-energy features contribute to performance.

Critical weaknesses
  1. My main concern is whether the generated trajectories capture meaningful protein–ligand dynamics or simply predict plausible conformations. The evaluation mainly uses RMSD, RMSF, and clash scores. These are useful, but they do not show whether the model reproduces the conformational distributions or interaction patterns observed in MD simulations. For example, comparing ligand torsional distributions or protein–ligand contact occupancies would provide stronger evidence. I also have some concerns about the interaction-energy validation. Figure 5 shows selected predicted and reference trajectories, but the two curves use different vertical scales, making the agreement difficult to assess. A quantitative comparison of energy correlations would be more convincing. In addition, the predicted interaction energies should not be confused with binding free energies.

  2. It is unclear whether the improved affinity prediction is useful for actual virtual screening. In practical screening, we are often more interested in ranking compounds correctly than predicting their absolute affinities. On PDBbind, D-PLBA reduces RMSE from 1.218 to 1.151, but Pearson R remains essentially unchanged (0.824 versus 0.825). This does not necessarily mean ranking is unchanged, but it raises the question of whether the improvement leads to better compound prioritization. The OpenBind experiment is particularly relevant because it involves multiple ligands against the same target. Although D-PLBA achieves a much lower RMSE, the paper does not report whether it ranks the strongest binders more accurately. I would like to see Spearman or Kendall correlations, pairwise ranking accuracy, and top-K recovery of high-affinity compounds. If only 10 or 20 compounds could be selected for experimental testing, would D-PLBA identify more strong binders than existing methods? The current ablations also do not fully establish whether temporal information is necessary. A comparison with multiple independently perturbed conformations or static energy features would help clarify whether the gains really come from modeling dynamics.

Questions for authors
  1. Can the authors provide additional evidence that the predicted trajectories reproduce meaningful conformational distributions, such as ligand torsions or protein–ligand contact occupancies? Could they also report quantitative correlations between predicted and MD-derived interaction energies?

  2. Does the improvement in RMSE translate into better compound ranking on OpenBind? In particular, how does D-PLBA compare with existing methods in Spearman correlation, pairwise ranking accuracy, and top-K recovery of high-affinity compounds?

  3. Can the ranking results be compared to AlphaFold3's iPTM ranking for a more comprehensive benchmarking?

  4. How much of the improvement comes from temporal dynamics rather than simply using additional conformations or energy features? Would similar gains be obtained using independently sampled conformations without temporal ordering?

Reviewer_dQa2OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes D-PLBA, a dynamics-aware framework for protein–ligand binding affinity prediction. A pretrained trajectory module predicts pocket motion, ligand rigid-body displacement, ligand torsion, and interaction energies. The predicted conformations and energy trajectories are then incorporated into a downstream affinity predictor. The method is evaluated on PDBbind, CSAR-HiQ, and OpenBind, together with trajectory-level metrics such as RMSD, RMSF, clashes, and interaction-energy error. The paper reports improvements over structure-based baselines and MD-based strategies, with substantially faster inference than conventional molecular dynamics simulation.

Critical strengths
  1. The paper addresses a real limitation of static protein–ligand representations and integrates conformational motion, pocket flexibility, and interaction-energy evolution into a coherent framework.\n\n2. The evaluation includes standard PDBbind benchmarks, cross-protein LBA splits, an external CSAR-HiQ benchmark, a single-target OpenBind setting, trajectory-quality metrics, and component ablations.
Critical weaknesses
  1. Binding affinity is an ensemble-level property, but the proposed model mainly produces a short, deterministic autoregressive trajectory. It is unclear whether the generated trajectory represents a meaningful conformational ensemble or primarily a learned average trajectory.\n\n2. The interaction-energy supervision and evaluation use different simulation packages and protocols. This weakens claims about physically meaningful absolute interaction-energy prediction. The data-isolation and sampling protocols also need more detail, including how the five frames are selected and whether PDBbind splits are independent at the protein and ligand levels.
Questions for authors
  1. What exact rule is used to select the five conformations from the 100 MISATO frames? How sensitive are results to random, contiguous, or different-number frame sampling?\n\n2. Are all downstream test proteins and ligands excluded from trajectory pretraining at the sequence and chemical-identity levels?\n\n3. How does the affinity model perform using real MD frames, predicted frames, and static frames under otherwise identical conditions?\n\n4. Can the authors report stability and uncertainty for longer or stochastic rollouts rather than only the first five predicted nanoseconds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.