acceptodds
Under review as a conference paper at ICLR 2027

Distodyn: Using Distograms to Learn Dynamic Protein Representations at Scale

Abstract

Protein function and other biological properties often depend on structural dynamics, yet most machine-learning predictors rely on static representations. Physics-based molecular simulations can describe conformational variability but remain computationally prohibitive at scale. Generative models provide a more efficient alternative, though their ability to produce accurate conformational ensembles is still limited. In this work, we bypass expensive simulations by leveraging residue–residue distance probability distributions (distograms) from structure predictors such as AlphaFold2. Our approach provides a scalable way to encode dynamic information into protein representations, aiming to improve function prediction without explicit conformational sampling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Representation learning, Protein structure, Dynamics

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_DLgYOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

DistoDyn proposes to use pairwise distance distributions produced by structure predictors as cheap alternative to explicit conformational sampling. The distogram is used either to construct additional graph edges or directly as edge features. Then relational GNNs are used to combine distance edges, with these distogram-derived relationships.

Critical strengths
  1. The paper is dealing with how to incorporate conformational info into scalable protein representation without requiring MD trajectories. The motivation & idea of reusing info from structure predictor is appealing.
  2. The evaluation is complete, covering residue classification, affinity, mutational stability, PPI, RNA tasks, etc. Some of the effect sizes are large enough and I believe there is a real and useful signal in those distograms.
Critical weaknesses
  1. My major concern is on the interpretation of those signals in the distograms. A structure predictor's distogram is not, in general, a Boltzmann distribution over physical conformation (of course, the authors may argue that can be learned by large corpus of training). It can contain different confounding factors including but not limited to aleatoric structural variability, evolution information, training-set priors, etc.. I think there could be some more experiments to directly demonstrate the claim on "dynamics" (especially, distinguish "dynamics" from a very strong learned soft-contact representation). Additionally, I guess there are architecture/capacity confounds, particularly in the PPI experiment. Adding a relation to an R-GNN changes more than the graph: it adds relation-specific parameters. Thus a distance-only model and a distance+distogram model are generally not parameter matched.

Basically, I hold the positive attitude towards this work. Additional illustrations are welcome from the authors.

Questions for authors
  1. Can the authors distinguish “dynamics” from a very strong learned soft-contact representation?
  2. (Minor) Does the quoted 30-second figure include MSA search, preprocessing, model initialization, and full inference? If not, the comparison is not end-to-end.
1 comment

We thank the reviewer for their comments and their positive appreciation of our work.

> W1 My major concern is on the interpretation of those signals in the distograms. A structure predictor's distogram is not, in general, a Boltzmann distribution over physical conformation (of course, the authors may argue that can be learned by large corpus of training). It can contain different confounding factors including but not limited to aleatoric structural variability, evolution information, training-set priors, etc.. I think there could be some more experiments to directly demonstrate the claim on "dynamics" (especially, distinguish "dynamics" from a very strong learned soft-contact representation)

We thank the reviewer for raising this important point. We believe that strong evidence that distograms encode dynamic information was provided in previous publications by several research groups. For example, in 10.1038/s41467-026-69172-y the authors used distograms to generate conformational ensembles of intrinsically disordered proteins, whose accuracy was validated against a large set of experimental data. A similar approach was proposed in 10.1038/s41467-025-56572-9. More recently, in 10.1002/qub2.70062, the authors demonstrated that multi-modal distograms encode information about protein alternative states, well beyond local fluctuations around the ground state.

All these works provide a solid foundation of our interpretation of distograms capturing protein dynamic information. We will include a section in the related work dedicated to the established links between distograms and dynamic.

Beside the literature cited above, the comparison of the performances obtained in our tasks where Molecular Dynamics data is available, and specifically the comparison between distances+distograms vs distances+distograms+MD data, indicates that MD is not providing additional information compared to distograms. This point, along with the references cited above, indicates that most of the predictive improvement is attributable to dynamics. We will put forward this important comparison by clarifying the paragraph around L316 and adding an explicit reference to Table S4.

>W2 Additionally, I guess there are architecture/capacity confounds, particularly in the PPI experiment. Adding a relation to an R-GNN changes more than the graph: it adds relation-specific parameters. Thus a distance-only model and a distance+distogram model are generally not parameter matched.

In the PPI experiment, our baseline is external and represents the current SOTA for this task. SpatialPPIv2 was retrained from scratch on our dataset in order to compare its performance with our approach. For this reason, we didn’t alter the architecture and number of parameters of the model.

We carried out additional experiments on both the binding-site prediction and binding-affinity prediction tasks, matching the number of model parameters and the average number of edges per graph across graph constructions by adjusting the distance and distogram thresholds. The results show consistent improvements with the distance + distogram construction across both tasks and architectures, with the exception of a slight decrease in accuracy for EGNN on the binding-site prediction task. This suggests that the observed imp

Affinity prediction task:

ModelGraph typeParametersDistance thresholdMean edges per graphRMSE ↓MAE ↓Pearson r ↑Spearman ρ ↑
REGNNDistance + Distogram270,0810.017,437.051.6741.2530.6530.703
REGNNDistance270,72157,754.791.7451.4020.6190.604
RGATDistance + Distogram79,6810.017,437.051.5151.2350.7260.707
RGATDistance77,95357,754.791.6061.2800.6880.671

Binding-site prediction task:

ModelGraph typeParametersThresholdMean edges per graphAccuracy ↑Precision ↑Recall ↑F1 ↑
RGATDistance + Distogram7,0100.000154,925.810.8090.2780.5770.375
RGATDistance6,8182257,295.160.7960.2520.5330.342
REGNNDistance + Distogram255,8740.000154,925.810.8310.3310.6850.446
REGNNDistance256,5142257,295.160.8380.3020.4800.371

We hope to have addressed your questions, and thank you again for helping us improve our submission.

Reviewer_fSUrOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DistoDyn, a computationally efficient approach for incorporating information about protein conformational variability into graph-based protein representations without explicitly generating molecular dynamics (MD) trajectories. The central idea is to leverage residue–residue distance probability distributions (distograms) produced by structure prediction models such as Boltz-2. The authors construct additional graph edges based on predicted distance probabilities and optionally incorporate the full distogram distributions as edge features. These representations are integrated into relational graph neural networks, including R-GCN, R-GAT, and R-EGNN, to support downstream biological property prediction. Experiments on the MISATO dataset demonstrate improvements over conventional distance-based graphs and MD-correlation-based representations for binding-site and binding-affinity prediction. The method is further evaluated on protein stability prediction using ThermoMPNN, protein–protein interaction prediction using SpatialPPIv2, and RNA property prediction. Other results include improving binding-site prediction and protein–protein interaction. Overall, the paper demonstrates the potential of predicted distograms as scalable, information-rich features for biomolecular representation learning.

Critical strengths

The primary strength of this work is its practical and computationally efficient alternative to simulation-dependent representations, addressing an important limitation in incorporating protein dynamics into large-scale machine learning. Rather than requiring expensive MD trajectories, the proposed approach exploits existing outputs from structure predictors and provides a relatively straightforward mechanism for augmenting established graph representations.

Critical weaknesses

The main concern is that the paper does not conclusively establish that the observed improvements originate from physically meaningful protein dynamics, rather than additional structural information or prediction uncertainty encoded by Boltz-2. Predicted distograms are not necessarily calibrated equilibrium distance distributions, and the experiments do not directly validate their correspondence to experimentally observed conformational fluctuations. Although the rewiring experiments partially address this concern, the controls are not entirely conclusive.

Questions for authors
  1. Can the authors directly validate whether the predicted distograms capture biologically meaningful conformational fluctuations using experimentally characterized dynamic proteins or sufficiently sampled MD ensembles? How much of the predictive improvement is attributable to dynamics rather than structural uncertainty?

  2. How sensitive is the method to the probability threshold and the amino-acid-dependent distance cutoffs? Do the optimal thresholds transfer across protein sizes, downstream tasks, and structure prediction models?

  3. Would distograms generated by AlphaFold2 or other structure predictors provide comparable improvements? How robust is the proposed framework to differences in distogram calibration, structural prediction quality, and protein length?

5 comments

We thank the reviewer for their thoughtful comments and their positive appreciation of our work. We offer a point-by-point answer to their question below.

> Q1. Can the authors directly validate whether the predicted distograms capture biologically meaningful conformational fluctuations using experimentally characterized dynamic proteins or sufficiently sampled MD ensembles?

We thank the reviewer for raising this important point. We believe that strong evidence that distograms encode dynamic information was provided in previous publications by several research groups. For example, in 10.1038/s41467-026-69172-y the authors used distograms to generate conformational ensembles of intrinsically disordered proteins, whose accuracy was validated against a large set of experimental data and compared with extensive MD simulations as well as other state-of-the-art ensemble generators. A similar approach was proposed in 10.1038/s41467-025-56572-9. More recently, in 10.1002/qub2.70062, the authors demonstrated that multi-modal distograms encode information about protein alternative states, well beyond local fluctuations around the ground state.

All these works provide a solid foundation of our interpretation of distograms capturing protein dynamic information. We will include a section in the related work dedicated to the established links between distograms and dynamic.

> Q2. How much of the predictive improvement is attributable to dynamics rather than structural uncertainty?

This is also an excellent point. Beside the literature cited at the previous point, the comparison of the performances obtained in our tasks where Molecular Dynamics data is available, and specifically the comparison between distances+distograms vs distances+distograms+MD data, indicates that MD is not providing additional information/advantages compared to using only distograms. This point, along with the references cited above, indicates that most of the predictive improvement is attributable to dynamics. We will put forward this important comparison by clarifying the paragraph around L316 and adding an explicit reference to Table S4.

>Q3 How sensitive is the method to the probability threshold and the amino-acid-dependent distance cutoffs? Do the optimal thresholds transfer across protein sizes, downstream tasks, and structure prediction models?

We investigated the sensitivity of our method to the probability threshold used for graph construction and assessed the contribution of the amino-acid-dependent distance cutoffs. These experiments cover two downstream tasks (binding-site and binding-affinity prediction), using different graph neural network architectures (RGAT and REGNN) and two structure prediction models (Boltz-2, the primary structure predictor used in our study, and ESMFold2). All results are obtained on the distance + distogram model without distogram features.

Sensitivity to the probability threshold. Across experimental settings, low thresholds (0.0001–0.01) generally yielded the best performance, while higher thresholds (0.1–0.5) led to degradation. For binding-site prediction, 0.0001 achieved the best F1 across both architectures and structure prediction models. For binding-affinity prediction, the optimal threshold varied with the architecture and metric, but 0.0001 remained competitive. Overall, the method is robust across a range of low thresholds, although the precise optimum depends on the task and evaluation metric.

Contribution of amino-acid-dependent distance cutoffs. To assess the contribution of the amino-acid-dependent distance cutoffs used in our graph construction, we compared our approach with a simpler alternative based on a uniform hard distance cutoff of 8 Å, applied to all residue pairs regardless of their amino-acid identities. For the binding-site prediction task, the hard-cutoff configuration yielded slightly lower F1 scores than our standard approach. For binding-affinity prediction, the advantage of amino-acid-dependent distance cutoffs was more pronounced with RGAT, whereas REGNN achieved broadly comparable performance with the amino-acid-dependent and uniform 8 Å hard-cutoff strategies.

Robustness across architectures, structure prediction models, and downstream tasks. The results show that low probability thresholds consistently provide strong performance across the evaluated architectures and structure prediction models. In particular, the threshold of 0.0001 used throughout the main experiments performs best for binding-site prediction and remains competitive for binding-affinity prediction. This indicates that the method does not rely on a narrowly tuned probability threshold and that the use of a common low threshold provides a reasonable trade-off across tasks. However, the precise optimal value may vary depending on the downstream task and evaluation metric.

Binding affinity prediction task:

RGAT

DatasetRMSE ↓MAE ↓Pearson ↑Spearman ↑
boltz2_0.00011.5111.2260.7280.716
boltz2_0.0011.4921.1610.7370.723
boltz2_0.011.5141.2170.7310.718
boltz2_0.11.5861.2390.6930.674
boltz2_0.51.5981.2950.6890.661
boltz2_0.0001_hard_cutoff1.5411.2260.7180.704

REGNN

DatasetRMSE ↓MAE ↓Pearson ↑Spearman ↑
boltz2_0.00011.6311.2370.6740.715
boltz2_0.0011.6571.2600.6580.688
boltz2_0.011.5861.2400.6930.694
boltz2_0.11.5961.2700.6900.703
boltz2_0.51.6221.3180.6930.673
boltz2_0.0001_hard_cutoff1.5611.2570.7180.713

Binding-site prediction task:

RGAT:

InputAccuracyPrecisionRecallF1
boltz2_0.00010.8010.2700.5860.370
boltz2_0.0001_hard_cutoff0.7980.2660.5870.366
boltz2_0.0010.7990.2630.5650.359
boltz2_0.010.7850.2390.5320.330
boltz2_0.10.7790.2230.4940.307
boltz2_0.50.7300.1920.5370.283
esmfold2_0.00010.7930.2720.5190.356
esmfold2_0.0010.7770.2590.5450.351
esmfold2_0.010.7730.2480.5230.337
esmfold2_0.10.7620.2320.5010.317
esmfold2_0.50.7480.2160.4890.300

REGNN:

InputAccuracyPrecisionRecallF1
boltz2_0.00010.8480.3540.6410.456
boltz2_0.0001_hard_cutoff0.8720.3920.5310.451
boltz2_0.0010.8510.3490.5790.436
boltz2_0.010.8130.2980.6500.409
boltz2_0.10.8120.2850.5940.385
boltz2_0.50.8000.2790.6360.388
esmfold2_0.00010.8280.3500.6520.456
esmfold2_0.0010.7800.2960.7240.421
esmfold2_0.010.8030.3130.6590.424
esmfold2_0.10.7870.2910.6430.400
esmfold2_0.50.7880.2850.6120.389

Continuation of Q3

Threshold vs protein size

To investigate whether the optimal threshold transfers across protein sizes, we evaluated performance across five equal-count size bins for both binding site (F1 score) and affinity (MAE) prediction tasks.

For binding-site prediction, low thresholds (0.0001 and 0.001) consistently achieved the best F1 scores across protein sizes for both RGAT and REGNN. Performance generally decreased at higher thresholds, suggesting that the preference for low thresholds is robust across protein sizes.

For binding-affinity prediction, the optimal threshold varied more across size groups and architectures. Nevertheless, low thresholds (0.0001–0.01) generally remained competitive, while higher thresholds (0.1 and 0.5) did not provide consistent improvements.

Overall, these results suggest that low distogram probability thresholds transfer reasonably well across protein sizes, particularly for binding-site prediction, although the exact optimal threshold may vary depending on the downstream task and architecture.

Binding site:

InputRGAT <162RGAT 162–275RGAT 276–427RGAT 428–653RGAT ≥654REGNN <162REGNN 162–275REGNN 276–427REGNN 428–653REGNN ≥654
distance + distogram_0.00010.5420.4730.3750.3160.2240.5400.5260.4790.4280.304
distance + distogram_0.0010.5470.4630.3730.3120.2160.5490.5310.4720.4120.280
distance + distogram_0.010.4980.4380.3620.3070.2010.5310.5040.4310.3900.254
distance + distogram_0.10.4720.4080.3370.2810.1930.5190.5020.4160.3570.215
distance + distogram_0.50.4670.4040.3330.2690.1670.5430.4930.4090.3540.239

Binding Affinity:

InputRGAT <310RGAT 310–356RGAT 357–394RGAT 395–439RGAT ≥440REGNN <310REGNN 310–356REGNN 357–394REGNN 395–439REGNN ≥440
distance + distogram_0.00011.1911.0841.1431.4331.2841.3670.9051.1541.3441.410
distance + distogram_0.0011.0921.1221.1451.1351.3011.4301.1141.1581.2031.386
distance + distogram_0.011.2400.9491.0301.2731.5001.3750.9911.1631.3821.295
distance + distogram_0.11.2981.0081.2261.1791.4661.3810.9821.2091.2651.500
distance + distogram_0.51.2691.1841.0831.3011.5821.2510.9961.2421.3951.686
Reviewer_oyk7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes using distograms predicted by Boltz-2 as a cheap substitute for MD-derived dynamic information in graph-based protein representations, by adding distogram-based edges to relational GNNs. Distogram edges improve over MD-correlation edges on binding-site and affinity prediction, and improve over existing baselines on protein stability, protein-protein interaction and RNA tasks where MD data is not available.

Critical strengths
  • The approach is well motivated, replacing costly MD simulations with cheap distograms from structure predictors, which makes dynamic information practical to use at scale.
  • The paper is easy to read and follow, with clear language and no unnecessarily complicated terminology.
  • The evaluation is broad. It covers five tasks (binding site, affinity, stability, PPI, RNA) and three relational GNN architectures.
  • The controls are extensive. Experiments with random edges and topology-aware rewiring strengthen the claims.
Critical weaknesses
  • No variance or seed information is reported for the main results. The binding-site, affinity and PPI results (Tables 1, 2, 4 and S3 to S6) come without error bars, which makes it difficult to assess how robust the smaller gains are, including those over the random and rewired edge baselines. Reporting means and standard deviations over several seeds would strengthen the conclusions.
  • The claim that the best-performing method uses distograms “across all models, metrics and tasks” is slightly overstated. It is largely supported, but there are exceptions on accuracy for binding-site prediction (e.g. correlation edges for R-GCN in Table S3, random edges for R-GAT in Table S6). Given the small gains on several metrics and the absence of variance estimates, the claim should be softened.
  • The conceptual novelty is somewhat limited. The framework closely follows Guo et al. (2025), replacing MD-correlation edges with Boltz-2 distograms and using standard relational GNNs. The contribution is mainly empirical.
Questions for authors
  • How do results change for a distance graph with a larger radius, or a k-NN graph, chosen to have the same number of edges as the distogram graph?
  • Why are the effects of edge features so architecture-dependent? Section 5 states that “using distograms as explicit edge features yields mixed, architecture-dependent effects” (line 307). Is this mainly due to architectural differences or to the data?

Recommendations more than questions:

  • The motivation and contributions could be merged into the introduction, with the contributions as bullet points, to improve clarity and flow.
  • Figure 3D would benefit from a different color palette, since it is hard to see where the edge types overlap in the heatmap.
3 comments

We thank the reviewer for their thoughtful comments and their positive appreciation of our work. We offer a point-by-point answer to their question below.

> W1 No variance or seed information is reported for the main results. The binding-site, affinity and PPI results (Tables 1, 2, 4 and S3 to S6) come without error bars, which makes it difficult to assess how robust the smaller gains are, including those over the random and rewired edge baselines. Reporting means and standard deviations over several seeds would strengthen the conclusions.

We acknowledge the importance of multi-seed analysis and statistical analysis, which we strived to provide (RNA and thermostability tasks). We originally did not include them for the binding site/affinity prediction due to the large number of experimental setups, nor for the PINDER experiment due to the large size of this dataset. In response to your review, we have conducted additional experiments using five different random seeds for both the binding-site prediction and binding-affinity tasks. We now report the mean and standard error across the five runs, providing a better assessment of the variability and robustness of our results.

Affinity prediction task:

ModelGraph typeRMSE ↓MAE ↓Pearson rr ↑Spearman ρ\rho ↑
RGCNDistance1.634 ±\pm 0.0121.297 ±\pm 0.0140.671 ±\pm 0.0070.645 ±\pm 0.008
+ Correlation1.681 ±\pm 0.0171.346 ±\pm 0.0250.650 ±\pm 0.0070.619 ±\pm 0.008
+ Distogram1.627 ±\pm 0.0181.294 ±\pm 0.0130.675 ±\pm 0.0090.659 ±\pm 0.011
RGATDistance1.627 ±\pm 0.0261.273 ±\pm 0.0250.677 ±\pm 0.0120.657 ±\pm 0.008
+ Correlation1.694 ±\pm 0.0191.347 ±\pm 0.0190.645 ±\pm 0.0090.623 ±\pm 0.013
+ Distogram1.565 ±\pm 0.0261.249 ±\pm 0.0100.707 ±\pm 0.0130.690 ±\pm 0.013
+ Distogram + Features1.544 ±\pm 0.0251.225 ±\pm 0.0210.714 ±\pm 0.0110.697 ±\pm 0.010
EGNNDistance1.674 ±\pm 0.0151.345 ±\pm 0.0140.660 ±\pm 0.0080.635 ±\pm 0.010
+ Correlation1.798 ±\pm 0.0171.443 ±\pm 0.0130.577 ±\pm 0.0110.551 ±\pm 0.015
+ Distogram1.570 ±\pm 0.0251.258 ±\pm 0.0220.712 ±\pm 0.0100.705 ±\pm 0.013
+ Distogram + Features1.598 ±\pm 0.0381.295 ±\pm 0.0320.696 ±\pm 0.0160.681 ±\pm 0.018

Binding site prediction task:

ModelGraph typeAccuracy ↑Precision ↑Recall ↑F1 ↑
RGCNDistance0.742 ±\pm 0.0140.173 ±\pm 0.0050.420 ±\pm 0.0180.244 ±\pm 0.003
+ Correlation0.788 ±\pm 0.0030.201 ±\pm 0.0030.382 ±\pm 0.0090.263 ±\pm 0.003
+ Distogram0.787 ±\pm 0.0030.245 ±\pm 0.0020.550 ±\pm 0.0100.339 ±\pm 0.002
RGATDistance0.749 ±\pm 0.0040.187 ±\pm 0.0010.456 ±\pm 0.0100.265 ±\pm 0.002
+ Correlation0.769 ±\pm 0.0060.223 ±\pm 0.0020.530 ±\pm 0.0160.313 ±\pm 0.002
+ Distogram0.803 ±\pm 0.0040.278 ±\pm 0.0040.618 ±\pm 0.0090.384 ±\pm 0.002
+ Distogram + Features0.811 ±\pm 0.0040.295 ±\pm 0.0050.643 ±\pm 0.0070.404 ±\pm 0.003
EGNNDistance0.809 ±\pm 0.0070.267 ±\pm 0.0050.524 ±\pm 0.0200.352 ±\pm 0.003
+ Correlation0.837 ±\pm 0.0080.306 ±\pm 0.0080.502 ±\pm 0.0460.376 ±\pm 0.016
+ Distogram0.846 ±\pm 0.0040.347 ±\pm 0.0070.622 ±\pm 0.0140.445 ±\pm 0.004
+ Distogram + Features0.849 ±\pm 0.0090.356 ±\pm 0.0140.624 ±\pm 0.0220.451 ±\pm 0.006

These results strengthen our previous conclusions, with distograms outperforming other methods in 23/24 settings and a tie (we now win on all r-EGNN binding site metrics). These results also help better resolve the impact of edge features as detailed in answer to Q3 below.

Unfortunately, we were unable to perform the same analysis for the rewiring experiments and the PPI prediction task within the rebuttal period, primarily due to limited GPU availability. We nevertheless commit to conduct these additional experiments and include the corresponding statistics in the final version of the paper.

We would also like to point out that, for the PPI task, the performance gap between SpatialPPIV2 and our proposed model was relatively substantial. Therefore, although evaluating multiple seeds would provide a more comprehensive assessment of the variability, we do not expect that this analysis would substantially alter the conclusions regarding the comparative performance of the models. We agree that reporting variability across multiple seeds is important and appreciate the reviewer's suggestion, which has helped us strengthen the experimental evaluation as well as our results.

>W2 The claim that the best-performing method uses distograms “across all models, metrics and tasks” is slightly overstated. It is largely supported, but there are exceptions on accuracy for binding-site prediction (e.g. correlation edges for R-GCN in Table S3, random edges for R-GAT in Table S6). Given the small gains on several metrics and the absence of variance estimates, the claim should be softened.

Thanks for pointing this out, we will replace this statement with “across most..”. However, as mentioned above, our results are now more favorable to Distodyn.

>W3. The conceptual novelty is somewhat limited. The framework closely follows Guo et al. (2025), replacing MD-correlation edges with Boltz-2 distograms and using standard relational GNNs. The contribution is mainly empirical.

We agree that the idea of integrating protein dynamic information into graph representations is not novel and that our framework closely follows Guo et al. However, we think that the “simplicity” of our contribution can also be viewed as an advantage: MD simulations are computationally expensive at scale while our approach is extremely easy to deploy to obtain strong empirical results.

> Q1. How do results change for a distance graph with a larger radius, or a k-NN graph, chosen to have the same number of edges as the distogram graph?

To investigate whether the performance gains brought by distograms could simply be explained by an increase in the number of edges, we compared distance-only graphs with larger distance thresholds against distance + distogram graphs with comparable numbers of edges on the binding site prediction task.

We selected two pairs of configurations: a distance threshold of 16 Å compared with a distogram probability threshold of 0.001, and a distance threshold of 22 Å compared with a distogram probability threshold of 0.0001. These pairs yield similar average numbers of edges per graph, as reported in the following tables.

ModelGraph TypeThresholdAvg. # edgesAccuracyPrecisionRecallF1
REGNNDistance16 Å27,895.560.8630.3420.4150.375
REGNNDistance + distogram0.00128,018.940.8510.3490.5790.436
REGNNDistance22 Å57,295.160.8530.3440.5260.416
REGNNDistance + distogram0.000154,925.810.8480.3540.6410.456
RGATDistance16 Å27,895.560.7830.2250.4850.307
RGATDistance + distogram0.00128,018.940.7990.2630.5650.359
RGATDistance22 Å57,295.160.7810.2360.5390.329
RGATDistance + distogram0.000154,925.810.8010.2700.5860.370

For both architectures, the distance + distogram configurations consistently outperform their distance-only counterparts in F1, despite having comparable numbers of edges. The improvements are particularly noticeable in recall, suggesting that distogram information helps identify binding-site residues that are missed when connectivity is determined solely by a geometric distance threshold. Please note that the regnn/distance network have inflated accuracy by predicting more negative. This is due to label imbalance, as can be seen through their lower recall values. F1 represents a more meaningful readout as it is not sensitive to calibration.

These results indicate that the performance gains cannot be attributed solely to increased graph connectivity. Rather, they support the contribution of the additional structural information encoded by distogram features.

> Q3. Why are the effects of edge features so architecture-dependent? Section 5 states that “using distograms as explicit edge features yields mixed, architecture-dependent effects” (line 307). Is this mainly due to architectural differences or to the data?

After rerunning the experiments with 5 seeds for both the binding site and binding affinity prediction tasks, we have a clearer understanding of the impact of edge features.

  1. Features help the RGAT model, with clear gains for binding site prediction and smaller ones for binding affinity.

  2. Features help RGAT more than REGNN. This could be explained by the ability of the GAT to better make sense of non-geometric edge features through attention.

  3. Features help more on the binding site than on the binding affinity task. This can be explained by the reduced size training set of the latter (roughly 5000 complexes), which might not be enough to train a model on these more complicated inputs. On the thermostability and PPI prediction tasks (which contain hundreds of thousands of proteins), using distogram edge features yielded a clear performance boost.

We hope to have addressed your questions, and thank you again for helping us improve our submission.

Thank you for this detailed rebuttal and for running new experiments.

Regarding W1 (the variance). The five-seed results for the binding site and affinity experiments address and resolve my concern. The standard errors are small when compared to the reported gaps (or at least most of them), which make the gains credible. However, I would describe the gains as consistent rather than substantial, as several are within tow SEs. Compared to that gains on binding sites are larger and separate. I, of course, accept the GPU-availability explanation given the short time window.

W.r.t. W2, "Across most" is better, and it should be used consistently, including the two claims (italicised).

W3, novelty. I appreciate the open response and do fully agree that the "cheapness" and deployment ease are a legitimate contribution. However, my assessment stands that the contribution is mainly empirical, which is reflected in my rating.

Q1, the edge-matched control. This is a truly useful addition and it directly addresses my question.

Q3, edge features. The architectur- and data-size based explanation is plausible, but with three architectures, I think it should remain a hypothesis rather than a mechanism, I think in potential future revisions, it should be presented as such.

I will maintain my rating of 3, even though the rebuttal has strengthened the empirical case, and I would suggest the ACs take this under consideration. However, my reservation/concern regarding methodological novelty remains and given the empirical nature of the paper, a topical conference or journal may be the better fit. I leave this decision to the ACs.

Again, thank you for your thorough answers and further experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.