henhaobio
Sign in

bioRxivBioinformaticsPreprint

Learning Dynamic Protein Representations at Scale with Distograms

Nicolas PortaliD, Wissam KarrouchaiD, Vincent MalletiD, Massimiliano BonomiiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

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 Boltz-2. Our approach provides a scalable way to encode dynamic information into protein representations, aiming to improve function prediction without explicit conformational sampling. All code required to reproduce the experiments presented in this work is publicly available at https://github.com/anonymous-research-world/distodyn.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (14 Oct 2026) · CC BY-NC-ND 4.0

Adding Boltz2 distogram edges and features to protein and RNA graph networks beats MD-correlation graphs on binding tasks and improves stability, interaction and RNA predictions without simulation.

Function predictors mostly use static structures, and encodings of protein motion that rely on molecular dynamics (MD) trajectories need costly simulations. The paper instead uses the residue-pair distance distributions that structure predictors output as byproducts (around 30 GPU seconds, versus several GPU hours for MD).

Key results

  1. ShownOn MISATO ligand binding site and binding affinity tasks, distogram-augmented relational graphs outperform graphs built from MD motion correlations, which need MD simulations.

    Distogram approach ranks first in 21/24 settings; when second it trails by 2 accuracy points (R-GCN); when first, gains reach 9.6 F1 points on binding site prediction. Using distograms as edge features has mixed, architecture-dependent effects but can help strongly (affinity with R-EGNN: +5 Spearman points). · Performance analysis; Experimental setup; Tables

  2. ShownThe gain is not an artifact of how MD information is encoded, nor of graph rewiring alone. Most of the information from short MD appears to be already contained in distograms.

    An empirical distogram from MD performs like correlation edges; correlation plus distogram edges performs like distograms alone. Random edges and topology-aware rewiring (1,000-10,000 loops) improve less than distogram edges in most cases, and the two effects are synergistic. · Performance analysis; Rewiring effects; B.2. Complementarity to Correlation; B.3. Additional rewiring results

  3. ShownDistogram edge features improve ThermoMPNN protein stability (ΔΔG) prediction on the mega-scale dataset and reach the SPURS baseline without architectural changes.

    R^2 rises from 0.524 to 0.567 and RMSE falls from 0.722 to 0.689; Pearson and Spearman also improve. Results are means over five seeds; only edge features were added because graphs are small (56 nodes on average). · 6.1. Prediction of mutation effects; Tables

  4. ShownDistograms improve protein-protein interaction prediction on larger graphs (PINDER heterodimers, over the SpatialPPIv2 baseline) and RNA property prediction (RNA-CM, RNA-Site), where distogram edge features also beat RBF-encoded distance features.

    PPI F1 rises from 0.732 to 0.815 and accuracy from 0.777 to 0.835. For RNA, distogram edge features beat RBF-encoded distance across distance, 2D+ and 2.5D graphs (balanced accuracy, three seeds; values are in Table 5, not in the text). · 6.2. Prediction of protein-protein interactions; 6.3. Prediction of RNA properties; Tables

  5. ShownGains from distogram features do not track Boltz2 prediction quality or MSA depth in the tests reported.

    On RNA tasks, gain in balanced accuracy (MCC) correlates with TM-score at Pearson 0.04 (-0.04). On 1610 MISATO binding-site systems, gain vs MSA depth has all |r| < 0.08; F1 gain for the 10% shallowest MSAs is comparable to the full test set. · 6.4. Robustness to Boltz2 Prediction Quality; B.4. Correlation with MSA depth

Methods, data and reproducibility

Residue graphs are enriched with edges between pairs whose Boltz2 distogram assigns enough probability to short distances (threshold 10^-4), plus optional distogram edge features. Relational GNNs (R-GCN, R-GAT, R-EGNN) are trained on distance, correlation and distogram relations, and ThermoMPNN and SpatialPPIv2 are modified.

  • MISATO protein-ligand complexes (binding site task, MISATO split; MD-derived correlations) · n = 19443 complexes
  • PDBbind 2020 refined set (train) and core set (test) for binding affinity · n = 5318 train and 285 test complexes
  • Mega-scale protein stability dataset, single-mutant ΔΔG for ThermoMPNN · n = 577313 samples after filtering; 298 wild types, 443,906 single-mutant sequences
  • PINDER heterodimers for protein-protein interaction prediction · n = 450000 train, 325 validation, 357 test pairs
  • RNAglib RNA-CM and RNA-Site tasks · n = 57 and 64 nodes on average per graph
Methods & tools
Boltz2, R-GCN, R-GAT, R-EGNN, ThermoMPNN, ProteinMPNN, SPURS, SpatialPPIv2, RNAglib, MMseqs, AdamW
Assay conditions & compute
Single 80GB A100 or 40GB A40 GPU. R-GAT needed over 40GB and took about 20 hours on binding site tasks; R-GCN and R-EGNN under 10 hours; ThermoMPNN fine-tuning 14 hours. Distograms take around 30 GPU seconds (Figure 1), compared with several GPU hours for MD.
Reproducibility gaps
Code is hosted at anonymized repository addresses; no software versions or environment files are given. Boltz run with default parameters.
Limitations (5)
  • Method depends on high-quality distograms, whose accuracy rests on coevolutionary signal in the MSA; it is uncertain whether informative features arise for systems with low-accuracy predictions. · stated by the authors
  • Generating distograms becomes costly for very large proteins and for multi-chain macromolecular assemblies. · stated by the authors
  • Distograms are marginal pairwise distance distributions, so they give an entangled description of variability and may miss higher-order multi-residue dynamic correlations. · stated by the authors
  • The authors' reproduction of the MD-correlation baseline gave stronger distance-only results than previously reported, and correlation edges helped binding site but not affinity prediction. · stated by the authors
  • Dropout, rewiring steps and learning-rate strategies were tuned per model and the best setups reported; the text does not describe a separate selection set for these choices. · noted in this summary

Read this if you build graph-based protein or RNA property predictors and want cheap dynamics-aware features from structure-predictor distograms instead of MD trajectories.

Citations
2

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTang '24Audagnotto '21Fan '25Kleiman '26Liu '25Xiang '26
68% semantic matchBiophysics
Yike Tang, Mendi Yu, Ganggang Bai+3 more·bioRxiv·Jun 27, 2024· 5 citations