Learning Protein-Protein Interaction Interfaces from Multimodal Representations

Abstract

Background: Protein-protein interactions govern essential cellular processes, making the identification of interface residues critical for protein engineering and molecular design. Although complex deep learning algorithms like AlphaFold3 generate full atom resolution assemblies, they are not optimized for residue-level interface identification. Existing methods face inherent limitations: sequence-based models lack explicit 3D context, while structure-based approaches can overlook evolutionary signals. Results: By combining evolutionary context with 3D geometry, the multimodal protein language model ESM3 offers the ideal foundation for residue-level interface classification. This results in the ESM3-PPISites model, which outperforms existing state-of-the-art methods across diverse training and benchmark sets. Generalization is assessed through stringent sequence- and structure-based redundancy filtering to ensure evaluation on previously unseen proteins. Notably, finetuning the open-weight ESM3 architecture meaningfully narrows the performance gap with its proprietary full-scale counterpart. Furthermore, integrating predicted interaction sites as spatial restraints into HADDOCK significantly improves the recovery of near-native binding poses while reducing computational search time by an order of magnitude compared with blind docking. Conclusions: ESM3-PPISites enables accurate interfaces identification without relying on complex structural inputs at inference time, establishing a reliable and computationally efficient strategy for mapping protein-protein interaction sites.

CHORUS: A proteome-scale atlas of membrane interaction interfaces constructed using cross-protein language model embeddings from sequences alone
J. Zhao, S. Huang, J. Sun
bioRxiv4 Oct 2026PreprintCC BY-NC-NDFull textBioinformatics & computational biologybioRxiv
Towards a Unified Framework for Protein Interaction Site Prediction Using Hierarchical Protein Representations
V. T. D. Nguyen, J. K. Johnstone, T.-S. Hy
bioRxiv18 Oct 2026PreprintCC BYFull textBioinformatics & computational biologybioRxiv
Learning Dynamic Neural Evidence Representations for Time-Adaptive Brain-Computer Interfaces
Beining Cao, Ziyi Zhao, Xiaowei Jiang +5
arXiv1 Sep 2026PreprintCC BY 4.0Full textNeurosciencearXiv
CUBE: Multimodal Representation Learning Reveals Biological Structure Across Histomorphology, Spatial Protein Phenotypes, and Transcriptome-Associated Signals
Z. Ge, H. Cai
bioRxiv14 Sep 2026PreprintAll rights reservedFull textBioinformatics & computational biologybioRxiv
Multimodal Protein Retrieval via Joint Representation Learning from Sequences and Cryo-EM Density Maps
A. Tulsani, A. Maddur Guruprakash, S. S. Prasad
bioRxiv2 Sep 2026PreprintCC BY-NC-NDFull textSystems & synthetic biologybioRxiv

Cite

BibTeX RIS

BibTeX

@misc{gardinazzi2026learning,
  title = {{Learning Protein-Protein Interaction Interfaces from Multimodal Representations}},
  author = {Y. Gardinazzi and E. N. Villegas Garcia and S. Senci and D. Di Vora and A. Feltrin and F. Cuturello},
  howpublished = {bioRxiv},
  year = {2026},
  month = may,
  doi = {10.64898/2026.05.29.728739},
  url = {https://www.biorxiv.org/content/10.64898/2026.05.29.728739v3},
  note = {CC BY},
}

RIS

TY  - UNPB
TI  - Learning Protein-Protein Interaction Interfaces from Multimodal Representations
AU  - Y. Gardinazzi
AU  - E. N. Villegas Garcia
AU  - S. Senci
AU  - D. Di Vora
AU  - A. Feltrin
AU  - F. Cuturello
PY  - 2026
DA  - 2026/05/29
PB  - bioRxiv
DO  - 10.64898/2026.05.29.728739
UR  - https://www.biorxiv.org/content/10.64898/2026.05.29.728739v3
AB  - Background: Protein-protein interactions govern essential cellular processes, making the identification of interface residues critical for protein engineering and molecular design. Although complex deep learning algorithms like AlphaFold3 generate full atom resolution assemblies, they are not optimized for residue-level interface identification. Existing methods face inherent limitations: sequence-based models lack explicit 3D context, while structure-based approaches can overlook evolutionary signals. Results: By combining evolutionary context with 3D geometry, the multimodal protein language model ESM3 offers the ideal foundation for residue-level interface classification. This results in the ESM3-PPISites model, which outperforms existing state-of-the-art methods across diverse training and benchmark sets. Generalization is assessed through stringent sequence- and structure-based redundancy filtering to ensure evaluation on previously unseen proteins. Notably, finetuning the open-weight ESM3 architecture meaningfully narrows the performance gap with its proprietary full-scale counterpart. Furthermore, integrating predicted interaction sites as spatial restraints into HADDOCK significantly improves the recovery of near-native binding poses while reducing computational search time by an order of magnitude compared with blind docking. Conclusions: ESM3-PPISites enables accurate interfaces identification without relying on complex structural inputs at inference time, establishing a reliable and computationally efficient strategy for mapping protein-protein interaction sites.
ER  -

Abstract

Related

Cite