Protein–Protein Interface Residue Prediction by Integrating Protein Language Representations and Structural Graph Attention
- 1Shanxi Agricultural University , , ,
- 2Shanxi Institute of Biology Co., Ltd. , ,
Abstract
Accurate residue-level prediction of protein–protein interaction interfaces is important for elucidating molecular recognition mechanisms and guiding targeted drug design. However, interface residues are relatively sparse and governed by complex spatial dependencies, making their accurate identification challenging. In this study, we developed EvoStruct-GAT, a multimodal graph attention model based on the PDBbind v2020.R1 data set. To reduce information leakage arising from sequence homology, receptor and ligand sequences were clustered independently, and the resulting clusters were used to construct training, validation, and test sets with no overlap in either receptor or ligand clusters. EvoStruct-GAT represents the target protein as a residue-level graph and integrates ESM-2 protein language model representations, DSSP-derived structural descriptors, and distance-aware graph attention, together with a combined strategy for addressing class imbalance. Using a minimum cross-partner heavy-atom distance of less than 5 Å to define interface residues, the model achieved an AUROC of 0.8081, an AUPRC of 0.4475, and an F1 score of 0.4562 on the independent test set. Alternative labels based on 4.0 Å, 4.5 Å, and ΔSASA definitions produced comparable AUROC values, indicating that the residue-ranking capability was reasonably robust across the tested interface definitions. Ablation experiments identified ESM-2 representations as the principal contributor to performance, with DSSP information and inter-residue message passing providing additional value. Under a common evaluation protocol, EvoStruct-GAT outperformed simplified internal baselines and showed complementary performance characteristics in external comparisons with publicly available models. These results demonstrate that EvoStruct-GAT can effectively rank potential interface residues without requiring information about a known binding partner, thereby supporting candidate-site selection, experimental validation, and downstream structural and functional analyses.
Related
- Cross-tissue graph attention networks for semi-supervised gene expression prediction
- Pep2Net: an integrated peptide knowledge base and cross-attention model for multi-task prediction of peptides
- KAT-Net: a hyperbolic routing framework for joint lysine acylation site prediction with shared sequence–structure representations
- Structural heterogeneity of CTE Type I tau filaments with an interface-shifted polymorph
- Integrated web interface for pentatricopeptide repeat (PPR) motif discovery and target prediction in plants
BibTeX
@article{hou2026proteinprotein,
title = {{Protein–Protein Interface Residue Prediction by Integrating Protein Language Representations and Structural Graph Attention}},
author = {Zhen Hou and Wei He and Weinan Cao and Jiarui Li and Hongquan Li and Shuo Sun and Hua Yang},
journal = {ACS Omega},
year = {2026},
month = oct,
doi = {10.1021/acsomega.6c08819},
url = {https://pubs.acs.org/acsodf/article/11/41/64022/5446641/Protein-Protein-Interface-Residue-Prediction-by},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Protein–Protein Interface Residue Prediction by Integrating Protein Language Representations and Structural Graph Attention
AU - Zhen Hou
AU - Wei He
AU - Weinan Cao
AU - Jiarui Li
AU - Hongquan Li
AU - Shuo Sun
AU - Hua Yang
PY - 2026
DA - 2026/10/07
JO - ACS Omega
DO - 10.1021/acsomega.6c08819
UR - https://pubs.acs.org/acsodf/article/11/41/64022/5446641/Protein-Protein-Interface-Residue-Prediction-by
AB - Accurate residue-level prediction of protein–protein interaction interfaces is important for elucidating molecular recognition mechanisms and guiding targeted drug design. However, interface residues are relatively sparse and governed by complex spatial dependencies, making their accurate identification challenging. In this study, we developed EvoStruct-GAT, a multimodal graph attention model based on the PDBbind v2020.R1 data set. To reduce information leakage arising from sequence homology, receptor and ligand sequences were clustered independently, and the resulting clusters were used to construct training, validation, and test sets with no overlap in either receptor or ligand clusters. EvoStruct-GAT represents the target protein as a residue-level graph and integrates ESM-2 protein language model representations, DSSP-derived structural descriptors, and distance-aware graph attention, together with a combined strategy for addressing class imbalance. Using a minimum cross-partner heavy-atom distance of less than 5 Å to define interface residues, the model achieved an AUROC of 0.8081, an AUPRC of 0.4475, and an F1 score of 0.4562 on the independent test set. Alternative labels based on 4.0 Å, 4.5 Å, and ΔSASA definitions produced comparable AUROC values, indicating that the residue-ranking capability was reasonably robust across the tested interface definitions. Ablation experiments identified ESM-2 representations as the principal contributor to performance, with DSSP information and inter-residue message passing providing additional value. Under a common evaluation protocol, EvoStruct-GAT outperformed simplified internal baselines and showed complementary performance characteristics in external comparisons with publicly available models. These results demonstrate that EvoStruct-GAT can effectively rank potential interface residues without requiring information about a known binding partner, thereby supporting candidate-site selection, experimental validation, and downstream structural and functional analyses.
SN - 2470-1343
ER -