Handshake: Partner-Specific Protein-Protein Binding Site Prediction at Scale Using ProstT5 and Cross-Chain Attention

Abstract

Partner-specific protein-protein binding site prediction, identifying which residues of a protein form the interface when bound to a specific partner, remains a challenging task with significant implications for drug discovery and understanding of protein structure and function. Existing computational methods are limited by small training datasets, inconsistent redundancy filtering, and reliance on three-dimensional structural information at test time. Here we present a sequence-only, partner-specific protein-protein interface predictor called HandShake. It combines ProstT5, a protein language model pre-trained on structural data, with Low-Rank Adaptation (LoRA), a cross-chain attention mechanism and a contact supervision head. Our method can detect both binding interfaces and pairwise contact matrices. We trained our model on very large datasets of non-redundant protein-protein pairs derived from the PPInterface dataset, the most comprehensive structural protein-protein database to date, and evaluated it on systematically filtered benchmarks at four redundancy thresholds (30%-90% sequence identity). We demonstrate that sequence redundancy inflates reported AUROC by up to 0.079 and MCC by up to 0.145 on identical models, representing a substantial methodological confound in the field. Even at 30% redundancy threshold, our results (AUROC=0.811, MCC=0.367, F1=0.45) exceed the best published sequence-only result on this convention. Our method also achieves comparable performance to existing partner-specific methods that use explicit structural information. The comprehensive training and evaluation dataset, in addition to the systematic redundancy inflation, can help gain insight into protein-protein interactions and the abilities and limitations of current detection methods. Data availabilityThe code and data can be found at http://github.com/nurith/Handshake. Author summaryUnderstanding which residues of a protein make contact with a specific protein partner is fundamental for designing drugs and understanding cellular processes, but predicting these interfaces remains challenging. We developed a deep learning method that takes only the amino acid sequences of two proteins and predicts, for each residue, whether it lies on their binding interface. The method combines ProstT5, a protein language model trained to translate between sequence and structure, with a cross-chain attention mechanism that lets each proteins residues attend to its partner. We train and evaluate on up to 32,503 non-redundant protein pairs across four sequence redundancy thresholds. To the best of our knowledge, this is by far the largest dataset trained and tested for this specific task. Our key methodological finding: sequence redundancy in evaluation benchmarks inflates reported metrics by up to 0.079 AUROC and 0.145 MCC on identical models when measured internally on PPInterface. Inference experiments on four independent benchmarks show that this internal inflation does not transfer cleanly to externally-curated datasets, where switching from 30%-trained to 90%-trained models changes performance only modestly. At 30% filtering, our method achieves AUROC values from 0.784 to 0.828 across five datasets, confirming genuine generalization. Through systematic negative controls and comparisons with the ESM-2 family of language models at matched parameter counts, we show that explicit structural pre-training, not just model scale, is what enables sequence-only binding site prediction to work.