HandShake Model Predicts Partner-Specific Protein Binding Interfaces From Sequences Alone

Researchers led by Nurit Haspel have developed HandShake, a sequence-only computational predictor designed for partner-specific protein-protein interface prediction. Identifying which residues of a protein form an interface when bound to a specific partner is a central challenge with significant implications for drug discovery and understanding protein structure and function. Existing computational methods have been constrained by small training datasets, inconsistent redundancy filtering, and reliance on three-dimensional structural information at test time. HandShake overcomes these constraints by operating strictly on sequence data while predicting which specific residues participate in binding interfaces.

To predict these partner-specific interfaces without structural inputs at inference, the HandShake architecture 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. Through this integrated computational design, the model detects both protein binding interfaces and pairwise contact matrices. The system processes interacting sequences directly, providing partner-specific predictions without requiring experimental or modelled structural coordinates at test time.

The model was trained on very large datasets of non-redundant protein-protein pairs derived from the PPInterface dataset, the most comprehensive structural protein-protein database to date. To evaluate interface prediction capabilities, the investigators tested the model against systematically filtered benchmarks established at four distinct redundancy thresholds ranging from 30 percent to 90 percent sequence identity. This systematic evaluation demonstrated that sequence redundancy inflates reported AUROC values by up to 0.079 and Matthews correlation coefficient (MCC) values by up to 0.145 on identical models, representing a substantial methodological confound across the field.

Even under the stringent evaluation setting of a 30 percent sequence redundancy threshold, HandShake demonstrated strong predictive performance. The sequence-only method achieved an AUROC of 0.811, an MCC of 0.367, and an F1 score of 0.45, exceeding the best published sequence-only results under this benchmarking convention. HandShake also achieved performance comparable to existing partner-specific prediction methods that rely on explicit three-dimensional structural information.

The authors note that the comprehensive training and evaluation dataset, alongside the systematic redundancy analysis, provides valuable insight into protein-protein interactions. Exposing how sequence redundancy inflates evaluation metrics helps clarify the capabilities and limitations of current interface detection methods. By demonstrating that effective partner-specific binding site prediction can be achieved from sequence data alone, the study advances computational tools for drug discovery and structural biology.

Subject of Research: Partner-specific protein-protein binding site prediction using sequence-only computational models

Article Title: Handshake: Partner-specific protein-protein binding site prediction at scale using ProstT5 and cross-chain attention

Article References: Haspel, N. (2026). Handshake: Partner-specific protein-protein binding site prediction at scale using ProstT5 and cross-chain attention. PLOS Computational Biology, 22(11), e1014899. https://doi.org/10.1371/journal.pcbi.1014899

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014899

Keywords: HandShake, protein-protein interactions, binding site prediction, ProstT5, cross-chain attention, Low-Rank Adaptation, contact supervision head, pairwise contact matrices, sequence redundancy, PPInterface dataset, protein language models, sequence-only prediction, computational biology, drug discovery, AUROC, Matthews correlation coefficient