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

HandShake, a sequence-only partner-specific protein-protein interface predictor, achieves AUROC 0.811 and MCC 0.367 at 30% redundancy, exceeding the best prior sequence-only benchmark.

• Why it matters: Partner-specific interface prediction is hindered by small training sets, inconsistent redundancy filtering, and test-time dependence on 3D structure, risking misleading performance estimates.
• What they did: The model combines ProstT5 with LoRA and cross-chain attention plus a contact supervision head, trained on large non-redundant PPInterface-derived pairs and evaluated across redundancy thresholds from 30% to 90%.
• The result: Sequence redundancy inflates AUROC by up to 0.079 and MCC by up to 0.145 on identical models, and HandShake’s results are comparable to structural-information-based partner-specific methods.

AI summary of the abstract · may contain errors · not medical advice · read the paper.