Decoding the Grammar of Protein-Protein Interaction Interfaces with Multimodal Representations

Protein-protein interactions (PPI) govern essential cellular processes, making the computational identification of interacting sites a central challenge in structural biology, with important implications for protein engineering and the development of targeted therapeutics. Existing prediction algorithms include sequence-based methods, which lack structural information, or structure-based approaches, which often struggle to effectively integrate evolutionary context. Here, we present ESM3-PPISites, a supervised model for residue-level classification of PPI interfaces, leveraging the multimodal representations of the ESM3 Protein Language Model. To ensure a bias-free evaluation, we adopt a stringent redundancy filtering protocol, systematically eliminating latent homology between the training data and a curated benchmark set in both sequence and structural space. Our findings demonstrate that while ESM3 largest proprietary version yields the highest predictive power, targeted fine-tuning of its small open-weight counterpart significantly narrows the performance gap. Requiring only primary sequence data at inference, ESM3-PPISites achieves unprecedented accuracy, vastly outperforming current approaches. Crucially, we demonstrate the practical impact of these predictions by integrating them as spatial restraints within the HADDOCK3 docking platform. When evaluated on an independent subset of 12 complexes from the Docking Benchmark v5, our prediction-guided pipeline strongly enhances the identification of near-native binding poses over ab initio blind docking, while reducing computational runtime by an order of magnitude. This framework establishes a scalable paradigm for high-throughput structural interactomics.