A paired sequence language model for protein-protein interaction modeling

Abstract

Understanding protein–protein interactions (PPIs) is crucial for deciphering cellular processes and guiding therapeutic discovery. While recent protein language models have advanced sequence-based protein representation, most are designed for individual chains and fail to capture inherent PPI patterns. Here, we introduce a Protein Pair Language Model (PPLM) that jointly encodes paired sequences, enabling direct learning of interaction-aware representations beyond what single-chain models can provide. Building on this foundation, we develop PPLM-PPI, PPLM-Affinity, and PPLM-Contact for binary interaction, binding affinity, and interface contact prediction. Large-scale experiments show that PPLM-PPI achieves state-of-the-art performance across different species on binary interaction prediction, while PPLM-Affinity outperforms both ESM2 and structure-based methods on binding affinity modeling, particularly on challenging cases including antibody–antigen and TCR–pMHC complexes. PPLM-Contact further surpasses existing contact predictors on inter-protein contact prediction and interface residue recognition, including those deduced from cutting-edge complex structure predictions. Together, these results highlight the potential of co-represented language models to advance computational modeling of PPIs.