Expanding the scope of protein language modeling to protein-protein interactions with MSA Pairformer.

Protein-protein interactions underlie biological complexity, and modeling their coevolution is essential for characterizing and engineering molecular assemblies. While protein and genomic language models have excelled at modeling individual proteins, extending these capabilities to protein complexes remains challenging. We present multiple sequence alignment (MSA) Pairformer, a protein language model that builds on AlphaFold2/3's bidirectional refinement between sequence and pairwise residue representations to accurately model the evolution of protein-protein interactions, despite training exclusively on individual chains. MSA Pairformer achieves nearly 3-fold improvement over existing methods in predicting protein-protein interface contacts and better distinguishes binding from non-binding sequences. A learned attention mechanism selectively weights sequences by their inferred evolutionary relevance, enabling discovery of subfamily-specific contacts. On single-protein benchmarks, it achieves state-of-the-art contact prediction and strong variant effect prediction using only 111 million parameters, over two orders of magnitude smaller than frontier models. These results offer an evolutionarily grounded, computationally efficient alternative to the scaling paradigm.