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

MSA Pairformer predicts protein-protein interface contacts nearly 3 times more accurately than existing methods, despite training only on individual chains.

• Why it matters: Understanding protein-protein interactions is crucial for characterizing biological complexity and engineering molecular assemblies, but current models struggle to extend from individual proteins to complexes.
• What they did: The authors developed MSA Pairformer, a protein language model based on AlphaFold2/3's bidirectional refinement, trained on single chains, and incorporating an attention mechanism to weight sequences by evolutionary relevance.
• The result: MSA Pairformer achieves state-of-the-art contact prediction and distinguishes binding from non-binding sequences with only 111 million parameters, offering a computationally efficient alternative to larger models.