CHORUS: A proteome-scale atlas of membrane interaction interfaces constructed using cross-protein language model embeddings from sequences alone

Abstract Membrane proteins drive vital cellular signaling and bioenergetics yet resolving their physical interaction interfaces remains hindered by the paucity of experimental complexes and the breakdown of sequence alignments across hydrophobic bilayers. Here, we present CHORUS, a sequence-only deep-learning framework that decodes interaction sites across membrane proteins without requiring 3D structures, multiple sequence alignments (MSAs), or predefined partner identities. By systematically synthesizing complementary latent representations across diverse protein language model (pLM) families and functional layer depths, CHORUS outperforms 18 established sequence- and structure-based methods across 4,655 homology-isolated test contexts (average precision: 0.643), demonstrating amplified predictive capacity within the lipid-embedded core (0.746). We show that CHORUS delineates quaternary contact boundaries across megadalton respiratory supercomplexes directly from sequences alone (0.984). We deliver the CHORUS Atlas ( https://chorus.atlas.sunnylab.org ), providing precomputed, residue-resolved interaction landscapes for 80,619 reviewed transmembrane proteins spanning 4,918 species and 33.7 million residues to empower interactome-scale discovery and allosteric therapeutic design.