Skip to main content
Back to timeline
bioRxivSource publication:

Membrane interfaces without complex structures: sequence-only cross-family language model ensemble ranks interface residues at 0.643

Synopsis

Predicting membrane-protein interaction interfaces has been limited by scarce experimental complexes and by sequence alignments that break down across hydrophobic bilayers. CHORUS combines three protein language model families, ESM-2, ESM-C and ProtT5, across layer depths and three downstream networks into a six-model ensemble that outputs per-residue interface probabilities from a single sequence, reaching a chain-macro average precision of 0.643 on 4,655 homology-isolated test chains, above the complete-coverage structure method ScanNet-MSA at 0.613. It also releases precomputed interface landscapes for 80,619 reviewed transmembrane proteins across 4,918 species.

Source-provided article image: CHORUS: A proteome-scale atlas of membrane interaction interfaces constructed using cross-protein language model embeddings from sequences alone
Fig. 1 · bioRxiv · Page 43

Interpretation

The physical interaction interface of a membrane protein can now be inferred from its amino-acid sequence alone, without 3D structures, multiple sequence alignments, or a nominated binding partner. Earlier interface prediction depended either on experimentally resolved complexes or on sequence alignments that degrade across hydrophobic transmembrane regions, while membrane proteins account for only about 3 to 5 percent of high-resolution PDB structures. On 4,655 independent test chains sharing less than 30 percent sequence identity with training data, the six-model ensemble reached a chain-macro average precision of 0.643, above the complete-coverage structure method ScanNet-MSA at 0.613, GPSite at 0.583, PeSTo at 0.581 and Seq-InSite at 0.556.

Accuracy is actually higher inside the lipid bilayer, indicating that the sequence representations capture the tight packing physics of transmembrane helices. In cytoplasmic and extracellular regions the two models were close, at 0.626 versus 0.621 and 0.632 versus 0.637. Within the intramembrane core, CHORUS reached a context-macro average precision of 0.746 against ScanNet-MSA's 0.645, with pooled average precision of 0.754 versus 0.623 and AUROC of 0.889 versus 0.814.

The ensemble keeps two best configurations from each of three language model families rather than scaling one model. Controls that varied training seeds, feature coordinates, within-family layer depth, duplicated prediction streams or downstream parameter count could not reproduce the accuracy. The six-model ensemble scored 0.643 on the test set, while seed variation gave 0.621, coordinate basis 0.618, within-family depth 0.588, identical stream replication 0.575 and a larger single model 0.557, with paired cluster-bootstrap intervals excluding zero.

On megadalton respiratory supercomplexes, sequence alone delineates inter-subunit contact boundaries. That evaluation supplies no spatial coordinates and no quaternary assembly context, with each chain sequence inferred independently. Across 12,337 residues of the porcine I2III2IV2 respirasome (PDB 8UGN, 140 chains), entity-macro average precision was 0.984 with AUROC 0.970, and entity-macro average precision stayed between 0.945 and 0.984 across six cryo-EM assemblies.

Perspective

The demonstrated scope is partner-agnostic, per-residue interface prediction for alpha-helical transmembrane proteins, evaluated on 4,655 homology-isolated test chains, 167 external membrane targets and six cryo-EM respiratory supercomplex assemblies. The atlas covers 80,619 reviewed UniProt transmembrane records, 70,247 non-redundant sequences, 4,918 species and more than 33.7 million residues, with web querying and structure mapping. Intended readers are experimental and computational researchers who need to prioritize candidate interface residues in membrane proteins without structures. The source proposes compressing multi-pLM ensembles and developing partner-conditioned representations as next steps.

The 5 angstrom heavy-atom cutoff is a geometric proxy for contact, not a thermodynamic contribution to binding, and partner-agnostic inference cannot capture dynamic conformational transitions or lipidic environmental shifts. Accuracy declines on polytopic architectures with seven or more transmembrane segments and on transient interfaces with soluble non-transmembrane partners, and querying three language models carries substantial computational overhead. Zero-shot transfer to soluble proteins is mixed, and the fixed calibration threshold fails to balance precision and recall on some soluble classes. Overlap between predicted interfaces and BioLiP small-molecule binding sites is close to random expectation, and enrichment at CATH FunFam conserved positions is modest overall, clearest in cytoplasmic regions.