High-throughput prediction of protein-protein interactions uncovers hidden molecular networks in biosynthetic gene clusters

Biosynthetic gene clusters (BGCs) are contiguous genomic regions that encode diverse proteins responsible for natural product biosynthesis. These proteins collectively produce various secondary metabolites with complex chemical structure, including antibiotics and mycotoxins, yet the complete biosynthetic pathways have been experimentally resolved for only a limited number of compounds. Protein-protein interactions within BGCs have recently been recognized as key determinants of intermediate transfer, enzymatic regulation, and structural stability. However, many BGCs still contain proteins of unknown function that cannot be predicted by conventional sequence-based bioinformatics tools, hindering a comprehensive understanding of their biosynthetic pathways. To address this challenge, we built a high-throughput complex prediction pipeline by replacing AlphaFold3s multiple sequence alignment generation with a faster MMSeqs2. We systematically screened 487,828 protein pairs derived from 2,437 BGCs registered in the Minimum Information about a Biosynthetic Gene cluster (MIBiG) database and predicted 15,438 heteromeric interactions with an ipTM [≥] 0.6. Among them, 1,390 protein pairs exhibited structural homology with an RMSD [≤] 2.0 [A]. These predictions highlight interesting molecular mechanisms involving proteins previously annotated as "uncharacterized" or "potentially dysfunctional". Our analysis further showed that the ipSAE metric can distinguish correct heterocomplex pairs when multiple functionally homologous proteins are present within a BGC. Overall, our computational analysis revealed molecular interaction networks among proteins encoded by each BGC and identified enzyme complexes that are likely functional only when assembled. These predicted complexes may represent previously unrecognized links in their biosynthetic pathways. The complete results are available in a reusable format at https://doi.org/10.5281/zenodo.17451667 to support future experimental validation.