Protein-protein interactions (PPIs) underlie virtually all biological processes. To screen for new PPIs in silico, we trained a classifier (KIRC) to score the interaction likelihood of all 200 million possible human protein pairs. The top 1.6 million KIRC pairs were then modeled using AlphaFold-Multimer (AF-M), and the likely functional relevance of each binary structure prediction was assessed using a second classifier called SPOC.

This pipeline identified ~16,000 high confidence PPIs (90% precision) or ~112,000 lower confidence PPIs (50% precision). This human "predictome" can be browsed here. We have also performed more focused in silico PPI screens, including an analysis of endosomal and genome maintenance proteins.

Coupled with experimental validation, in silico PPI screening has led to new mechanistic insights in DNA replication, DNA repair, ubiquitin signaling, and other pathways. The human predictome is expected to greatly advance our understanding of the PPIs that underlie cell physiology, laying the foundation for the development of new drugs to fight human disease.