<i>Alphafuser</i> : a parsimonious approach to predicting higher-order protein complexes

Many biological processes rely on the activity of multiple proteins acting in concert as part of higher-order oligomeric complexes. However, the majority of structural studies have routinely simplified these systems to examine only one or a few protein partners of a multi-component system. Artificial intelligence (AI) methods implemented in AlphaFold, RoseTTAFold, ESMFold and others have revolutionized our ability to not only predict the fold of individual proteins, but also to predict that of complex assemblies. A major bottleneck in exploiting AI-based protein fold prediction for multiprotein complexes, however, is identifying likely interacting partners from a given interactome. Here, we present a protein complex prediction pipeline based on AlphaFold, called Alphafuser, that combines experimental interaction data with systematic querying of possible combinations of protein partners in a computationally parsimonious manner. This versatile protein complex prediction pipeline creates a queue of all potential complexes, up to a user defined number of partners. Because a complete search of all permutations is not typically practical, we implemented a simple dead-end trimming algorithm from the dimer step onwards based on the interface probability template modeling (ipTM) score to remove low-probability subcomplexes from the queue of higher-order complexes. Experimentally structurally characterized multiprotein complexes in the Protein Data Bank, PP2A-B56γ1 holoenzyme-PME-1 complex, the human γ-secretase complex, the TFIIIC complex and the TRAPP I complex, were used to obtain a general ipTM cutoff parameter for the pipeline and to confirm that complexes without extensive direct contacts between all subunits could be identified. We applied Alphafuser to two test cases starting from yeast two-hybrid and co-immunoprecipitation/mass spectrometry (co-IP/MS) interaction data. These predictions were tested experimentally using pull-down assays, confirming direct interactions of the proteins identified computationally by the Alphafuser pipeline.