Alphafuser: a parsimonious approach to predicting higher-order protein complexes.

Area of Science:

• Structural Biology
• Computational Biology
• Biochemistry

Background:

• Biological processes often involve complex assemblies of multiple proteins.
• Previous protein structure studies simplified multi-protein systems.
• Artificial intelligence (AI) methods like AlphaFold have advanced protein fold prediction for individual proteins and complexes.

Purpose of the Study:

• To develop a computational pipeline, Alphafuser, for predicting multi-protein complexes.
• To integrate experimental interaction data with AI-based predictions for identifying protein partners.
• To address the bottleneck of identifying interacting partners from large interactomes for AI-driven complex prediction.

Main Methods:

• Developed Alphafuser, a protein complex prediction pipeline utilizing AlphaFold.
• Combined experimental interaction data with systematic querying of protein partner combinations.
• Implemented a dead-end trimming algorithm using the interface probability template modeling (ipTM) score to optimize predictions.
• Validated the pipeline using known multi-protein complexes from the Protein Data Bank.

Main Results:

• Alphafuser successfully predicted multi-protein complexes, including those without extensive direct contacts between all subunits.
• A general ipTM cutoff parameter was established using known complexes.
• Application of Alphafuser to yeast two-hybrid and co-immunoprecipitation/mass spectrometry data yielded testable predictions.
• Experimental validation confirmed direct interactions for computationally identified protein partners.

Conclusions:

• Alphafuser provides a computationally efficient method for predicting multi-protein complexes by integrating experimental data and AI.
• The pipeline is versatile and can handle varying numbers of protein partners.
• The study demonstrates the utility of Alphafuser in identifying functionally relevant protein interactions for further experimental validation.