Experimental assessment of AI-based interactome mapping.

Area of Science:

• Molecular Biology
• Computational Biology
• Genetics

Background:

• Genotype-phenotype relationships depend on macromolecular interactions.
• Understanding the protein-protein interactome is crucial for genetics and cellular biology.
• AI advancements, like AlphaFold, enable accurate tertiary protein structure prediction, reviving interest in computational interactome prediction.

Purpose of the Study:

• To experimentally assess the impact and performance of AI-driven interactome predictions.
• To compare AI predictions against experimental interactome maps for yeast and humans.
• To evaluate the discovery potential of AI in identifying novel protein-protein interactions (PPIs).

Main Methods:

• Development of a comprehensive experimental framework.
• Systematic assessment of AI-driven interactome predictions (e.g., AlphaFold).
• Proteome-wide screening and comparison with experimental reference interactome maps.

Main Results:

• High-confidence AI predictions match the quality of established experimental methods.
• AI approaches underperform in discovering novel PPIs compared to experimental maps.
• The yeast interactome map identified over 40-fold more novel PPIs than AI.
• AlphaFold provided structural models for many experimentally identified PPIs missed by AI screens.

Conclusions:

• AI's current strength lies in modeling the quaternary structures of experimentally identified PPIs.
• Experimental methods remain superior for discovering novel PPIs at a proteome-wide scale.
• AI and experimental approaches are complementary in interactome research.