ProCyon: A multimodal foundation model for protein phenotypes

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Genomics

Background:

• A significant portion of human proteins lack experimentally validated functions or context-specific phenotypic insights.
• Accurate protein characterization is crucial for understanding biological processes and disease mechanisms.

Purpose of the Study:

• To develop a universal approach for modeling protein phenotypes across diverse knowledge domains.
• To introduce ProCyon, a multimodal foundation model for generating and predicting protein phenotypes.

Main Methods:

• ProCyon integrates protein sequence, structure, and natural language processing.
• The model is trained on a novel dataset, ProCyon-Instruct, comprising 33 million protein phenotype instructions.
• Benchmarking involved evaluating ProCyon on dozens of tasks against existing single-modal and multimodal models.

Main Results:

• ProCyon demonstrates competitive performance across various benchmarking tasks.
• The model successfully retrieves proteins based on drug mechanisms and disease contexts.
• ProCyon generates phenotypic descriptions for uncharacterized proteins, including those linked to Parkinson's disease.
• Experimental validation in multiple sclerosis identified novel genes and elucidated pathway mechanisms.

Conclusions:

• ProCyon offers a powerful, generalizable method for generating functional insights into the human proteome.
• The model's multimodal approach enhances the prediction and understanding of protein phenotypes.
• ProCyon has the potential to accelerate biological discovery and disease research.