Context-informed subgraph foundation models enable interpretable protein-function prediction.

Area of Science:

• Genomics
• Computational Biology
• Biochemistry

Background:

• Protein function prediction is vital for understanding biological processes and developing therapeutics.
• Current computational tools struggle with predicting rare, uncharacterized, or indirect protein functions.
• Accurate protein annotation is essential for advancing biological and medical research.

Purpose of the Study:

• To introduce COSMOS, a novel context-aware Gene Ontology (GO) subgraph mining system for enhanced protein-function prediction.
• To address limitations in predicting functions for GO classes with sparse or no experimental annotations.
• To provide an interpretable and complementary tool for protein annotation.

Main Methods:

• Leveraged inductive subgraph foundation models and an enriched knowledge graph of protein-GO relationships.
• Utilized 7,923,952 functional semantic relationships for model training and validation.
• Employed zero-shot, few-shot, and low-homology learning paradigms for robust prediction.

Main Results:

• Achieved state-of-the-art predictions for sparsely annotated GO classes.
• Generated interpretable functional subgraphs, enabling transparent rationale analysis.
• Demonstrated complementary benefits when integrated with existing embedding-based prediction methods.

Conclusions:

• COSMOS offers a robust and interpretable approach for protein-function prediction, particularly for underexplored GO classes.
• The system serves as a valuable complementary tool to conventional protein annotation methods.
• COSMOS has the potential to significantly advance genomics and therapeutic research through improved protein function elucidation.