Domain-PFP allows protein function prediction using function-aware domain embedding representations.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein domains are key functional and structural units.
• Characterizing protein domains aids in understanding protein function.
• Existing methods for protein function prediction have limitations.

Purpose of the Study:

• To develop a self-supervised protocol for creating functionally consistent domain representations.
• To improve protein function prediction accuracy using these domain embeddings.
• To introduce and evaluate a novel function prediction method, Domain-PFP.

Main Methods:

• Employing a self-supervised protocol to learn domain-Gene Ontology (GO) co-occurrences and associations.
• Constructing domain embeddings based on learned associations.
• Developing the Domain-PFP method for function prediction using domain embeddings.

Main Results:

• Domain embeddings are effective for protein function prediction tasks.
• Protein representations using domain embeddings outperform large-scale protein language models in GO prediction.
• Domain-PFP significantly surpasses state-of-the-art function predictors.
• Domain-PFP achieved top performance in the CAFA3 evaluation.

Conclusions:

• Self-supervised learning of domain-GO associations provides powerful representations for protein function prediction.
• Domain-PFP offers a superior approach to predicting protein functions compared to existing methods.
• The domain embedding strategy represents a significant advancement in bioinformatics for functional genomics.