Domain-PFP: Protein Function Prediction Using Function-Aware Domain Embedding Representations.

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Science

Background:

• Proteins perform biological functions via distinct structural and functional units called domains.
• Accurate characterization of protein domains is crucial for understanding protein function.
• Existing methods for protein function prediction have limitations in capturing domain-specific information.

Approach:

• Developed a self-supervised protocol to learn domain representations by analyzing domain-Gene Ontology (GO) co-occurrences.
• Constructed domain embeddings that capture functional consistency.
• Utilized these embeddings for protein function prediction tasks.

Key Points:

• Domain embeddings effectively represent protein functions.
• Protein representations derived from domain embeddings surpass large-scale protein language models in GO prediction.
• The Domain-PFP method, based on domain embeddings, significantly outperforms state-of-the-art function predictors.
• Domain-PFP achieved top performance in the CAFA3 evaluation.

Conclusions:

• Self-supervised learning of domain embeddings provides a powerful approach for protein function prediction.
• Domain-PFP offers a superior alternative to existing methods for predicting protein functions.
• This work advances the field of computational protein analysis and functional genomics.