ProtGO: universal protein function prediction utilizing multi-modal gene ontology knowledge.

Area of Science:

• Biochemistry and Molecular Biology
• Bioinformatics
• Artificial Intelligence in Life Sciences

Background:

• Protein function prediction is a critical challenge in life sciences, exacerbated by the increasing number of AI-designed proteins and the need to process multi-modal data.
• Traditional low-throughput experimental methods are insufficient to keep pace with the growing volume of known protein sequences and the demand for functional prediction.

Purpose of the Study:

• To develop a universal multi-modal method for protein function prediction that addresses the limitations of current approaches.
• To leverage the Gene Ontology (GO) knowledge base and integrate diverse data modalities for enhanced prediction accuracy.

Main Methods:

• ProtGO framework utilizes pre-trained protein language models (PLMs) for sequence representation.
• Integrates specialized modules for text descriptions (text alignment), species-specific taxonomy (taxonomy encoding), and biological relationships (GO graph embedding).
• Combines four distinct knowledge representations to maximize the utility of GO resources.

Main Results:

• ProtGO demonstrated significant improvements in protein function prediction, with an 8% to 27% increase in the maximum F1 measure (Fmax) compared to baseline models.
• The framework effectively enhances the performance of standard PLMs and biological language models (LMs) in GO prediction tasks.
• Achieved outstanding performance by integrating functional and evolutionary knowledge from multiple data sources.

Conclusions:

• ProtGO offers a robust and adaptable framework for accurate protein function prediction.
• The multi-modal approach effectively harnesses biological knowledge, advancing the field of bioinformatics.
• This work provides a valuable tool for researchers in life sciences and biomedicine facing big data challenges.