Learning maximally spanning representations improves protein function annotation

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Automated protein function annotation is vital for understanding biological processes, medicine, and biotechnology.
• Existing methods struggle with imbalanced data, leading to poor performance on understudied protein functions.
• This imbalance stems from biases in data collection and protein evolution.

Purpose of the Study:

• To develop MSRep, a novel deep learning framework to address data imbalance in protein function annotation.
• To improve prediction accuracy for both well-represented and underrepresented protein functions.
• To enhance the generalizability of protein function prediction models.

Main Methods:

• MSRep refines a pre-trained protein language model using a novel loss function inspired by neural collapse (NC).
• The framework induces an NC-like structure to ensure balanced representation of all function classes in the embedding space.
• Evaluated across four diverse protein function annotation tasks (EC numbers, Gene3D, Pfam, GO terms).

Main Results:

• MSRep demonstrated superior predictive performance across all tested annotation tasks.
• The framework significantly improved accuracy for both well-studied and understudied protein functions.
• MSRep outperformed several existing state-of-the-art protein annotation tools.

Conclusions:

• MSRep effectively addresses the challenge of imbalanced data in protein function annotation.
• The approach enhances the annotation of understudied functions and uncharacterized proteins.
• MSRep holds promise for advancing protein function studies and accelerating biological discovery.