NetGO 3.0: Protein Language Model Improves Large-scale Functional Annotations.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Automated Function Prediction (AFP) methods like NetGO 2.0 are crucial for understanding protein functions.
• Current AFP methods primarily rely on experimentally annotated proteins, overlooking the vast potential of unannotated protein data.
• Protein language models (PLMs) offer a novel approach to learn sequence-based representations through self-supervision.

Purpose of the Study:

• To develop an improved AFP method by incorporating information from unannotated proteins.
• To evaluate the efficacy of using protein language model embeddings for AFP.
• To enhance the NetGO 2.0 framework with a new PLM-based component.

Main Methods:

• Representing proteins using Evolutionary Scale Modeling (ESM)-1b embeddings.
• Training a logistic regression (LR) model, termed LR-ESM, for AFP using ESM-1b representations.
• Integrating the developed LR-ESM model into the existing NetGO 2.0 framework to create NetGO 3.0.

Main Results:

• The LR-ESM model demonstrated performance comparable to the top components of NetGO 2.0.
• The integration of LR-ESM into NetGO 2.0 resulted in substantial performance improvements in automated function prediction.
• NetGO 3.0 effectively leverages information from unannotated proteins, expanding the scope of AFP.

Conclusions:

• Protein language models, exemplified by ESM-1b, provide valuable representations for improving automated function prediction.
• NetGO 3.0 represents a significant advancement in AFP by successfully integrating PLM-based features.
• The enhanced NetGO 3.0 tool offers improved performance and broader applicability in protein function annotation.