Using InterLabelGO+ for Accurate Protein Language Model-Based Function Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The advent of large language models for protein sequences has accelerated the development of deep learning approaches for predicting protein functions, primarily through Gene Ontology (GO) terms.
• Accurate protein function prediction is crucial for understanding biological systems and disease mechanisms.

Purpose of the Study:

• To introduce InterLabelGO+, a novel deep learning-based model for protein GO term prediction.
• To demonstrate the effectiveness of InterLabelGO+ in the Critical Assessment of Function Annotation (CAFA5) challenge.
• To provide accessible tools (webserver and standalone package) for protein function prediction and model retraining.

Main Methods:

• Utilized the ESM2 protein language model to extract sequence-based features.
• Developed a deep learning model trained with a specialized loss function accounting for inter-term relationships.
• Integrated deep learning predictions with GO terms derived from sequence homology searches for consensus predictions.

Main Results:

• InterLabelGO+ achieved top performance in the CAFA5 challenge for protein function prediction.
• The model effectively leverages sequence features and homology information for accurate GO term assignment.
• The developed methods provide a robust framework for predicting protein functions.

Conclusions:

• InterLabelGO+ represents a significant advancement in deep learning-based protein function prediction.
• The accessibility of the InterLabelGO+ webserver and package facilitates broader research applications.
• The ability to retrain the model ensures its continued relevance with evolving biological data.