A unified graph-based approach for protein function prediction using AlphaFold structures and sequence features.

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Predicting protein function is crucial for understanding biological systems and diseases.
• Traditional methods often overlook protein structure, relying mainly on sequence and interaction data.
• Advances in protein structure prediction, like AlphaFold, enable new approaches.

Purpose of the Study:

• To develop a novel hybrid model, StructSeq2GO, for enhanced protein function prediction.
• To integrate structural information from AlphaFold with sequence data for improved accuracy.
• To predict Gene Ontology (GO) labels for proteins.

Main Methods:

• StructSeq2GO utilizes graph representation learning on AlphaFold-predicted structures.
• It combines structural features with sequence embeddings from the ProteinBERT language model.
• The model predicts GO labels across Biological Process, Cellular Component, and Molecular Function ontologies.

Main Results:

• StructSeq2GO achieved state-of-the-art performance in predicting protein function across three GO domains.
• Key performance metrics include Fmax (up to 0.681), AUC (up to 0.939), and AUPR (up to 0.763).
• Results underscore the importance of structural context, which complements sequence-only information.

Conclusions:

• Integrating protein structure and sequence data significantly enhances function prediction accuracy.
• StructSeq2GO demonstrates the power of combining structural insights with advanced language models like ProteinBERT.
• Future work could involve improving structure confidence modeling and extending predictions to pathways or diseases.