Struct2GO-Enhanced: Multimodal Graph Attention Improves Protein Function Prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein function prediction is crucial for understanding biological systems.
• Current models struggle with multimodal feature fusion and attention for structure-function relationships.
• AlphaFold2 structural information has advanced prediction but requires better integration.

Purpose of the Study:

• To develop an enhanced framework for protein function prediction.
• To improve multimodal feature fusion and attention mechanisms.
• To leverage AlphaFold2 structural data more effectively.

Main Methods:

• Introduced Graph-CBAM for graph neural network attention.
• Implemented complete multimodal fusion of Node2vec embeddings and one-hot encodings.
• Utilized a dual-head self-attention pooling module for node importance stabilization.

Main Results:

• The enhanced model consistently outperformed existing benchmarks on human protein datasets.
• Achieved a 2.9% increase in Fmax for Biological Process (BP) and 3.9% AUPR enhancement for Cellular Component (CC).
• Demonstrated the independent contributions of structural embeddings, one-hot encodings, and Graph-CBAM via ablation studies.

Conclusions:

• The proposed framework offers a more complete and practical solution for AlphaFold2-based protein function prediction.
• The model shows particular advantages for proteins lacking protein-protein interaction data.
• This work advances the field by improving the capture of structural-functional relationships.