Struct2GO: protein function prediction based on graph pooling algorithm and AlphaFold2 structure information.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Deep learning models for protein function prediction typically rely on sequence data and protein-protein interaction networks.
• These models struggle with newly sequenced proteins lacking interaction data.
• Recent advances in protein structure prediction, like AlphaFold2, provide accurate atomic-level structural information.

Purpose of the Study:

• To develop a novel deep learning model, Struct2GO, for enhanced protein function prediction.
• To improve the precision and generality of protein function prediction, especially for proteins not present in interaction networks.
• To leverage both protein structure and sequence data for more robust predictions.

Main Methods:

• Utilized graph representation learning to obtain amino acid residue embeddings from protein structures.
• Employed a self-attention-based graph pooling algorithm to capture global structural features.
• Fused structural features with sequence features derived from a protein language model.

Main Results:

• The Struct2GO model demonstrated superior performance compared to traditional protein sequence-based function prediction models.
• Integration of structural information significantly improved prediction accuracy.
• The model showed enhanced generality for proteins with limited or no existing interaction data.

Conclusions:

• Combining protein structure and sequence data offers a powerful approach for accurate protein function prediction.
• Struct2GO provides a more generalizable solution for predicting functions of novel proteins.
• This study highlights the potential of integrating structural bioinformatics with deep learning for advancing biological discovery.