GGN-GO: geometric graph networks for predicting protein function by multi-scale structure features.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• High-throughput sequencing generates vast genomic and transcriptomic data, but most protein functions remain unannotated.
• Traditional experimental methods for protein function annotation are resource-intensive.
• Existing deep learning methods struggle to capture fine-grained atomic geometric features and long-range dependencies in protein structures.

Purpose of the Study:

• To develop a novel geometric graph network (GGN-GO) for accurate protein function prediction.
• To address the limitations of current methods in capturing multi-scale geometric structural features and identifying key residues.
• To improve the efficiency and accuracy of protein function annotation.

Main Methods:

• Proposed a geometric graph network (GGN-GO) incorporating multi-scale geometric structural features at atomic and residue levels.
• Utilized a geometric vector perceptron for feature representation and aggregation.
• Implemented a graph attention pooling layer and contrastive learning for enhanced graph representation and discriminability.

Main Results:

• GGN-GO outperformed six comparative methods in protein function prediction tasks for both experimental and predicted structures.
• The model demonstrated interpretability by identifying functionally relevant residues that correspond to experimentally confirmed sites.
• Achieved superior performance in tasks with a high number of labels.

Conclusions:

• GGN-GO offers a significant advancement in protein function prediction by effectively leveraging geometric structural information.
• The method provides a more interpretable and accurate approach to annotating protein functions, aiding in the understanding of biological processes.
• GGN-GO's ability to pinpoint key functional residues highlights its potential for guiding experimental validation.