Accurate protein function prediction via graph attention networks with predicted structure information

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Science

Background:

• Experimental protein function annotation is a bottleneck, with less than 0.1% of protein sequences experimentally characterized.
• Existing computational methods for protein function prediction lack sufficient accuracy.
• Advances in protein structure prediction and protein language models offer new opportunities for improving prediction accuracy.

Purpose of the Study:

• To develop an advanced computational method, GAT-GO, for substantially improving protein function prediction.
• To leverage predicted protein structure information and protein sequence embeddings within a graph attention network (GAT) framework.
• To evaluate the performance of GAT-GO against state-of-the-art sequence- and structure-based methods.

Main Methods:

• Development of GAT-GO, a graph attention network (GAT) model.
• Integration of predicted protein structure information and protein sequence embeddings as input features.
• Evaluation on PDB-mmseqs and PDB-cdhit test sets with varying sequence identity thresholds.

Main Results:

• GAT-GO significantly outperforms homology-based methods like BLAST on the PDB-mmseqs dataset (Fmax 0.508 vs 0.117 for MFO).
• On the PDB-cdhit dataset, GAT-GO surpasses the performance of DeepFRI, a recent method utilizing experimental structures (Fmax 0.637 vs 0.542 for MFO).
• GAT-GO demonstrates superior performance across multiple Gene Ontology (GO) domains (MFO, BPO, CCO) in terms of Fmax and AUPRC.

Conclusions:

• GAT-GO represents a substantial advancement in computational protein function prediction.
• The method effectively utilizes predicted structural information and sequence embeddings to achieve high accuracy.
• GAT-GO offers a scalable and accurate solution to the challenge of annotating functions for large protein sequence databases.