Hierarchical graph transformer with contrastive learning for protein function prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• High-throughput sequencing provides vast protein data, but functional annotation is slow and costly.
• Computational models are crucial for accelerating protein function prediction.
• Existing graph neural networks struggle with long-range structural correlations and identifying key residues.

Purpose of the Study:

• To introduce a novel deep learning model, Hierarchical graph Transformer with contrastive Learning (HEAL), for enhanced protein function prediction.
• To address limitations in capturing structural semantics and long-distance dependencies in protein graphs.
• To improve the accuracy and efficiency of computational protein function prediction.

Main Methods:

• Developed HEAL, a hierarchical graph Transformer model incorporating super-nodes to mimic functional motifs.
• Employed graph contrastive learning for network optimization and improved graph representation.
• Integrated AlphaFold2 predicted structures to enhance predictions for proteins lacking experimental data.

Main Results:

• HEAL-PDB achieved comparable performance to state-of-the-art methods with less training data.
• HEAL significantly outperformed DeepFRI on PDBch test set when using AlphaFold2 predicted structures.
• HEAL demonstrated superior performance on AFch test set compared to DeepFRI and DeepGOPlus using predicted structures.
• Class activation mapping enabled HEAL to identify functional sites within proteins.

Conclusions:

• HEAL offers a powerful and accurate approach for protein function prediction, particularly when experimental structures are unavailable.
• The model's ability to capture hierarchical structural semantics and utilize predicted structures represents a significant advancement.
• HEAL provides a valuable tool for accelerating functional annotation in the post-genomic era.