Fast protein structure comparison through effective representation learning with contrastive graph neural networks.

Area of Science:

• Structural bioinformatics
• Computational biology
• Machine learning in structural biology

Background:

• Protein structure alignment algorithms are computationally intensive, hindering large-scale similarity-based retrieval.
• The rapid growth of protein structure databases necessitates more efficient comparison methods.

Purpose of the Study:

• To introduce GraSR, an effective graph-based representation learning method for fast and accurate protein structure comparison.
• To address the limitations of existing time-consuming alignment-based approaches.

Main Methods:

• Constructing protein tertiary structures into graphs based on intra-residue distances.
• Employing deep graph neural networks (GNNs) with short-cut connections for representation learning under a contrastive framework.
• Incorporating a dynamic training data partition strategy and length-scaling cosine distance.

Main Results:

• GraSR achieved 7%-10% improvement over state-of-the-art methods on SCOPe v2.07 and a PDB test set.
• GraSR demonstrated significantly faster performance compared to traditional alignment-based methods.
• Analysis revealed that learned discriminative residue-level and global descriptors contribute to GraSR's superiority.

Conclusions:

• GraSR offers a computationally efficient and accurate solution for protein structure comparison and similarity retrieval.
• The method's effectiveness stems from its graph-based representation learning and novel training strategies.
• GraSR provides a valuable tool for the structural bioinformatics community, with publicly available code and a web server.