GraSR

GraSR encodes protein tertiary structures as graphs and applies graph-based representation learning to enable fast, accurate protein structure similarity search and comparison.

Key Features:

• Graph-Based Representation Learning: Constructs graphs from intra-residue distances derived from protein tertiary structures to represent structural information.
• Deep Graph Neural Networks (GNNs): Employs deep GNNs with short-cut connections to learn residue-level and global structural representations.
• Contrastive Learning Framework: Trains representations under a contrastive learning objective to enhance discriminative power for structure comparison.
• Dynamic Training Data Partition Strategy: Adapts the training data partitioning during training to improve model performance and robustness.
• Length-Scaling Cosine Distance: Uses a length-scaling cosine distance metric to refine distance calculations between protein structure representations.

Scientific Applications:

• Protein Structure Similarity Searches: Facilitates identification of structurally similar proteins across large structural datasets.
• Functional Annotation: Supports prediction of protein function by comparing query structures to annotated structures with similar folds.
• Drug Discovery: Assists in identifying potential drug targets by comparing structures of proteins involved in disease pathways.

Methodology:

Protein tertiary structures are converted into graphs using intra-residue distances as edges. Deep GNNs with short-cut connections are trained under a contrastive learning framework with a dynamic training data partition strategy. Length-scaling cosine distance is applied to refine comparisons between learned representations. Performance was evaluated on SCOPe v2.07 and an independent PDB test set, reporting a 7%–10% improvement in accuracy over state-of-the-art methods and faster runtime than alignment-based approaches.