DeepFRI

DeepFRI predicts protein function and identifies functional residues by integrating protein structural data with sequence features from a protein language model using Graph Convolutional Networks (GCNs).

Key Features:

• Structure-based prediction: Performs protein function prediction using protein structural data.
• Graph Convolutional Networks (GCNs): Implements GCNs to learn from protein structural graphs.
• Protein language model features: Utilizes sequence features extracted by a protein language model to capture sequence-level signals.
• Sequence–structure integration: Combines protein language model-derived sequence features with structural information for prediction.
• Class activation mapping: Applies class activation mapping techniques to localize predicted functions to specific residues.
• Residue-level annotations: Produces site-specific, residue-resolution functional annotations.
• Robustness to homology models: Maintains predictive accuracy when experimental structures are substituted with homology models, indicating de-noising capability.
• Performance versus CNNs: Demonstrates improved predictive performance compared with sequence-based Convolutional Neural Networks (CNNs).
• Scalability: Scales to large protein sequence repositories for high-throughput annotation.

Scientific Applications:

• PDB annotation: Annotates structures from the Protein Data Bank (PDB) to generate new functional predictions.
• SWISS-MODEL annotation: Annotates homology models from SWISS-MODEL to extend functional coverage.
• Site-specific functional mapping: Identifies residue-level functional sites for structure–function studies.
• Experimental versus modeled structure analysis: Enables analyses using experimental structures and homology models with only minor drops in accuracy.
• Large-scale annotation: Facilitates high-throughput annotation of protein sequences and structures.

Methodology:

Uses Graph Convolutional Networks (GCNs) integrated with sequence features extracted from a protein language model and protein structural data, and employs class activation mapping for residue-level localization.