GraphBepi

GraphBepi predicts B-cell epitopes by integrating AlphaFold2-predicted protein structures with ESM-2 residue embeddings, an edge-enhanced graph neural network (EGNN), BiLSTM sequence modeling, and a multilayer perceptron for final epitope classification.

Key Features:

• Integration of Structural Information: Uses AlphaFold2-predicted 3D structures to construct protein graphs with nodes representing residues encoded by ESM-2 embeddings.
• Graph Neural Network Architecture: Employs an edge-enhanced deep graph neural network (EGNN) to capture geometric and spatial relationships from 3D structures.
• Sequence Dependency Analysis: Applies bidirectional long short-term memory (BiLSTM) networks to model long-range dependencies in amino acid sequences.
• Multilayer Perceptron for Prediction: Integrates low-dimensional representations from EGNN and BiLSTM via a multilayer perceptron for final epitope prediction.

Scientific Applications:

• B-cell Epitope Prediction and Benchmarking: Improves prediction on curated epitope datasets, reporting over 5.5% improvement in AUC and 44.0% improvement in AUPR versus state-of-the-art methods.
• Vaccine and Immunotherapy Design: Provides residue-level epitope predictions to inform rational vaccine development and immunotherapy design.

Methodology:

AlphaFold2 predicts protein 3D structures; predicted structures are transformed into protein graphs with residues encoded by ESM-2; EGNN processes the graphs to extract spatial information; BiLSTM analyzes sequence dependencies; a multilayer perceptron combines EGNN and BiLSTM outputs for final epitope prediction.