EGNA

EGNA predicts protein-ligand binding affinities by representing bound complexes as graphs and applying graph convolutional neural network models to capture intermolecular interaction patterns for structural bioinformatics and drug discovery.

Key Features:

• Graph-Based Representation: Constructs graphs representing proteins, ligands, and their interactions based on distinct regions within each bound complex.
• Graph Convolutional Layers: Applies graph convolutional layers to capture interaction patterns between protein and ligand molecules and to simulate empirical scoring functions (SFs).
• Differential Contribution Modeling: Models the differential contribution of atoms across various protein regions to binding affinity.
• Interatomic Relationship Exploration: Explores relationships between atom properties and intermolecular distances to inform binding strength.
• Transparency in Factor Contribution: Provides means to investigate the contributions of various factors to predicted binding affinity.
• Deep Learning without Complex Feature Engineering: Leverages deep learning approaches that avoid reliance on extensive manual feature engineering.

Scientific Applications:

• Drug Discovery and Design: Predicts and ranks protein-ligand binding affinities to support identification and optimization of candidate therapeutic compounds.
• Scoring Function Evaluation: Serves to evaluate and compare empirical and machine learning–based scoring functions for protein-ligand complexes.
• Atomic Contribution Analysis: Enables analysis of how individual atomic properties and interatomic distances influence binding affinity.
• Benchmarking and Generalization Assessment: Can be used in comparative evaluations across benchmark datasets to assess robustness and generalization of affinity prediction models.

Methodology:

Constructs graphs from bound protein-ligand complex regions, processes those graphs with graph convolutional layers that simulate empirical scoring functions to capture interaction patterns, and models differential atomic contributions while exploring atom-property versus intermolecular-distance relationships using deep learning without complex feature engineering.