StructureNet

StructureNet predicts protein-ligand binding affinity using a graph neural network model that processes solely structural descriptors.

Key Features:

• Structure-based representation: Represents protein and ligand structures as graphs processed using a GNN-based ensemble deep learning model.
• Geometric and topological descriptors: Incorporates geometric and topological descriptors to improve structural representations of protein-ligand complexes.
• Ensemble conformer processing: Processes ensembles of binding complex conformers generated using molecular dynamics simulations to capture binding site flexibility.
• Hybridization capability: Can be hybridized with interaction- and sequence-based models for comparative performance analysis.

Scientific Applications:

• Binding affinity prediction: Predicts protein-ligand binding affinity for drug discovery applications.
• Virtual screening: Distinguishes between active and decoy ligands for rapid virtual screening of drug candidates.
• De novo drug design: Addresses pattern recognition challenges in de novo applications by avoiding data memorization from sequence and interaction features.
• Binding site flexibility analysis: Improves model accuracy by incorporating multiple conformations of the same complex to capture binding site flexibility.

Methodology:

StructureNet represents protein and ligand structures as graphs and processes them using a GNN-based ensemble deep learning model. An ablation study identified geometric descriptors as key drivers of model performance.