DENVIS

DENVIS performs high-throughput ligand virtual screening using graph neural networks (GNNs) to predict protein–ligand interactions without relying on docking, enabling rapid early-stage drug discovery.

Key Features:

• Graph Neural Networks (GNNs): Uses GNNs to model protein–ligand interactions without requiring pre-determined binding poses from docking.
• Protein Pocket Modeling: Represents protein pockets with both atomic and surface features to improve binding-site characterization.
• Docking-Free End-to-End Pipeline: Eliminates intermediate docking steps to provide substantially faster virtual screening compared to docking-based and hybrid methods.
• Model Ensembles and Data Augmentation: Employs model ensembles and data augmentation, including artificial negative sampling during training, to improve robustness and predictive performance.
• Lower-Level Protein Representation: Processes protein information at a representation level below sequence-based methods to support high-throughput screening.
• Scalability: Designed to handle large-scale screenings, including libraries on the order of billions of molecules.

Scientific Applications:

• Early-Stage Drug Discovery: Prioritizes and identifies potential hit compounds for specified protein targets in early discovery campaigns.
• Large-Scale Virtual Screening: Enables screening of very large compound libraries without computationally intensive docking.
• Protein–Ligand Interaction Prediction: Provides predictive scoring of ligand binding propensities using learned structural representations of pockets and ligands.

Methodology:

Integrates graph neural networks with traditional computational techniques; models protein pockets using atomic and surface features; avoids reliance on protein–ligand complex binding poses; trains with model ensembles and data augmentation including artificial negative sampling; validated on two benchmark databases.