RealVS

RealVS improves prioritization of active compounds in ligand-based virtual screening by optimizing top-k hit precision for drug discovery.

Key Features:

• Transferable Information Integration: Leverages transferable information from source domains to mitigate the insufficiency of inactive ligands for specific drug targets.
• Adversarial Domain Alignment: Employs adversarial domain alignment to align the distribution of generated features between training datasets and screening databases to enhance generalization.
• Novel Objective Function: Optimizes classification loss, regression loss, and adversarial loss simultaneously, prioritizing exclusion of most inactive ligands before activity regression prediction.
• Top-k Precision Focus: Concentrates on improving precision of top-k hits rather than predicting bioactivities across all compounds to better select promising candidates.
• Graph Attention Networks (GATs): Uses Graph Attention Networks to identify key ligand substructures associated with bioactivity and to enhance interpretability.

Scientific Applications:

• Ligand-based Virtual Screening: Prioritizes candidate compounds in large screening libraries by improving top-k hit precision in virtual screens.
• Lead Prioritization in Drug Discovery: Refines selection of potential drug leads by reducing false positives among top-ranked compounds.
• Structure–Activity Interpretation: Identifies substructures correlated with bioactivity via GATs to inform structure–activity relationship analyses.
• Benchmark Validation: Demonstrates improved top-k precision across multiple benchmark datasets and varying k values.

Methodology:

Integrates transferable information from source domains; applies adversarial domain alignment to match feature distributions between training and screening sets; uses a joint objective optimizing classification, regression, and adversarial losses to exclude inactive ligands before regression; and employs Graph Attention Networks to detect key substructures.