Deep Docking

Deep Docking applies deep learning–based QSAR models to predict docking scores and accelerate virtual screening of ultra-large chemical libraries (e.g., 1.36 billion ZINC15 molecules) against protein targets.

Key Features:

• Deep learning QSAR models: Trains QSAR models on docking scores from subsets of chemical libraries to predict docking outcomes for unprocessed entries.
• Iterative filtering: Iteratively removes unfavorable molecules to prioritize candidates for further docking.
• Docking program integration: Integrates with existing docking programs such as FRED for docking calculations.
• Scale demonstration: Demonstrated on 1.36 billion ZINC15 molecules against 12 target proteins, reporting up to a 100-fold data reduction and up to a 6000-fold enrichment of high-scoring molecules.
• Accuracy retention: Maintains identification of favorably docked entities despite large-scale reduction and enrichment.

Scientific Applications:

• Ultra-large virtual screening: Enables screening and prioritization of billions of compounds from libraries such as ZINC15.
• Hit enrichment: Enriches high-scoring docking hits to increase the fraction of favorably docked molecules for downstream testing.
• Large-scale docking campaigns: Supports multi-target docking campaigns (e.g., demonstrated on 12 target proteins) using established docking engines like FRED.

Methodology:

Train deep learning QSAR models on docking scores from subset dockings, predict scores for remaining entries, iteratively remove unfavorable molecules, and perform docking with programs such as FRED.