DeepCoy

DeepCoy generates property-matched decoy molecules for virtual screening to mitigate dataset bias and improve benchmarking of molecular recognition methods.

Key Features:

• Bias Mitigation: Generates decoy molecules tailored to match or introduce specific biases in physicochemical properties to address bias in actives/decoys datasets.
• Deep Learning Generation: Uses a deep learning–based approach to produce decoys that closely resemble active molecules in physicochemical properties without increasing false negatives.
• Validation and Performance: Validated on DUD-E and DEKOIS 2.0 benchmarks, improving the Deviation from Optimal Embedding (DOE) score by an average of 81% for DUD-E (0.166 → 0.032) across all 102 DUD-E targets and 66% for DEKOIS 2.0 (0.109 → 0.038) for 80 of 81 targets.
• Docking Challenge: Produces decoys that are more challenging for docking methods, demonstrated by a decrease in Area Under the Receiver Operating Characteristic Curve (AUC ROC) for Autodock Vina from 0.70 to 0.63.

Scientific Applications:

• Virtual screening benchmarking: Construction of property-matched decoy sets for more rigorous evaluation of virtual screening algorithms.
• Molecular recognition method development: Generation of datasets that reduce confounding biases to aid development and validation of predictive models for molecular interactions.
• Drug discovery dataset curation: Creation of decoy collections matched to active compounds for use in target-specific benchmarking and method comparison.

Methodology:

Generates decoys using deep learning to match physicochemical properties of actives and was validated on DUD-E and DEKOIS 2.0 using the Deviation from Optimal Embedding (DOE) score and docking evaluations with Autodock Vina reporting AUC ROC.