Generating property-matched decoy molecules using deep learning.

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Established active and decoy sets are crucial for virtual screening method development.
• Common decoy sets possess inherent biases, leading to methods that exploit these biases rather than learning molecular recognition.
• This bias hinders the generalization capabilities of virtual screening methods and impedes progress in the field.

Purpose of the Study:

• To develop a deep learning method, DeepCoy, for generating unbiased decoy molecules.
• To enable the construction of decoy sets with user-defined specifications to remove or introduce bias.
• To improve the reliability and accuracy of virtual screening benchmarks.

Main Methods:

• Developed DeepCoy, a deep learning model for generating decoy molecules.
• Validated DeepCoy using two standard benchmarks: DUD-E and DEKOIS 2.0.
• Assessed decoy quality by comparing physicochemical properties with active molecules and evaluating docking performance.

Main Results:

• DeepCoy successfully generated decoy molecules that closely matched the physicochemical properties of active molecules across 102 DUD-E and 80 DEKOIS 2.0 targets.
• The generated decoys significantly improved the Deviation from Optimal Embedding (DOE) score, reducing it from 0.166 to 0.032 for DUD-E and 0.109 to 0.038 for DEKOIS 2.0.
• Virtual screening performance, measured by AUC ROC, decreased from 0.70 to 0.63 when using DeepCoy decoys, indicating they are more challenging for docking software.

Conclusions:

• DeepCoy effectively generates high-quality decoy molecules, addressing the critical issue of bias in virtual screening benchmarks.
• The method provides a powerful tool for creating customized decoy sets, thereby advancing the development and validation of more robust virtual screening techniques.
• The improved decoy sets facilitate more accurate assessment of molecular recognition capabilities, ultimately aiding in more effective drug discovery.