Efficient decoy selection to improve virtual screening using machine learning models.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Machine learning models for drug discovery rely heavily on protein-ligand interaction fingerprints.
• The performance of these models is critically dependent on the decoy selection strategies used.
• Protein per Atom Score Contributions Derived Interaction Fingerprint (PADIF) is a key feature for model development.

Purpose of the Study:

• To analyze various decoy selection strategies for enhancing machine learning models based on PADIF.
• To evaluate the effectiveness of different decoy sources, including random databases, high-throughput screening non-binders, and docking-generated conformations.
• To validate model performance using experimentally determined inactive compounds.

Main Methods:

• Explored three decoy selection workflows: random selection (ZINC15), recurrent non-binders (dark chemical matter), and data augmentation (docking conformations).
• Trained and tested machine learning models using PADIF with active molecules from ChEMBL and the selected decoy approaches.
• Validated model performance against experimentally determined inactive compounds from the LIT-PCBA dataset.

Main Results:

• Models trained with random ZINC15 selections and dark chemical matter compounds demonstrated performance comparable to models using actual non-binders.
• All developed models exhibited improved exploration of novel chemical spaces for specific targets.
• The models enhanced the selection of top active compounds compared to classical scoring functions, increasing molecular docking screening power.

Conclusions:

• Random selection from ZINC15 and utilization of dark chemical matter are effective decoy strategies when specific inactivity data is scarce.
• Appropriate decoy selection maintains model accuracy and expands applicability to new targets.
• These strategies significantly boost screening power in molecular docking for drug discovery.