Applying Machine Learning to Ultrafast Shape Recognition in Ligand-Based Virtual Screening.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Ligand-Based Virtual Screening (LBVS) methods like Ultrafast Shape Recognition (USR) use molecular descriptors for similarity calculations.
• Existing USR-derived methods can be computationally intensive and may have limitations in detecting novel leads.

Purpose of the Study:

• To explore machine learning techniques for improving similarity detection in LBVS using USR descriptors.
• To enhance the performance and efficiency of virtual screening for identifying potential drug leads.

Main Methods:

• Trained machine learning models (Gaussian Mixture Models, Isolation Forests, Artificial Neural Networks) using USR descriptors from the Directory for Useful Decoys-Enhanced dataset.
• Evaluated models using full conformer models and Lowest Energy Conformations (LECs).
• Assessed model performance with reduced training dataset sizes to simulate real-world virtual screening scenarios.

Main Results:

• Machine learning models, particularly GMMs, showed significant performance gains over ElectroShape 5D, with up to 430% better mean Enrichment Factor.
• Models maintained performance within 10% of the mean even with reduced training data.
• Retrospective screening times were reduced by an average factor of 10 compared to standard USR.

Conclusions:

• Machine learning techniques substantially improve the performance and efficiency of USR-based virtual screening.
• The developed models offer a faster and more effective approach for identifying potential drug leads.
• This study highlights the potential of integrating ML with established cheminformatics methods for drug discovery.