A Novel Automated Screening Method for Combinatorially Generated Small Molecules.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Enumerating chemical spaces for drug design requires efficient methods to distinguish feasible from infeasible molecules.
• Current screening techniques, such as 2D heuristics and 3D force fields, struggle to balance speed and accuracy.

Purpose of the Study:

• To develop an automated, fast, and accurate approach for screening small molecules.
• To improve upon existing methods for assessing molecular feasibility in computational drug design.

Main Methods:

• Computed 2D descriptors for feature encoding of molecules.
• Generated 3D-based feasibility targets on a subset for classification.
• Trained a machine learning model to predict 3D-based feasibility using 2D descriptors.
• Screened the full set of molecules using the trained predictive model.

Main Results:

• The developed approach is approximately 8 times faster than 3D simulation-based screening.
• Achieved high accuracy comparable to 3D methods without significant sacrifice.
• Demonstrated superior accuracy and coverage of feasible molecules compared to 2D-based pruning rules.

Conclusions:

• The automated machine learning approach offers a significant improvement in speed and accuracy for small molecule screening.
• This method enhances the efficiency of exploring chemical spaces in drug design.
• Full automation is achievable once topological features and 3D conformer evaluation are standardized.