Discovering Highly Potent Molecules from an Initial Set of Inactives Using Iterative Screening

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Quantitative structure-activity relationships (QSAR) and similarity searching excel at interpolating compound activity within known ranges.
• Their performance in early-stage drug discovery, particularly extrapolating from inactive to active compound data (extrapolation), remains underexplored.

Purpose of the Study:

• To evaluate and compare the effectiveness of various computational methods for virtual screening in early drug discovery scenarios with limited active data.
• To identify the optimal strategy for identifying highly potent compounds through iterative screening when starting with a large set of inactive molecules.

Main Methods:

• An iterative virtual screening strategy was developed and tested on 25 diverse ChEMBL bioactivity datasets.
• Methods benchmarked included random forest (RF), multiple linear regression, ridge regression, similarity searching, and random selection.
• Performance was measured by the number of iterations needed to find a highly active compound.

Main Results:

• Linear and ridge regression methods demonstrated superior performance, reducing the iterations required to find active compounds by over 50% compared to RF and similarity searching.
• Simple regression models showed better extrapolation capabilities to high-bioactivity ranges than RF.
• Scaffold diversity influenced performance, with similarity searching and RF sometimes requiring twice as many iterations as random selection.

Conclusions:

• Linear and ridge regression are highly effective for early-stage drug discovery virtual screening when only inactive data is available.
• The developed iterative screening framework can be extended to multi-target drug discovery, as demonstrated with COX-1 and COX-2 data.
• This study provides a validated approach and identifies optimal computational setups for discovering potent compounds in data-scarce early drug discovery phases.