Improving classical substructure-based virtual screening to handle extrapolation challenges.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Substructure-based virtual screening (sSBVS) is a key computational approach in drug discovery.
• Existing sSBVS methods face challenges in extrapolating predictions to molecules dissimilar to the training set.
• Evaluating the extrapolation capabilities of traditional methods like Tanimoto coefficient (MTC) and Naive Bayes (NB) is crucial.

Purpose of the Study:

• To assess the predictive performance of classic sSBVS methods in extrapolation tasks.
• To introduce and validate a novel approach, the Shadow method, for enhanced extrapolation in sSBVS.
• To improve the detection of novel drug candidates with low similarity to known active molecules.

Main Methods:

• Evaluation of classic virtual screening methods: similarity searching using Tanimoto coefficient (MTC) and Naive Bayes (NB).
• Development of the Shadow approach, which accounts for inclusion relationships between molecular substructures.
• Discarding contributions from substructures that are included in ('shaded' by) other substructures within the molecule.

Main Results:

• Classic methods (MTC and NB) demonstrated better performance in interpolation than in extrapolation.
• The Shadow approach significantly outperformed MTC (pValue = 3.1 × 10⁻¹⁶) and NB (pValue = 3.5 × 10⁻⁹) in extrapolation tasks.
• The Shadow classifier successfully identified active molecules with low similarity to the training set.

Conclusions:

• Traditional sSBVS methods have limitations when applied to extrapolation scenarios in drug discovery.
• The Shadow approach offers a significant improvement for extrapolative virtual screening by considering substructure hierarchies.
• This method enhances the ability to discover novel chemical entities distant from known active compounds.