Kernel-based partial least squares: application to fingerprint-based QSAR with model visualization.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are crucial for predicting biological activity but often function as "black boxes."
• A need exists for interpretable QSAR models that can be visualized within chemical structures.
• Existing methods lack direct visualization of model contributions at the atomic level.

Purpose of the Study:

• To develop interpretable QSAR models by combining kernel partial least squares (PLS) with Canvas 2D fingerprints.
• To enable visualization of QSAR model predictions directly onto chemical structures.
• To validate the approach using diverse protein-ligand binding affinity data.

Main Methods:

• Integration of direct kernel-based PLS with Canvas 2D fingerprints.
• Projection of developed QSAR models onto atomic representations of chemical structures.
• Validation using binding affinities across 10 protein targets from 7 protein families.

Main Results:

• Predictive QSAR models (test set Q(2) > 0.5) were successfully developed for 6 out of 10 datasets.
• Canvas 2D fingerprints consistently outperformed classical physicochemical and topological descriptors.
• A bootstrapping technique provided reliable uncertainty estimates for predictions.

Conclusions:

• The developed method provides interpretable QSAR models, bridging the gap between modelers and chemists.
• Atomic-level visualization aids in identifying favorable and unfavorable structural characteristics for activity.
• This approach offers a powerful tool for drug discovery and optimization.