MolSHAP: Interpreting Quantitative Structure-Activity Relationships Using Shapley Values of R-Groups

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Optimizing lead compounds is crucial in drug discovery.
• Current machine learning methods often lack interpretable explanations for predictions.
• Existing interpretation techniques may not align with medicinal chemists' fragment-based optimization principles.

Purpose of the Study:

• To develop an interpretation framework, MolSHAP, for quantitative structure-activity relationship (QSAR) analysis.
• To estimate the contributions of R-groups to molecular activity, aligning with fragment-based drug design.
• To provide a model-agnostic tool for interpreting and optimizing drug candidates.

Main Methods:

• Proposed MolSHAP, an interpretation framework estimating R-group contributions to molecular activity.
• MolSHAP is model-agnostic, handling various input formats and model architectures.
• Evaluated MolSHAP on an R-group ranking task using representative activity regression models.

Main Results:

• MolSHAP demonstrated significantly superior interpretation power compared to existing methods.
• Developed a compound optimization algorithm leveraging MolSHAP's insights.
• Validated the reliability of MolSHAP-optimized compounds through an independent case study.

Conclusions:

• MolSHAP provides a valuable tool for accurate QSAR interpretation.
• The framework facilitates rational optimization of compound activities in drug discovery.
• MolSHAP bridges the gap between machine learning predictions and medicinal chemistry practices.