Interpretation of Ligand-Based Activity Cliff Prediction Models Using the Matched Molecular Pair Kernel

Area of Science:

• Medicinal Chemistry
• Computational Drug Discovery
• Cheminformatics

Background:

• Activity cliffs (ACs) represent potent drug candidates with minor structural modifications causing significant potency changes.
• Accurate AC prediction is crucial for optimizing lead compounds and guiding drug discovery efforts.
• Existing predictive models often lack interpretability, hindering practical application in early-stage research.

Purpose of the Study:

• To develop interpretable molecular fingerprints for predicting activity cliffs.
• To enhance the interpretability of machine learning models used in activity cliff analysis.
• To compare the performance and interpretability of a model-specific approach against a model-independent method.

Main Methods:

• Development of interpretable matched molecular pair (MMP) fingerprints.
• Application of a support vector machine (SVM) model with an MMP kernel.
• Comparison of feature importance using a model-specific interpretation approach and SHapley Additive exPlanations (SHAP).

Main Results:

• The SVM-based interpretation approach successfully differentiated between activity cliffs and non-activity cliffs.
• SHAP assigned high weights to features not present in test instances, indicating limitations in this context.
• SVM-derived feature weights for specific MMPs aligned with established X-ray crystallographic binding data, demonstrating chemical intuition.

Conclusions:

• Interpretable MMP fingerprints and SVM-based interpretation offer a chemically intuitive approach to activity cliff prediction.
• This method enhances decision-making in early drug discovery by providing understandable insights into structure-activity relationships.
• The developed approach addresses the interpretability challenge in predictive modeling for drug development.