Explainable machine learning to identify cannabinoid 1 receptor agonists.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Pharmacology

Background:

• Synthetic cannabinoid receptor agonists (SCRAs) interact with CB1 receptors, potentially causing greater psychoactivity than natural agonists.
• Identifying novel SCRAs via traditional bioassays is often time-consuming and inefficient.
• Machine learning (ML) offers a promising complementary approach for prioritizing compounds in drug discovery.

Purpose of the Study:

• To develop and evaluate machine learning algorithms for predicting CB1 receptor agonist activity in molecules.
• To assess the performance of ML models on a moderately imbalanced dataset.
• To utilize explainability techniques to interpret model predictions and their relation to chemical structure.

Main Methods:

• Exploration of various machine learning algorithms for binary classification of CB1 agonist activity.
• Model training and validation using a dataset with a 22% minority class representation.
• Application of Shapley values for model interpretability and identification of key structural features.

Main Results:

• Achieved median prediction scores of 0.937 (accuracy), 0.814 (precision), and 0.933 (recall).
• Identified limitations of the developed predictive models through extensive evaluation.
• Shapley values demonstrated consistency with traditional structure-activity relationship (SAR) studies.

Conclusions:

• Machine learning models can effectively predict CB1 receptor agonist activity, aiding in the discovery of novel SCRAs.
• Explainable AI methods, like Shapley values, provide insights into the chemical drivers of CB1 agonism.
• This ML-driven approach enhances the efficiency of identifying potential therapeutic or illicit SCRAs.