Explainable machine learning to identify cannabinoid 1 receptor agonists

Synthetic cannabinoid receptor agonists (SCRAs) are molecules that interact with the CB1 receptor and may exhibit greater psychoactive effect than CB1's natural agonists. Determining new SCRAs by conducting bioassays can be time-consuming and may or may not yield positive results. Recent development in machine learning has demonstrated its usefulness as a complementary tool for prioritised bioassay testing and thus, improving the chances of getting a positive hit. In this study, we explored the use of different machine learning algorithms to predict specifically if a molecule is an active CB1 agonist. We achieved median scores of 0.937, 0.814 and 0.933 in terms of accuracy, precision and recall, respectively, for a moderately imbalanced dataset with the minority class occupying 22% of the test dataset. We carried out extensive evaluation and identified limitations to our model. We also demonstrated the use of Shapley values for explainability which appeared consistent with traditional structure-activity relationship (SAR) studies that accurately pinpoints importance of linkages to CB1 agonists' activity.