Predicting affinity and potency of new psychoactive substances at cannabinoid 1 receptor with explainable artificial intelligence.

Introduction: Binding affinity and functional potency are two distinct but pharmacologically related properties in ligand-receptor interactions. The molecular features influencing these two endpoints may differ, reflecting distinct physicochemical and conformational requirements for receptor binding and activation. Consequently, understanding the molecular features that determine both endpoints is essential, especially for the regulation of fast-moving markets of recreational drugs such as new psychoactive substances (NPS), where manufacturers are able to circumvent the stringent national and international regulations by exploiting loopholes within drug regulation. Synthetic cannabinoid receptor agonists form one class of NPS and are generally full agonists of the cannabinoid 1 receptor (CB1). Methods: We compiled publicly available data on CB1 affinity and potency. We used molecular descriptors and fingerprints to train five machine learning (ML) classification models. Using explainable Artificial Intelligence, particularly SHAP values, we analyzed the features driving affinity and potency. Results: Especially XGBoost (XGB) and Random Forest (RF) with molecular descriptors and Extended Connectivity Fingerprints (ECFP), respectively, outperformed the other models in predicting binding affinity. For potency, XGB and RF showed excellent performance with either molecular descriptors or ECFP. All these models had recall, precision, and F1 scores >90%. Further, we mapped SHAP values to selected NPS to highlight structural features having a positive or negative impact on high affinity and potency. Discussion: Affinity relies mainly on lipophilicity and membrane-partitioning descriptors, whereas potency depends on a broader combination of lipophilicity, shape, branching, and electronic descriptors.