(84bj) Data-Driven Design of Selective Partial Agonist for Cannabinoid Receptors

Drug discovery is a time-consuming and expensive process consisting of a series of steps from target identification to the final approval. The molecule space of potential drug candidates is estimated to be more than 10⁶⁰. However, screening the entire space with existing computational techniques (e.g., docking) is impossible. Moreover, existing techniques suffer from low hit-rate due to a lack of force-field accuracy. In this work, using deep learning based data-driven modeling, our objective is to design selective partial agonists for cannabinoid receptor 1. These designed molecules will be able to tackle the selectivity and overstimulation issues of existing cannabinoid ligands. The selectivity of the selective partial agonists is decided based on the difference of the predicted binding affinity (pK_i) of cannabinoid receptors 1 and 2 (CB1 and CB2). The functionality of these molecules is determined based role of the molecules in activating the receptor (agonist/antagonist). Furthermore, partial agonism is decided based on the downstream efficacy of the molecules (E_max).

We create a predictive data-driven workflow for selectivity, functionality and partial agonism using the existing database of cannabinoid assays for binding and function. The latent space obtained from the junction tree variational autoencoder (JTVAE) is used to featurize the molecules for predicted models. This VAE is pretrained on all class A GPCR ligands (~160K molecules). Based on the stochasticity of the latent space, a data augmentation strategy is used for training of prediction models. This strategy is shown to improve the accuracy of the prediction results significantly. The random forest algorithm is used with ten-fold cross-validations for classification and regression. The overall performance (R²score) of our regression models is more than 0.96, and the accuracy of the classification models is above 99%.

Selective partial agonists are defined as molecules with more than a 100-fold affinity difference between CB1 and CB2 and being agonists with an Emax value of less than 70%. We screen around one billion compounds in the ZINC20 and ChEMBL databases using our strategy. Around 250 molecules are obtained from the screening process as selective partial agonists for cannabinoid receptors. Further analysis with interpretability shows features crucial for distinguishing partial agonists from unselective molecules. This assists us in generating molecules from pre-trained JTVAE moving along the critical features in the VAE latent space. The predicted and generated selective molecules will be tested experimentally by our collaborator. Ultimately, this research will lead to the discovery of new scaffolds for developing selective partial agonists for therapeutic drug development.