(197ao) Selective Agonist Design Targeting Cannabinoid Receptor 2

Cannabinoid receptors (CB1 and CB2) are important drug targets for inflammation, obesity, and other central nervous system disorders. However, drug design targeting these receptors has been a costly and time-consuming process with a high failure rate due to the off-target side effect. CB2 selective agonists can potentially treat pain and inflammation without the psychological side effects of CB1 agonism. In this project, with the combination of molecular simulation and data-driven modeling, we design selective molecules for CB2.
We hypothesize that the subtype selectivity of selective ligands can be explained by ligand binding to the conformationally distinct states between CB1 and CB2. ~700 µs of unbiased simulations are performed to study the activation mechanism of both receptors in the absence of ligands. Using neural network-based VAMPnets, we discretize the conformational ensembles of both receptors into metastable states. Structural and dynamic comparisons of these metastable states allow us to decipher a coarse-grained view of protein activation by revealing sequential conversion between these states. Specifically, we observe the difference in the binding pocket volume of different metastable states of CB1, whereas there are minimal changes observed in CB2. Docking analysis reveals that differential binding pocket volume leads to distinct binding poses and docking affinities of CB2 selective agonists in CB1. Only a few of the intermediate metastable states of CB1 shows high affinity towards CB2 selective agonists. On the other hand, all the CB2 metastable states show a similar affinity for CB2 selective agonists, explaining these ligands' overall higher affinity towards CB2. Furthermore, we performed a data-driven approach for designing CB2 selective ligands. The selectivity of the selective CB2 agonists is decided based on the 100-fold difference of the predicted binding affinity (pKi>2) between CB1 and CB2. The functionality of these molecules is determined based role of the molecules in activating the receptor (agonist/antagonist). We create a predictive data-driven workflow for selectivity 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 pre-trained on all class A GPCR ligands (~160K molecules). Based on the stochasticity of the latent space, a data augmentation strategy is used to train 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 regression and classification. The overall performance (R2 score) of our regression models is more than 0.96. With this network, the entire ZINC20 and ChEMBL databases will be screened. The screened molecules will be further validated by docking into CB1 and CB2 metastable states obtained from our simulations. Overall, with this computational study, we assist the CB2 drug discovery process by mechanistically explaining the subtype selectivity of CB2 selective ligands and designing a workflow for predicting selective molecules.