Prediction of ligand-receptor pharmacological activities using a combined docking and machine learning approach

G protein coupled receptors (GPCRs) form one of the largest families of proteins in humans, and are valuable therapeutic targets for a variety of different diseases. One central question of drug discovery surrounding GPCRs is what determines the agonism or antagonism exhibited by ligands which bind these important targets. Ligands exert their action via the interactions they make in the ligand binding pocket. We hypothesised that there is a common set of receptor interactions made by ligands of diverse structures that mediate their action. We reasoned that among a large dataset of different ligands, the functionally important interactions will be over-represented. To investigate this hypothesis, we assembled a database of ~2700 known {beta}2AR ligands and computationally docked them to multiple experimentally determined {beta}2AR structures, generating ca 75,000 docking poses. For each docking pose, we predicted all interactions between the atoms of the receptor and the atoms of the ligand. Using Machine Learning (ML) we identified specific interactions that correlated with the agonist or antagonist activity of these ligands, and developed ML-based predictors of agonist/antagonist activity with up to 90% accuracy. This approach can be readily applied to other GPCRs and drug targets beyond GPCRs.