Most contemporary drug discovery projects start with a ‘hit discovery’ phase where small chemicals are identified that have the capacity to interact, in a chemical sense, with a protein target involved in a given disease. To assist and accelerate this initial drug discovery process, ’virtual docking calculations’ are routinely performed, where computational models of proteins and computational models of small chemicals are evaluated for their capacities to bind together. In cutting-edge, contemporary implementations of this process, several conformations of protein targets are independently assayed in parallel ‘ensemble docking’ calculations. Some of these protein conformations, a minority of them, will be capable of binding many chemicals, while other protein conformations, the majority of them, will not be able to do so. This fact that only some of the conformations accessible to a protein will be ’selected’ by chemicals is known as ’conformational selection’ process in biology. This work describes a machine learning approach to characterize and identify the properties of protein conformations that will be selected (i.e., bind to) chemicals, and classified as potential binding drug candidates, unlike the remaining non-binding drug candidate protein conformations. This work also addresses the class imbalance problem through advanced machine learning techniques that maximize the prediction rate of potential protein molecular conformations for the test case proteins ADORA2A (Adenosine A2a Receptor) and OPRK1 (Opioid Receptor Kappa 1), and subsequently reduces the failure rates and hastens the drug discovery process.