AbstractDespite considerable advances obtained by applying machine learning approaches in protein–ligand affinity predictions, the incorporation of receptor flexibility has remained an important bottleneck. While ensemble docking has been used widely as a solution to this problem, the optimum choice of receptor conformations is still an open question considering the issues related to the computational cost and false positive pose predictions. Here, a combination of ensemble learning and ensemble docking is suggested to rank different conformations of the target protein in light of their importance for the final accuracy of the model. Available X-ray structures of cyclin-dependent kinase 2 (CDK2) in complex with different ligands are used as an initial receptor ensemble, and its redundancy is removed through a graph-based redundancy removal, which is shown to be more efficient and less subjective than clustering-based representative selection methods. A set of ligands with available experimental affinity are docked to this nonredundant receptor ensemble, and the energetic features of the best scored poses are used in an ensemble learning procedure based on the random forest method. The importance of receptors is obtained through feature selection measures, and it is shown that a few of the most important conformations are sufficient to reach 1 kcal/mol accuracy in affinity prediction with considerable improvement of the early enrichment power of the models compared to the different ensemble docking without learning strategies. A clear strategy has been provided in which machine learning selects the most important experimental conformers of the receptor among a large set of protein–ligand complexes while simultaneously maintaining the final accuracy of affinity predictions at the highest level possible for available data. Our results could be informative for future attempts to design receptor-specific docking-rescoring strategies.