In this work, we developed and applied a computational procedure for creating and validating predictive models capable of estimating the biological activity of ligands. The combination of modern machine learning methods, experimental data, and the appropriate setup of molecular descriptors led to a set of well-performing models. We thoroughly inspected both the methodological space and various possibilities for creating a chemical feature space. The resulting models were applied to the virtual screening of the ZINC20 database to identify new, biologically active ligands of RORγ receptors, which are a subfamily of nuclear receptors. Based on the known ligands of RORγ, we selected candidates and calculate their predicted activities with the best-performing models. We chose two candidates that were experimentally verified. One of these candidates was confirmed to induce the biological activity of the RORγ receptors, which we consider proof of the efficacy of the proposed methodology.
Keywords: Machine learning; Nuclear receptors; QSAR; RORγ; Virtual screening.
© 2023 The Author(s).