Artificial Intelligence for Virtual Screening and Drug Discovery: Integrated Machine Learning and Molecular Electron Density approach

Virtual screening and docking-scoring methods are essential in preselecting hit molecules for pharmacological targets among millions of compounds. Structure-based methods aim to select the best potential hits for in vitro assays. Docking and scoring remain currently a major challenge in retrieving the true ligands, in terms of pose and ranking. To address these issues, the goal of our project is to develop novel accurate empirical scoring functions, relying on physics-based terms to predict protein-ligand binding. The approach will combine chemically realistic descriptors of the interacting charge distributions and advanced machine-learning methods for scoring functions training. Multipolar atoms databases, compared to quantum chemical approaches, allow fast reconstruction of molecular electron distributions, well beyond point charge approximation. Furthermore, we have developed an efficient tool to polarize the transferred electron density. Accurately modelled electrostatic, induction, van der Waals energies, entropy and solvent effects will be incorporated in the machine learning process. The resulting original Artificial Intelligence (AI) tool for accurate binding free energy prediction will be implemented and automated in our MoProViewer software. The methodology will be applied on a promising anticancer target on which we have a strong expertise. This target can be inhibited both for its enzymatic and for its binding to a protein partner to affect cell proliferation. In vitro screening will be applied on both catalytic and PPI sites to cross-validate the in silico predictions and to identify new valuable small molecules. In silico screening of will be performed using targeted libraries for either PPI or enzymatic inhibitors. The most promising hits will be further characterised by biophysical, biochemical and structural studies