Project Summary A scalable computational quantum mechanics method for non-covalent protein-ligand interactions will be developed based on "extended" symmetry-adapted perturbation theory (XSAPT), a cubic-scaling, fragment-based approach that is specifically designed for large supramolecular complexes, and which affords a demonstrated accuracy of ;$ 1 kcal/mo! with respect to the best-available ab initio benchmarks. In Phase I of this work, we will enhance the efficiency of XSAPT via better parallelization that will enable routine application to protein-ligand models containing 300+ atoms, using only modest computational resources. A bootstrap procedure will be developed to assess the accuracy of the method and a data set will be generated that includes protein-ligand interaction energies and their components: electrostatics, steric repulsion, dispersion, polarization, and charge transfer. The data set will build upon standard ones derived from crystal structures but will also include nonequilibrium structures as well as small ligand fragments for which crystal structures and other experimental data are not available; the latter are representative of fragment-based drug discovery strategies. These are challenging cases for interaction energy computations that can only be addressed quantitatively by using the predictive power of quantum mechanics, not by empirical scoring functions or by fits to experimental data. In Phase II, this data set will be used to train a machine learning (ML) model that is capable of ranking-ordering ligand binding energies in a reliable and quantitative way, something that existing scoring functions ( even those based on ML) cannot do. Additional Phase II work will integrate the ML-XSAPT scoring function into virtual drug-discovery workflows (including flexible docking protocols), which will facilitate both lead generation and lead optimization in drug discovery, based on quantitative ab initio energetics.