We plan to employ a dual prediction strategy. First, we will use a physics-based protocol with high-throughput docking with Schrödinger Glide (optionally combined with an active learning approach for higher throughput) or a Phase Pharmacophore model built based on the active site in WD40 domains. Further prioritization may be performed by MM-PBSA and absolute free energy calculations with Schrödinger FEP+. Second, we plan to run generative models driven by structure-based scoring functions (e.g. Glide docking score or Phase Pharmacophore model based on active site) to predict actives. These will then be used in a ligand-based approach to extract similar structures from Enamine Real/ZINC database. These compounds will then be scored again. The hit sets will be combined and then filtered by properties based on machine learning models and unwanted substructures. If necessary the hits will be clustered to select a diverse set for submission.