Challenge #6

Our approach combines expertise of Kozakov Lab at Stony Brook and Tropsha Lab at UNC. Our workflow uses several complimentary modules for identification of high affinity hits for a given protein target with a known 3D structure. Identification of the binding site hot-spot information together with conventional structure-based virtual screening methods enhanvced by generative modeling are key enabling components of our hit selection approach. First, we will use FTMap, a computational mapping algorithm that identifies binding regions on the surface of the target protein with major contributions to the ligand binding free energy. FTMap samples all possible positions of small organic molecule probes and scores them using a physical energy function. The binding site regions that bind multiple probes identify the binding hot spots and the corresponding favorable chemical groups. In addition to hypothesis-free FTMap, we will use a different approach towards hot spot identification, LigTBM, that was inspired by structural similarity search methods. The basic idea is to match physico-chemical environment of the protein to the micro pockets containing small organic molecule probes extracted from PDB structures containing bound ligands.  Early version of LigTBM was the top performer by docking model accuracy at NIH sponsored D3R competition, and most recently in the CASP ligand experiment. This matching procedure will also provide us with possible fragment placement within the target protein, so the data will be presented in the same form as FTMap data, which will facilitate their comparative analysis and identification of consensus hot spots. Finally, we will employ a recently developed deep learning based extension of FTMap, which mimics the original FTMap but uses an improved scoring function during sampling, which is trained on fragment PDB data. The hot spot information will be used to create a pharmacophore model for the next stage of virtual screening. Using this model, we then will perform a pharmacophore-based screening of the entire Enamine REAL library (~48B with tautomers) to select a subset of the target specific compounds based on the fitting to our pharmacophore hypothesis. In addition, we will use pharmacophore models to bias our deep and reinforcement learning method termed ReLeaSE to generate target-specific novel hit compounds. These hit compounds will be employed as queries for another round of similarity searching against the entire Enamine library to identify purchasable compounds most similar to the generative hits. The combined set of pharmacophore- and generatve compound based virtual screening (typically, ~1M molecules) will be docked into the binding site using Glide by Schrödinger. The top scored docking hits then will be additionally prioritized using the hot spot information and the LigTBM type approach. These consensus hits will be nominated for the experimental testing.