d over the corresponding calculation on a CPU 14192894 alone. Methods Virtual screening using DARC An overview of the intended DARC workflow for virtual screening is diagrammed in rotatable dihedral 
angles as degrees of freedom. Finally, these minimized complexes are reranked on the basis of energetic considerations as well as structural considerations. The top scoring compounds can then be advanced for further characterization in biochemical or cell-based assays. Since DARC scoring considers solely shape complementarity, the intended use of DARC is not as a standalone tool for predicting binding free energies, or even for predicting whether any particular compound is likely to bind the target protein. Rather, DARC is intended to provide a fast, low-resolution tool for identifying the likely binding mode of a compound. Our intended workflow thus separates the extensive burden of sampling from the requirement of a detailed energy function to discriminate active from inactive compounds. This approach is in contrast to complementary methods such as RosettaLigand, which carries out detailed flexible-ligand docking via Monte Carlo simulations using the all-atom Rosetta energy function but is too computationally expensive 2578618 to enable routine screening of large compound libraries. Scoring with DARC DARC starts from a PDB file of a NVP-BKM120 site protein conformation, either from an experimentally derived structure or from biased “pocket optimization”simulations. The shape of a surface pocket is defined using a grid-based method described in detail elsewhere. Briefly, a grid is placed over the protein surface of interest. Based on the coordinates and radii of the atoms comprising the protein, grid points are marked either “protein” or “solvent”. Solvent points which lie on a line between two protein points are then marked as “pocket”; this approach was originally used in the LigASite software. The pocket “shell”is identified as those pocket grid points in direct contact with the protein. Additional grid points are then added around the perimeter of the pocket shell, used to mark regions outside the pocket where ligand binding will not lead to favorable interactions. The direction from the pocket center of mass to the protein center of mass is defined, and a point 30 A along this direction is defined as the origin from which rays will emanate. The angles and the distances expressing each of the shell points and forbidden points in spherical coordinates are calculated and saved. The number of shell points and “forbidden”points that define the pocket and thus the number of rays depends both the grid spacing and on the size of the surface pocket. In a typical use case, approximately 7,000 rays are used to define the protein pocket. This collection of vectors serves as a mapping of the protein surface topography that should be complemented by a well-docked ligand; the protein conformation and grid points are not directly used in docking beyond this point. Given the position and orientation of a ligand to be scored, a series of rays are cast from the origin along each of the directions used to map the surface topography. For each ray, the distance at which the first intersection with the ligand occurs is calculated and subtracted from the distance at which the same ray hit the protein surface. Each ray contributes to the total score as follows: Fast Docking on GPUs via Ray-Casting This approach to scoring is notably different from commonlyused docking tools, each of which e