created with babel and the Rosetta python app molfile_to_params, as follows: babel ipdb molecule.pdb opdb molecule.sdf molfile_to_params.py c nKHR 19088077 pmol molecule.sdf The Rosetta command line used to generate a set of rays that define a protein pocket topography is as follows: make_rayfiles.linuxgccrelease iinput_protein_ file 2YXJ.pdb central_relax_pdb_num 105 The Rosetta command line used to run DARC on a GPU using these input files is as follows: 5 Fast Docking on GPUs via Ray-Casting iterations. We marked the top-scoring 10% of the library as “hits,”then asked how many of these “hit”compounds would remain in the top 10% if docking was carried out using a reduced number of iterations and particles. We found that 94 of the 100 hit compounds were recovered in the topscoring 10% using our “typical use”parameters of 200 particles and 200 iterations, with little benefit associated with more extensive sampling. We MedChemExpress 2883-98-9 therefore carried forward these values for the further studies described below. DARC speedup on Graphics Processing Units All timing comparisons described below were carried out using a GeForce GTX 580 GPU, which can run 1024 threads concurrently, and a Dual Intel Xeon E5-2670 CPU using one thread. As a first timing benchmark, we evaluated the time needed to carry out docking using the same model system described earlier: a single conformer of ZINC00057615 docked against a pocket on the surface of the protein Bcl-xL. Based on our typical grid spacing and the size of the surface pocket we would typically use about 7,000 rays to describe this pocket; for benchmarking, we instead reduced the grid spacing to generate 93,000 initial rays then varied the number of rays used in docking by generating subsets of this large collection. As expected, the time required to complete this calculation scales approximately linearly with the number of rays and the number of particles, 10037488 whether carried out entirely on a CPU or with the help of a GPU. While the scaling is similar, however, the calculations are completed much more quickly using the GPU: in a typical uses case, the CPU takes 93 seconds to carry out the calculation and the GPU takes 3.4 seconds, corresponding to a 27-fold speedup. Similar behavior is observed when docking a single conformer to a surface pocket at the functional site of another protein, Mdm2. Due to the different size and shape of this pocket, the same grid spacing would lead to only 3,000 rays to describe this protein surface. Under these conditions, the calculation would take 47 seconds using the CPU alone, or 3.2 seconds using the GPU. We next tested the scaling of time with regards to the number of atoms in the ligand, docking to Mdm2 using 5,000 rays and 200 particles. We used a series of ligands containing 20, 25, 30, 35, and 40 non-hydrogen atoms. We find that the time required for this calculation on the CPU alone is not linearly related to the number of ligand atoms, because the geometry of the ligand dictates how much of the calculation can be avoided through the “ray elimination”step. In all cases, carrying out this calculation using the GPU results in a speedup of about 25-fold. While the typical-use speedup in the examples here is dramatic, we note that these data in fact downplay the true difference stemming from the use of the GPU for these calculations. In the timings we have reported above, the algorithm carried out on the CPU includes the “ray elimination”step that reduces the number of potential ray-atom