GPU Accelerated Protein Docking Software with Flexible Refinement

Herbordt, Vajda, and Kozakov groups, Acpharis Inc (Faculty startup by Vajda)

Supported by:
NIH R41GM101907-01 “GPU Accelerated Protein Docking Software with Flexible Refinement” (PI: Herbordt)

Based on the results of CAPRI (Critical Assessment of Predicted Interactions), a worldwide protein docking competition, PIPER, developed in the Vajda lab, is among the best protein-protein docking programs currently available. A major problem is that the flexible refinement of the PIPER-generated structures requires major computational resources. The general goal of this research is to develop efficient flexible refinement methods, and to implement the computationally expensive steps on GPUs. The refinement will employ two novel algorithms. First, given a putative interface defined by a cluster, we have developed  a program to identify the “key” variable side chains in the interface and their potential conformational states. Second, we are in the process of developing a Monte Carlo minimization algorithm for flexible refinement, which combines search in the space of the selected side chain rotamers with rigid body minimization based on manifold concepts. In addition to the use of more powerful flexible refinement algorithms, our goal is to further speed-up the calculations by implementing the time consuming components of both docking and refinement on GPUs. Profiling the algorithms we have found two such components, namely (1) correlation calculations that use fast Fourier transforms (FFTs) in docking, and (2) the non-bonded energy evaluation in the flexible refinement step. For the docking step we perform rotation and grid assignment on the CPU while the FFT and filtering are computed on the GPU. Acceleration of the energy evaluation steps requires changing the underlying data structures and statically mapping the work onto GPU threads in a way that allows parallel energy evaluations. The algorithm will also use the graph-theory based nonbonded list generation algorithm currently developed in the Vajda lab. With the above algorithmic and architectural speed-up, we can expect that a docking and refinement problem that previously required several hours on a 128 CPU cluster will be solved in the same amount of time by a single CPU and 2 NVIDIA Fermi GPU cards. Such a system can currently be assembled for $3500.