GPU-accelerated multi-scoring functions protein loop structure sampling

Abstract

Accurate protein loop structure models are important to understand functions of many proteins. One of the main problems in correctly modeling protein loop structures is sampling the large loop backbone conformation space, particularly when the loop is long. In this paper, we present a GPU-accelerated loop backbone structure modeling approach by sampling multiple scoring functions based on pair-wise atom distance, torsion angles of triplet residues, or soft-sphere van der Waals potential. The sampling program implemented on a heterogeneous CPU-GPU platform has observed a speedup of ~40 in sampling long loops, which enables the sampling process to carry out computation with large population size. The GPU-accelerated multi-scoring functions loop structure sampling allows fast generation of decoy sets composed of structurally-diversified backbone decoys with various compromises of multiple scoring functions. In the 53 long loop benchmark targets we tested, our computational results show that in more than 90% of the targets, the decoy sets we generated include decoys within 1.5A RMSD (Root Mean Square Deviation) from native while in 77% of the targets, decoys in 1.0A RMSD are reached.