基于CPU和GPU异构的蛋白质分子半柔性对接算法优化

分子对接是预测蛋白质复合物的有效手段。对于分子对接算法的优化旨在加速分子对接效率，降低计算成本，以及充分发挥计算资源的利用率。本文主要采用3个方案对半柔性对接算法进行优化：（1）方案一在CPU端进行优化；（2）方案二在方案一的基础上，利用CUFFT的移植工具CUFFTW为方案一提供部分GPU并行接口；（3）方案三利用GPU并行架构，通过CPU和GPU的协同处理，利用纯并行计算接口进行优化。3种方案对PDB code分别为1PEE，1B6C，4HX3和2SNI的测试蛋白进行结合态和自由态的对接，求得的最小均方根偏差L_RMSD小于5 Å，满足了复合物结构预测竞赛要求的中等精度结构标准，验证了对接结果的正确性。最后在保证结果正确性的前提下，测试了不同蛋白在不同方案下的运行速率；在保证不同蛋白对接效率相同的前提下，以1PPE为例，比较了不同方案下的对接速率。实验结果表明在同等旋转步长并保证程序运行结果正确性的前提下，最终的优化效果可提速近10倍，有效改进了半柔性对接算法的运行速率。

Optimization of Semi Flexible Docking Algorithm for Protein Molecules Based on CPU and GPU Heterogeneous

Molecular docking is an effective method to predict protein complexes. The optimization of molecular docking algorithm is aimed at accelerating the efficiency of molecular docking, reducing the computational cost, and making full use of computer resources. This paper mainly optimizes the semi flexible docking algorithm using three schemes:(1) The first scheme is optimized on the CPU side. (2) The second scheme, on the base of the first scheme, uses the CUFFT transplantation tool CUFFTW to provide partial GPU parallel interface for the solution. (3) The third scheme is based on the compute unified device architecture (CUDA) parallel structure, which is achieved by a pure CUDA interface and the coordinated operation of CPU and GPU. The three schemes use four different proteins to test the bound docking and unbound docking. The proteins'' PDB code is 1PEE, 1B6C, 4HX3 and 2SNI, respectively. The correctness of test is verified based on the minimum root mean square deviation L_RMSD, and its value is less than 5 Å, which satisfies the CAPRI medium precision structural standard. Then, under the premise of ensuring the correctness of the results, the running speed of different proteins under different schemes is tested. Finally, under the premise of ensuring the same efficiency of different protein docking, we take 1PPE as the final object to point out the rate of docking under different schemes. Experimental results show that under the same rotation step and the correctness of the program, the final optimization effect can increase the speed of the algorithm by nearly 10 times, thus greatly improving the speed of the algorithm.