The Parallel Analysis Of Protein Structure Based On Back-Propagation Neural Network

With the rapid development of bioinformatics and the completion of the human genome project, a large number of DNA and protein sequence information have been obtained, so the prediction of protein structure and function has become the main research direction.Back propagation is a multi-layer feed forward neural network, because of the non-linear mapping ability,BP neural network is widely used in the fields of bioinformatics, medical science and informatics.This paper introduces the application of improved BP algorithms in the prediction of protein secondary structure.Base on the idea of dataparallel and structureparallel, techniques are presented for the parallel implementation of BP neural network in the shared memory environment and distributed memory environment. The main researches of this paper are as follows:(1) The application of improved BP algorithms in the prediction of protein secondary structure. As the BP algorithm has some shortcoming in training, the paper introduces the batch learning mode to improve the convergence rate of neural network. Also, kinetic momentum method and learning rate self-adaptation adjusted strategy are employed to avoid the “oscillation phenomenon” and the local minimum. The precision of the protein secondary structure prediction is promoted to 73%.(2) Dataparallel and structureparallel algorithms are presented for the parallel implementation of BP neural network. The paper introduces the dataparallelism algorithm which parallelize the training of neural network with a shared error buffer and the structureparallel algorithm which parallelize the calculation of hidden layer with a shared output buffer. The experimental results show that the speed up rate of the algorithm is 7.551 in the 8 cores computing environment.(3) A hybrid mode parallel algorithm is presented for the parallel of BPneural network. The paper presented a hybrid mode parallel algorithm which based on both data parallel and structure parallel. Faced with massive dataset and large scale network, the algorithm can achieve better result. The experimental results show that the speed up rate of the hybrid mode parallel algorithm is 7.206 in 8 cores computing environment.(4) Based on the idea of distributed parallel and the hybrid mode parallel algorithm, a parallel strategy is presented for the prediction of protein secondary structure by using BP neural network in the distributed memory environment. The experimental results show that the speed up rate of the parallel strategy is 12.985 in the distributed memory environment.