基于BP神经网络的蛋白质结构并行分析

发布时间：2018-06-05 16:05

本文选题：蛋白质结构预测 + BP神经网络　；参考：《华南理工大学》2016年硕士论文

【摘要】：随着生物信息学的发展和人类基因组计划的完成,人们获得了大量的DNA和蛋白质序列信息,迫切需要效果更好的蛋白质结构和功能预测方法。本文采用BP神经网络算法对蛋白质二级结构预测进行研究,给出了面向共享内存环境及分布式内存环境下的BP神经网络并行方案。本文主要研究内容及成果如下:(1)给出了基于改进BP神经网络的蛋白质二级结构预测方法。该方法针对传统BP神经网络收敛速度慢、容易陷入局部极小值等问题,对网络的学习过程进行改进,采用批处理学习模式,提高网络收敛速度;基于梯度方法上进行优化,引入附带动量项和自适应学习率的学习规则,避免“振荡现象”和陷入局部极小值。实验结果表明,该方法对蛋白质二级结构的预测准确率达到73%;(2)提出面向共享内存环境的数据并行和结构并行算法。该方法在输出层加入共享误差缓存区,实现网络学习过程的并行化;在隐藏层加入共享输出缓存区,并对网络隐藏层结构进行划分,实现并行输出。实验结果表明,在8核的计算环境下,BP神经网络并行算法的加速比可达7.551;(3)提出面向共享内存环境的混合方式并行算法。该方法结合数据并行和结构并行思想,分别对隐藏层输出计算、输出层输出和误差计算进行并行化,在数据量较小或网络结构简单时,可以获得更好的加速效果。通过实验分析表明,在8核的计算环境下,BP神经网络混合方式并行算法的加速比达7.206;(4)给出了一种分布式内存环境下,应用BP神经网络进行蛋白质二级结构预测的并行策略。该策略采用基于主从模式的分布式并行算法,同时节点内采用面向共享内存环境的混合方式并行算法,实现BP神经网络算法的并行化。实验结果表明,在使用4节点、单节点4核的计算环境下,该并行算法的最大加速比达12.985。
[Abstract]:With the development of bioinformatics and the completion of the Human Genome Project, a large amount of DNA and protein sequence information has been obtained, and a better prediction method of protein structure and function is urgently needed. In this paper, BP neural network algorithm is used to study protein secondary structure prediction, and a parallel BP neural network scheme for shared memory environment and distributed memory environment is presented. The main contents and results of this paper are as follows: 1) the prediction method of protein secondary structure based on improved BP neural network is presented. Aiming at the problems of the traditional BP neural network, such as slow convergence speed and easy to fall into the local minimum, the learning process of the network is improved, the batch learning mode is adopted to improve the convergence speed of the network, and the optimization based on the gradient method is carried out. The learning rules with momentum term and adaptive learning rate are introduced to avoid "oscillating phenomenon" and falling into local minima. The experimental results show that the prediction accuracy of this method for protein secondary structure is 73 / 2) and a parallel data and structure parallel algorithm for shared memory environment is proposed. In this method, the shared error buffer is added to the output layer to realize the parallelization of network learning process, and the shared output cache is added to the hidden layer, and the structure of the network hidden layer is divided to realize the parallel output. The experimental results show that the speedup of the parallel algorithm of BP neural network can reach 7.551 / 3 under the computing environment of 8 cores) and a hybrid parallel algorithm for shared memory environment is proposed. Combining the idea of data parallelism and structural parallelism, the method parallelizes the hidden layer output, the output and the error calculation, and can obtain better acceleration effect when the data is small or the network structure is simple. The experimental results show that a parallel strategy for protein secondary structure prediction based on BP neural network under distributed memory environment is proposed. The strategy adopts a distributed parallel algorithm based on master-slave mode and a hybrid parallel algorithm for shared memory environment in the nodes to realize the parallelization of BP neural network algorithm. The experimental results show that the maximum acceleration of the parallel algorithm is Prida 12.985 under the computing environment of 4-node and single-node 4-core.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：Q51;TP183

【相似文献】