Protein Secondary Structure Prediction Based On Neural Network

Proteins carry out most of the basic functions of life at the molecular level. The function of protein mostly depends on its tertiary structure. Prediction of protein structure is very important to understand the relationship between the structure and the function of protein in respect that biologic function of the protein largely depends on its spatial structure. Protein secondary structure prediction becomes the most important step of predicting the space conformation from protein molecule. Neural Networks (NNs) as a most important method of machine learning, has been widely applied to bioinformatics and had considerable impact yielding.To improve the secondary structure prediction precision, this paper develop two NN ensemble models. First for the structure and function of a region in a sequence may strongly depend on both upstream and downstream of the region. Dual-layer BP network is taken as individual network in ensemble models. The "pruning method" and "early stop method" are used to avoid overfitting during training. For increasing the diversity, random noises are added to train sets resampled by bagging method. Five different networks are trained solely and final results are determined by voting rule. The neural network ensemble has provided with a good performance. But secondary structure prediction with BP networks is based on a local fixed-size window of amino acids, which cannot obtain relevant information in distant regions of the protein. And Bidirectional Neural Network (BRNN) can use information from the past and the future which is very useful for analysis and predictions. Therefore, the other neural network ensemble model with BRNN is proposed in this paper. BRNN can overcome the drawbacks of local fixed-window approaches, and it is used as individual network of ensemble model to predict protein secondary structure. Considering the complicated structure, slow convergence and too much parameters of BRNN, an improved BRNN structure is provided by deleting its right and left hidden layer. At the same time, Resilient Backpropagation (RPROP) is applied to train the network. The simulation results show that the proposed model has better generalization ability and faster convergence speed.