Research On Prediction Algorithm Of Protein Secondary Structure Based On Neural Network

Understanding the potential relationship between the amino acid sequence of protein and its structure is one of fundamental problem in the structural Bioinformatics. Since the direct prediction of the three-dimensional structure of protein from the amino acid sequence is extremely difficult, an intermediate but effective step is to predict the secondary structure of protein. In particular, the prediction information of protein secondary structure not only contribute to determine the three-dimensional structure of protein, and but also can be used for protein function and interaction prediction.Neural network is the main model in deep learning. Since the neural network has strong nonlinear modeling capabilities, which has been applied widely in many domains, including speech recognition, language modeling, object recognition and image classification. Based on neural network model and the latest deep learning technologies,we mainly study the prediction of eight classes protein secondary structure, the main contents are as follows:For the prediction of protein secondary structure with the sliding window, a deep feed-forward neural network algorithm is proposed. The algorithm adopts the 5-layer feedforward neural network and uses the Dropout regularization to overcome overfitting. Experimental results show that the proposed algorithm can significantly improve the classification accuracy of eight classes protein secondary structure.For the prediction of protein secondary structure in sequential manner, a prediction framework based on recurrent neural network is proposed. The framework utilizes a bidirectional recurrent neural network to model local and long-range interaction between amino acid residues in protein.The hidden layer outputs of the bidirectional recurrent neural network is further fed to the three-layer feedforward neural network for the prediction of eight classes protein secondary structure. Experimental results show that the prediction accuracy of protein secondary structure of the bidirectional LSTM and GRU is significantly better than that of the bidirectional vanilla RNN.