Prediction of protein secondary structure using multilayer feed-forward neural networks

Important progress has been achieved in predicting secondary structure of protein sequences using artificial neural network recently. However, most of the models they used were BP networks with single hidden layer. In this paper, we try to use feedforward neural network involving more hidden layers to train and test the data set. While it has better generation ability and higher accuracy rate than the network with single hidden layer, the time complexity to train the model is often high. Hence, we utilize the contrastive divergence algorithm to solve this problem, which give better initial values to the weights in the network. Then we adjust the weights in turns. Experiments show that our train strategy is more efficient than BP algorithm.