Fig. 1

The workflow of our method. Our method composes 5 components which are sequence pre-process layer, bidirectional LSTM encoder, 2 convolutional neural network and prediction layer. a Sequence pre-process layer processes protein sequences into one-hot encoding matrix and position-specific scoring matrix. One protein sequence is converted into two matrix which are the inputs of following deep learning neural networks. b Bidirectional LSTM encoder takes an one-hot encoding matrix as input and processes the items in the matrix sequentially. This encoder includes two LSTM layers. One layer processes the matrix from beginning to end while another layer processes the matrix backwards from end to beginning. After two direction encoding, the original one-hot encoding matrix is encoded to 256 values. c Convolutional neural network. The convolutional neural network is used to extract features from pssm matrix and encoded matrix. Two identical neural networks are used in our method. One network learns pssm matrix and another one learns encoded matrix. In the network, 4 convolution layers are included to filter out main features and 3 maxpooling layers are inserted among the convolution layers to choose outstanding features. The number of kernels in the 4 convolution layers are 256, 128, 64 and 32 respectively, and the kernel size of each layer is 4 × 3, 3 × 3, 3 × 3 and 3 × 3 correspondingly. d). Prediction layer produces the possibility of each subcellular location. At the beginning of this layer, the outputs from previous two convolutional neural networks are concatenated together, then the concatenated matrix is flattened into one dimensional array. A fully connected network integrate those features together and Sigmoid function computes the corresponding possibility at each location. Based on the output possibility, the subcellular locations of a protein can be decided