Fig. 4

Architectures of neural networks used in this study. The CNN is made of two convolution stacks (convolution layer + maxpooling). A filter size of five is used in the first convolution stack to serve as a motif detector. Thereafter, we used a larger filter size (32) in the next convolutional layer to capture larger patterns in the sequence. Following the convolution stacks, the features are flattened and batch normalized before passing through two dense layers using the ReLu activation function which are connected by a drop out layer. Finally, the output from the dense layer is passed to an output layer with a sigmoid activation to produce a final prediction value. Similar to the CNN model, we first used a convolution layer with a filter size of five to serve as a local motif detector for our CNN-LSTM model. After maxpooling, the output matrix is passed to an LSTM with 32 cells. Thereafter, the output from the LSTM is batch normalized and passed through two fully connected layers with the same configuration as our CNN model