Alignment-free methods implemented a Convolutional Neural Networks adopted from previous work with modifications. To identify the consensus features from different species, each sequence was first mapped to the feature space, F, to create the feature vectors according to a finite alphabet, A. After mapping, the similarity between feature vectors was calculated. We trained a simple CNN with a total of 8 hidden layers, including a one-dimensional convolution on top of the feature vectors. Consistent with the concept used in image processing, we also added zero padding around the feature vectors to get the same dimensions as input36, 37 so that no features would be shrinking. The maximum (Max) pooling layer was used in between to extract the sharpest features and reduce the complexity.38, 39 We applied 20 sliding filters with different kernel sizes to optimize performance. To prevent co-adaptation of the hidden layers and make the model more robust, we apply regularization by employing dropout layers.40, 41 Twenty percent of the hidden units were dropped out. Last, we added a dense layer with one as parameter to produce a single output node in the output vector.42, 43 Sources and versions of open-source software libraries and neural-network library are listed below.
