Classification of DNA Sequences with k-mers Based Vector Representations

Genes are biologically represented as a sequence of nucleotides which are labeled as Adenine, Thymine, Guanine, and Cytosine. DNA sequences, generally, need to be represented by numerical vectors to be able to use on machine learning algorithms, and how the sequences are represented affects the performance of sequence classification algorithms. In this paper, four different approaches (one-hot based with random and default dictionary, Voss and dna2vec) are used to represent DNA sequences of three different benchmark datasets (splice, promoter, and H3) and a deep convolutional neural network is used to see the effect of representation on classification. The random dictionary approach is presented as a new representation to show the effect of using equidistant vectors with a random order when using convolutional neural networks for classification. It is observed from the results that, representation approach which represents the similar subsequence with similar vectors is superior to other methods.