Discriminative DNA classification and motif prediction using weighted degree string kernels with shift and mismatch

There has been a growing interest in discovery of significant patterns in biological sequences that correspond to some structural and functional feature of the bio-molecule, known as motifs and has important application in determining regulatory sites and drug target identification. Identification of motif is challenging because it exists in different sequences in various mutated forms. Despite extensive studies over the last few years this problem is far from being satisfactorily solved. In this paper, the problem of finding a given a motif of length l with up to m number of mismatches in a given set of DNA sequences using kernel based approach is addressed. This paper presents Weighted Degree kernel with Shift and extends it to incorporate mismatches, for use with SVMs in a discriminative approach for DNA sequence classification and motif detection. It provides a biologically relevant computational way to compare DNA sequences without relying on family-based generative models, such as Hidden Markov models. Training SVMs is computationally expensive when using large sized training samples. We use the suffix tree based mismatch tree data structure to train the SVM using a scoring scheme as a speedup measure during implementation. The proposed kernel based method recovers motifs from the DNA sequences without relying on background information and generative models even using very few training examples.