Alignment free Dissimilarities for sequence classification

One way to represent a DNA sequence is to break it down into substrings of length L, called L-tuples, and count the occurence of each L-tuple in the sequence. This representation defines a mapping of a sequence into a numerical space by a numerical feature vector of fixed length, that allows to measure sequence similarity in an alignment free way simply using disssimilarity functions between vectors. This work presents a benchmark study of 4 alignment free disssimilarity functions between sequences, computed on their L-tuples representation, for the purpose of sequence classification. In our experiments, we have tested the classes of geometric-based, correlation-based and information-based dissimilarities, incorporating them into a nearest neighbor classifier. Results computed on three dataset of nucleosome forming and inhibiting sequences, shows that the geometric and correlation disssimilaritiess are more suitable for nucleosome classification. Finally, their use could be a valid alternative to the alignment-based similarity measures, which remains yet the preferred choice when dealing with sequence similarity problem