Normalized Maximum Likelihood Models with Memory for Genomics

This paper will review the recent applications of MDL techniques for genomic sequence analysis presented in detail in [12] and [13]. The normalized maximum likelihood (NML) model [5]-[8] for a class of Markov sources [13] was recently used for the compression of full genomes, obtaining for the human genome the best existing compression results [4]. We have shown that one of the underlying biological features implicitely uncovered by the compression algorithm is the existence of approximate gene duplication. We proposed a refined method based on the same NML models for the segmentation of DNA sequences for uncovering gene duplications [12]. Several analysis tasks in genomic sequences involve preliminary segmentation or clustering of the data, which can be performed by a number of techniques, based on various similarity measures. The process of sequence matching is used for solving the problem of uncovering gene duplications with the help of a preliminary segmentation of a complex DNA locus, known to have evolved through a series of duplications.