Mixture models of geometric distributions in genomic analysis of inter-nucleotide distances

Abstract

The mapping defined by inter-nucleotide distances (InD) provides a reversible numerical representation of the primary structure of DNA. If nucleotides were independently placed along the genome, a finite mixture model of four geometric distributions could be fitted to the InD where the four marginal distributions would be the expected distributions of the four nucleotide types. We analyze a finite mixture model of geometric distributions (f_2), with marginals not explicitly addressed to the nucleotide types, as an approximation to the InD. We use BIC in the composite likelihood framework for choosing the number of components of the mixture and the EM algorithm for estimating the model parameters. Based on divergence profiles, an experimental study was carried out on the complete genomes of 45 species to evaluate f_2. Although the proposed model is not suited to the InD, our analysis shows that divergence profiles involving the empirical distribution of the InD are also exhibited by profiles involving f_2. It suggests that statistical regularities of the InD can be described by the model f_2. Some characteristics of the DNA sequences captured by the model f_2 are illustrated. In particular, clusterings of subgroups of eukaryotes (primates, mammalians, animals and plants) are detected.