Abstract
Motivation Nucleosomes form the first level of DNA compaction and thus bear a critical role in the overall genome organization. At the same time, they modulate chromatin accessibility and, through a dynamic equilibrium with other DNA-binding proteins, may shape gene expression. A number of large-scale nucleosome positioning maps, obtained for various genomes, has compelled the importance of nucleosomes in the regulation of gene expression and has shown constraints in the relative positions of nucleosomes to be much stronger around regulatory elements (i.e. promoters, splice junctions and enhancers). At the same time, the great majority of nucleosome positions appears to be rather flexible. Various computational methods have in the past been used in order to capture the sequence determinants of nucleosome positioning but, as the extent to which DNA sequence preferences may guide nucleosome occupancy largely varies, this has proved to be rather difficult. In order to focus on highly specific sequence attributes, in this work we have analyzed two well-defined sets of nucleosome-occupied sites (NOS) and nucleosome-free-regions (NFR) from the genome of S. cerevisiae, with the use of textual representations.
Results We employed 3 different genomic sequence representations (Hidden Markov Models, Bag-of-Words and N-gram Graphs) combined with a number of machine learning algorithms on the task of classifying genomic sequences as nucleosome-free (NFR) or nucleosome-occupied NOS (to be further amended based on updated results). We found that different approaches that involve the usage of different representations or algorithms can be more or less effective at predicting nucleosome positioning based on the textual data of the underlying genomic sequence. More interestingly, we show that N-gram Graphs, a sequence representation that takes into account both k-mer occurrences and relative positioning at various lengths scales is outperforming other methodologies and may thus be a choice of preference for the analysis of DNA sequences with subtle constraints.
