A Deep Learning Network for Exploiting Positional Information in Nucleosome Related Sequences

Abstract

A nucleosome is a DNA-histone complex, wrapping about 150 pairs of double-stranded DNA. The role of nucleosomes is to pack the DNA into the nucleus of the Eukaryote cells to form the Chromatin. Nucleosome positioning genome wide play an important role in the regulation of cell type-specific gene activities. Several biological studies have shown sequence specificity of nucleosome presence, clearly underlined by the organization of precise nucleotides substrings. Taking into consideration such advances, the identification of nucleosomes on a genomic scale has been successfully performed by DNA sequence features representation and classical supervised classification methods such as Support Vector Machines and Logistic regression. The goal of this work is to propose a classification method for nucleosome positioning that, differently from the proposed method so far, does not make any use of a sequence feature extraction step. Deep neural networks (DNN) or deep learning models, were proved to be able to extract automatically useful features from input patterns. Under this framework, Long Short-Term Memory (LSTM) is a recurrent unit that reads a sequence one step at a time and can exploit long range relations. In this work, we propose a DNN model for nucleosome identification on sequences from three different species. Our experiments show that it outperforms classical methods in two of the three data sets and give promising results also for the other