Recurrent Deep Neural Networks for Nucleosome Classification

Nucleosomes are DNA-histone complex, each wrapping about 150 pairs of double-stranded eukaryote DNA. Several biological studies have shown that the nu- cleosome positioning influences the regulation of cell type-specific gene activities. In addition, bioinformatic studies have shown proof of sequence specificity in the DNA fragment wrapped into nucleosomes. The main consequence has been the adoption of sequence features representation for the automatic identification of nucleosomes on a genomic scale. In this work, we propose a recurrent deep neural network for nucleo- some classification. The proposed architecture stacks convolutional and Long Short- term Memories layers to automatically extract features from short and long-range de- pendencies in a sequence. The adoption of this network allows avoiding the feature extraction and selection steps while improving the classification performances. Results have been computed on eight data sets of three different organisms, from Yeast to Hu- man.