LeNup: learning nucleosome positioning from DNA sequences with improved convolutional neural networks

Abstract

MotivationNucleosome positioning plays significant roles in proper genome packing and its accessibility to execute transcription regulation. Despite a multitude of nucleosome positioning resources available on line including experimental datasets of genome-wide nucleosome occupancy profiles and computational tools to the analysis on these data, the complex language of eukaryotic Nucleosome positioning remains incompletely understood.ResultsHere, we address this challenge using an approach based on a state-of-the-art machine learning method. We present a novel convolutional neural network (CNN) to understand nucleosome positioning. We combined Inception-like networks with a gating mechanism for the response of multiple patterns and long term association in DNA sequences. We developed the open-source package LeNup based on the CNN to predict nucleosome positioning in Homo sapiens, Caenorhabditis elegans, Drosophila melanogaster as well as Saccharomyces cerevisiae genomes. We trained LeNup on four benchmark datasets. LeNup achieved greater predictive accuracy than previously published methods.Availability and implementationLeNup is freely available as Python and Lua script source code under a BSD style license from https://github.com/biomedBit/LeNup.Supplementary information Supplementary data are available at Bioinformatics online.

CORENup: a combination of convolutional and recurrent deep neural networks for nucleosome positioning identification

“…In nucleosome-linker classification one of the most recent and effective networks is the LeNup network [ 25 ]. This network, as many deep learning systems, has a structure inspired by the Google Inception network [ 33 ].…”

“…non periodic and periodic features. For sure the automatic identification of nucleosome positions seems to have attracted several machine learning researchers, and very effective models have been proposed so far [ 23 , 25 , 31 , 32 ]. Among the most performing ones, we have to mention iNuc-PseKNC [ 23 ].…”

“…the representation for the input data involve a specific preprocessing stage. Conversely, other authors have used the simple one hot representation obtaining greater accuracy, using deep neural network classifiers [ 25 , 26 , 28 ]. Actually the model called Le-Nup [ 25 ] is the top performer.…”

“…Conversely, other authors have used the simple one hot representation obtaining greater accuracy, using deep neural network classifiers [ 25 , 26 , 28 ]. Actually the model called Le-Nup [ 25 ] is the top performer. In this work, we will compare CORENup and LeNup in terms of performance metrics, the complexity of the models, and elapsed computation times.…”

Identifying complex motifs in massive omics data with a variable-convolutional layer in deep neural network

“…This CNN network treats each convolutional kernel as an individual motif scanner, extracts the major signal by global max-or average-pooling, and predicts whether the input sequence binds to the protein in question by a simple logistic regression or a two-layer neural network on such signals from all kernels. Most, if not all, subsequent CNN models for DNA/RNA sequences have followed DeepBind's kernel-as-motif-scanner strategy and successfully handled various DNA/RNA problems with huge volumes of input datasets (e.g., Zhou et al [2018], Kelley et al [2018], Angermueller et al [2017], , Zhang et al [2018] ).…”