NGC | Catalog
Welcome Guest


For downloads and more information, please view on a desktop device.
Logo for AtacWorks


Deep learning toolkit for coverage track denoising and peak calling from low-coverage or low-quality ATAC-Seq data



Use Case




Latest Version



September 29, 2020

Compressed Size

689.65 KB


AtacWorks is a deep learning toolkit for coverage track denoising and peak calling from low-coverage or low-quality ATAC-Seq data.

The ability to measure biologically-meaningful changes in accessible chromatin using ATAC-seq depends on both the signal-to-noise ratio and the depth of sequencing coverage. Technical parameters such as the overall quality of cells or tissues, the nuclei extraction method or over-digestion of chromatin can result in attenuated measurements of accessibility. Importantly, these issues are exacerbated in single-cell experiments, where primary tissues may vary in quality and key cell types may be exceedingly rare.

A well trained AtacWorks model can accurately predict both chromatin accessibility at base-pair resolution (a coverage track), and the genomic locations of accessible regulatory regions (peak calls). We apply AtacWorks to subsampled low-coverage bulk ATAC-seq and show that AtacWorks improves the resolution of the chromatin accessibility signal and the identification of regulatory elements. Further, AtacWorks is able to denoise signal from cell types not present in the training set, demonstrating that our deep learning models learn generalizable features of chromatin accessibility. Below is an image of a noisy track (Blue), a clean track (Black) and a noisy track denoised by AtacWorks (Green). As can be seen, AtacWorks successfully denoises the 1 million read track to approximate the quality of a 50 million read track.

Model Architecture

The model architecture of Atacworks is based on Residual Neural Networks. To train an AtacWorks model, you need a pair of ATAC-Seq datasets from the same cell type, where one dataset has lower coverage or lower quality than the other. The model learns to predict both the clean coverage track and the positions of peaks in the clean dataset from the noisy coverage track.

Model Architecture