Fig. 1: A deep learning approach to denoise ATAC-seq data.

a Schematic of the ResNet architecture. The zoomed-in region displays a residual block composed of 1-dimensional convolutional layers (green squares), nonlinear ReLU activation functions (purple squares), and a skip connection. b Schematic showing how to train and validate AtacWorks on subsampled bulk ATAC-seq data. Clean high-coverage bulk ATAC-seq data is subsampled to create noisy data. Matched pairs of clean and noisy data are used to train AtacWorks models, which are then applied to denoise and call peaks from subsampled noisy data derived from a different cell type. c ATAC-seq signal tracks near the erythroblast marker gene GYPA, for four cell types used to train an AtacWorks model (gray), high-coverage erythroblast data (50 million reads; black), and erythroblast data subsampled to 0.2 million reads before (blue) and after (green) denoising with AtacWorks. Red bars below the zoomed-in tracks show peak calls by MACS2 (for the 50 M and 0.2 M read tracks) and AtacWorks (for the denoised track). d Pearson correlation between a clean ATAC-seq signal track (50 million reads) and subsampled data for erythroblasts, before (blue) and after (green) denoising with AtacWorks. Solid lines show correlation over the genome; dotted lines show correlation over chromosome 10. e AUPRC for MACS2 (blue) and AtacWorks (green) showing their peak calling performance on subsampled data, using peaks called by MACS2 subcommands on the clean (50 million reads) signal track as ground truth. Solid lines show AUPRC over the genome; dotted lines show AUPRC over chromosome 10. AUPRC: Area Under the Precision-Recall Curve. AUROC: Area Under the Receiver Operating Characteristic. MSE: Mean Squared Error. ReLU: Rectified Linear Unit. Source data are provided as a Source Data file.