Abstract
Single-cell ATAC-seq (scATAC-seq) technology allows the interrogation of chromatin accessibility of individual cells. Dropout events occur while the sequencing data signals at some bona fide chromatin sites of individuals are not captured, and the curse of these dropouts in scATAC-seq data inevitably hinders downstream analysis. It remains a challenge to impute scATAC-seq data due to its high dimensionality, sparsity, and near-binarization properties. Herein, we propose IMATAC, a deep hierarchical network with denoising autoencoder for imputing scATAC-seq data in the form of peak by cell. The network embeds scATAC-seq data into a latent space by a deep hierarchical architecture at two different levels, including bottom level for local details and top level for global information, that helps to characterize the high-dimensional sparse scATAC-seq data. Besides, it is encouraged to learn to reconstruct the original scATAC-seq data from an artificially corrupted version through a denoising autoencoder, so as to acquire an ability to recover the missing values primarily relying on the cells under the same population with the help of a parallel multi-classifier. Using simulated and experimental data, the performance of IMATAC is demonstrated by a comparative analysis with the other competing methods. The results suggest that our method can achieve lower imputation errors, and benefit the downstream analysis, including heterogeneous clustering, differential analysis, and regulatory element discovery. Besides, the contributions of several important network modules in our IMATAC are investigated, and how well it can separate the dropout zeros from biological zeros are discussed.