To test the applicability of this computational workflow for single-cell analysis, we set out to measure the robustness of chromVAR outputs to data downsampling. To do this, we applied chromVAR to bulk ATAC-seq data from a deeply-sequenced set of hematopoietic cell types8, and compared the results of the analysis for the data across various degrees of downsampling. We found that the TF motif deviations using 106 to 5×103 fragments per sample are highly correlated to those determined using the full data set (Fig. 1b, Supplementary Fig. 7). The clustering accuracy using the bias corrected deviations is also largely preserved after downsampling, and compares favorably to clustering using PCA or other peak-based approaches (Supplementary Fig. 7; see methods).
