For clustering samples using chromVAR results, highly correlated motifs were first removed and then one minus the pearson correlation of the bias corrected deviations was used as the distance matrix for input into hierarchical clustering. For clustering samples using PCA, PCA was performed on the log of the fragment counts for all peaks after normalization for the total number of reads in peaks, and clustering was performed on the euclidean distance between the first five principal components. Hierarchical clustering was performed with complete linkage, and the resulting dendrogram was cut into 13 groups (the number of cell types). Clustering accuracy was measured using normalized mutual information 27.
