Supplementary Figure 8: Enhancer cytometry with CIBERSORT provides robust classification of most human hematopoietic cell types.

(a–c) Leave-one-out cross-validation of CIBERSORT, comparing the predicted fraction to the synthesized (ground truth) fraction in MPPs (a), CLPs (b), and monocytes (c). For these in silico validations, one replicate of each normal cell type was excluded from use in generating the signature matrix using CIBERSORT. Random synthetic mixtures of these ‘left-out’ samples were then used to assess the ability of CIBERSORT to correctly predict cellular composition. (d) Overall classification performance of deconvolving synthesized mixtures for each hematopoietic cell type shown in Figure 3d. r squared value is derived from correlation of the CIBERSORT-predicted fraction with the synthesized ground truth fraction of 100 random permutations. (e,f) Correlation of deviation between MPPs and HSCs (e) and MPPs and CMPs (f) derived from Figure 3d. Higher correlation of observed deviations implies that the two cell types are more frequently misclassified for one another using CIBERSORT. (g) Heat map representation of all pairwise correlations for deviation as shown in e and f. Heat map color represents the r² value from the correlation of deviation between two cell types (Online Methods). (h) Enhancer cytometry performed without manual removal of peaks corresponding to chromosome X, TSS peaks, and cancer-specific peaks shows a lower correlation with ground truth data (r² = 0.91 as compared to r² = 0.95 with manual curation of peaks). (i) Enhancer cytometry of DNase I hypersensitivity data using a signature matrix derived from our ATAC-seq data.