Fig. 5.

Validation of EPISCORE in breast cancer. a Barplots displaying for a number of mammary epithelial cell lines, their estimated epithelial (basal + luminal), endothelial, and stromal fractions as inferred using the imputed DNAm reference derived from the breast SmartSeq2 single-cell atlas (left panel). Right panel as left, but for the imputed DNAm reference derived from the breast tissue 10X single-cell atlas, which also contained references for lymphocytes and macrophages. b Beanplots displaying the estimated epithelial fraction (using one of the two imputed breast DNAm reference matrices, as indicated) for breast cancer cell lines and cancer cell lines derived from other epithelial tissues. P values derive from a one-tailed Wilcoxon rank sum test. c Beanplots displaying the Pearson correlation coefficient (PCC) between the estimated and true cell type fractions for each of the cell types in the 10X-derived DNAm reference matrix. There is one PCC value for each of 100 different Monte-Carlo runs, and for each run, the PCC is obtained by comparing estimated to true cell type fractions for 100 in silico mixtures. d Barplots comparing the fraction of DMCTs (FDR < 0.05) between triple-negative and ER+ breast cancer tissue occurring in each cell type, as inferred using CellDMC and the imputed breast DNAm reference (SmartSeq2). The number of DMCTs is expressed as a fraction of the total number of CpGs assayed. e The sensitivity (SE) of CellDMC to capture a gold standard set of epithelial-specific DMCTs derived by comparing triple-negative and ER+ breast cancer cell lines, in the corresponding breast cancer tissue DNAm dataset. SE is displayed at two different FDR thresholds, as indicated. Green datapoints represent the results for typical Monte-Carlo runs where the breast cancer subtype labels were randomized. f Scatterplot of t statistics between triple-negative and ER+ breast cancer tissue for significant (FDR < 0.05) epithelial DMCTs, as derived with CellDMC and the imputed DNAm reference (x-axis) vs. the corresponding t statistic as derived by comparing triple-negative to ER+ breast cancer cell lines (y-axis). The odds ratio (OR) and P value derived from a one-tailed Fisher test, quantifying the agreement between the DNAm changes in the epithelial cells of the tissue and epithelial cell lines is given. Barplot to the right compares the odds ratio for CellDMC to correctly identify epithelial DMCTs to that of an ordinary linear model which calls differentially methylated cytosines without regard to cell type