PMC

(a) Bar-plots represent the estimated expression profiles of 12 different genes within constituent cell types for each of the breast cancer intrinsic subtypes, as well as for the set of normal breast (control) samples. (b) Summary of main enriched gene sets among up- or down-regulated genes between cancer and normal breast in each cell type. See also .

(a) Enrichment of either OXPHOS or GLYCOLYSIS gene sets (hallmark gene sets MSigDB ()) among those up- or down-regulated in epithelial or stromal cells of breast cancer. Cell type specific differential expression analysis was performed with either by applying EDec to TCGA dataset, or in the LCM dataset. Dashed lines represent a p-value of 0.01. (b) Estimated stromal expression of either adipocyte or CAF markers across breast cancer subtypes. (c) Representative H&E staining images of matched tumor and normal breast samples from TCGA (TCGA-BH-A0B2). (d) Histogram of correlations between stromal expression of OXPHOS genes and stromal expression of marker genes of either adipocyte or CAF across breast cancer subtypes. (e) Histogram of correlations between epithelial expression of OXPHOS genes and stromal expression of marker genes of either adipocyte or CAF across breast cancer subtypes.

(a) Estimated versus true methylation levels for each constituent cell type and locus involved in the simulated mixtures dataset. (b) Estimated versus true proportions for each constituent cell type in each of the samples involved in the simulated mixtures dataset. (c) Estimated versus true methylation levels for each constituent cell type and locus profiled in the experimental mixtures dataset. (d) Estimated versus true proportions for each constituent cell type in each of the samples profiled in the experimental mixtures dataset. (e) Heat map representing the level of correlation between the estimated methylation profiles from the application of EDec to the targeted bisulfite sequencing dataset and the reference methylation profiles. Red boxes indicate the highest level of correlation for each estimated methylation profile. The estimated methylation profiles were labeled as cancer epithelial, normal epithelial, immune, or stromal based on what reference methylation profile was most correlated to each of them. (f) Proportion of constituent cell types estimated by EDec for samples in the targeted bisulfite sequencing dataset versus pathologist estimated proportions (H&E staining). Color key for all panels: orange (MCF-7), blue (HMEC), green (CAF), and red (T-cell).

(a) The EDec method has 2 main stages (Stages 1 and 2), preceded by a preparation stage (Stage 0). In Stage 0, a set of reference methylation profiles is used to select a set of genomic loci or array probes with distinct methylation levels across groups of references representing different constituent cell types. Methylation profiles of complex tissue samples over the set of loci/probes selected in Stage 0 are used as the input for the Stage 1 of the EDec method. In Stage 1, EDec estimates both the average methylation profiles of constituent cell types and the proportions of constituent cell types in each input sample using an iterative algorithm for constrained matrix factorization using quadratic programming. Stage 2 of EDec takes as input the gene expression profiles of the same tissue samples profiled for DNA methylation, as well as the proportions of constituent cell types for those samples, estimated in Stage 1, and outputs the gene expression profiles of constituent cell types. (b) Representation of the model associated with Stage 1 of EDec method. (c) Representation of the model used for gene expression deconvolution in Stage 2 of the EDec method.

(a) Heat map representing the methylation levels over the chosen set of array probes for the reference methylation profiles. (b) Heat map representing the correlation between the methylation profiles estimated by EDec and the reference methylation profiles. Red boxes indicate the highest correlation for each estimated methylation profile. (c) Scatterplot of EDec cell type proportion estimates for 9 TCGA samples based on targeted bisulfite sequencing (y-axis) and microarray (x-axis). (d) Scatterplot between EDec and pathologist (H & E) estimates of proportions of constituent cell types for a subset (six samples) of the TCGA dataset for which H&E staining-based estimates were available. (e) EDec estimated proportions of constituent cell types for samples in the TCGA dataset. Side bar represents separation of TCGA cancers samples into PAM50 expression subtypes. The red box highlights the samples best explained by the cancerous epithelial 2 profile which are almost exclusively classified as basal-like. (f) Kaplan-Meier plot indicating the significant difference in prognosis (p-value < 0.01) for patients within the group of samples best explained by the cancer epithelial 2 profile (red box in panel F; basal-like) with high versus low estimated immune cell type proportion. See also .