Fig. 2

Cell mixture design, characterization, and analysis by OLS. A Breast cancer cells (BT474, T47D, and MCF7), leukemia cells (THP1 and Jurkat), and human mesenchymal stem cells (hMSCs) were used to generate six populations of mixed cells (cell mixtures). Each cell line was profiled individually by bulk RNA-seq in triplicates, and each mixture was profiled by bulk RNA-seq and flow cytometry in triplicates as well as by a 10 × genomics Chromium controller. B Cell mixtures were composed of varying proportions of cancer cells, with leukemia cells accounting for 15% and hMSC accounting for 0.5% (M1 and M4) to 2% (M3 and M6) of each mixture. C The clusters derived from scRNA-seq data corresponded to composing cell types, as identified by cell-type biomarkers. D The integration of scRNA-seq profiles of our mixtures (in gray) and scRNA-seq profiles of BT474, T47D, MCF7, BT483, AU565, HCC70, and DU4475 (Gambardella et al., 2022) revealed a significant overlap between profiles of BT474, T47D, and MCF7 cells, while negative controls, including BT483, AU565, HCC70, and DU4475, clustered separately. E Cell counts at the time of mixture generation were significantly correlated with cellular composition estimates by flow cytometry (r = 0.97) and F by scRNA-seq analysis (r = 0.96). However, the correlation between the estimates by flow cytometry and scRNA-seq was significantly lower (r = 0.92, p < 0.05, Fisher’s transformation). G Ordinary least squares regression (OLS) using bulk RNA-seq profiles of composing cell types estimated the composition of our mixtures with high accuracy (r = 0.95). H OLS deconvolution abundance estimates using cell-type expression profiles from scRNA-seq analysis were also accurate (r = 0.72, p < 1E − 4) but significantly worse (p < 1E − 5, Fisher’s transformation)