Fig. 4

Single-nucleus (snRNA-seq) and single-cell (scRNA-seq)-informed-pseudobulks comparing simulated missing cell types, and real adipose bulks deconvolved with these as reference. For our simulation experiments, pseudobulks were created from snRNA-seq adipose tissue cell expression with realistic proportions as observed in the snRNA-seq data and deconvolved with either scRNA-seq (2 cells missing) or snRNA-seq (no cells missing). See Supplemental Fig. 7C for details on expression and references used for this simulation. The left panels (A–C) show the correlation between the residual’s NMF factors and each of the missing cell types’ proportions across pseudobulks for A non-negative least squares (NNLS), B BayesPrism, and C CIBERSORTx. Plots in the left column represent pseudobulks made with no noise added, and those in the right column represent pseudobulks with noise added. Pearson’s correlation (r) is noted in each plot. Panels on the right (D–F) show real vs. calculated proportions for pseudobulks of realistic proportions in each of the deconvolution methods. The left column represents pseudobulks with no noise added, and the right column represents bulks with noise added. D NNLS, E BayesPrism, and F CIBERSORTx deconvolution were used. The top plot of each panel represents the deconvolution with no cells missing (same cells as present in pseudobulks), and the bottom plot represents the proportions with two cells missing (no adipocytes or mesothelial cells). The red line in each plot represents the regression fit line. Each plot has root mean square error (RMSE), and r is noted. G We deconvolved 43 real bulk adipose tissue samples, and calculated each residual using both the snRNA-seq (x-axis) and scRNA-seq (y-axis)—which we hypothesize contains no missing cells compared to bulk. The mean is calculated for each gene across samples for both residuals, and these values are compared in a scatterplot for adipocyte and mesothelial cell markers, along with CIBERSORTx barcode genes