Extended Data Fig. 8: BayesPrism accurately recovers the heterogeneity in the expression of macrophages.

(a) UMAP visualization shows the expression of individual macrophages in the pseudo-bulk GBM28 dataset. The expression profile inferred by BayesPrism, shown as △, and the averaged expression profile from scRNA-seq for each patient, shown as ○, are projected onto the UMAP manifold. (b) Scatter plot shows Pearson’s correlation of reads summed across macrophages in the pseudo-bulk (ground truth) and gene expression deconvolved by BayesPrism (red) or undeconvovled pseudo-bulk (blue) as a function of the fraction of macrophages in the simulated pseudo-bulk (N=1,350). The correlation was computed using variance-stabilizing transformed reads. (c-f) Heatmap shows the pairwise Pearson correlation matrix between gene expression computed for each pair of simulated pseudo-bulk samples with macrophage fractions greater than 20% (c, e) and 50% (d, f). Simulated samples were obtained by drawing a random proportion of each cell type from the GBM-28 dataset, while sampling macrophages from an individual macrophage sub-cluster. Simulated pseudo-bulk samples are grouped by hierarchical clustering, as shown by the dendrogram. (c-d) Correlations estimated using the total gene expression without any correction. (e-f) Correlations over the same set of samples and over the same set of genes as in c and d, but using BayesPrism deconvolved expression profiles for macrophages in each sample.