Fig. 2

Residual non-negative matrix factorization (NMF) of distinct immune cell types from single-nucleus RNA-seq adipose tissue using non-negative least squares (NNLS). The panels on the left (A–D) show the deconvolution performance of NNLS in random-proportioned pseudobulks with A zero cell types missing, B one cell-type missing, C two cell types missing, and D three cell types missing from the NNLS deconvolution cell reference. Pearson’s correlation (r) and root mean square error (RMSE) are noted in each plot between the real and calculated proportions. Panels on the right show each missing cell type’s proportions (real proportions in pseudobulks) correlated with each residual’s NMF factor across the number of missing cell types; E reference with one cell type missing, F two cell types missing, and G three cell types missing. Each row represents one of the missing cell types, and each column represents each normalized NMF factor for each panel. The color represents Pearson's correlation (see color bar), and RMSE is noted in each box