Fig. 2

Evaluation of cell type proportion estimation in pancreatic transcriptomic profiles. A, C, E Pearson correlation coefficient (PCC) between true and estimated cell type proportions in 100 pseudo-bulk samples, averaged over cell types, is shown for different deconvolution methods. Results are shown for the Segerstolpe-H (A), Segerstolpe-T2D (C), and Enge-H (E) datasets. Category labels of bar charts indicate the reference signature, with “True” indicating the true underlying signature that is normally not available during deconvolution, “Baron” indicating the Baron signatures, and “NL-x” indicating Baron signatures with noise added at level x. For “NL-x”, results shown are mean with 95% confidence interval from evaluations using 11 variants of the Baron signature with noise added at level x. B, D, F, H, J, L PCC for each cell type separately is compared between the two best methods for respective datasets, when using NL-4 signatures (B, D, F) or Baron signatures (H, J, L). G, I, K Estimated and true proportions in the 100 pseudo-bulk profiles are directly compared, for a single cell type from each dataset, and for the two best methods for that dataset. BEDwARS performance is more robust to noise than the other methods in all datasets. All methods have comparable performance when the true signature is used. For the Baron signature, the performance of BEDwARS is similar to other methods, with a noticeable improvement for Segerstolpe-T2D dataset. BEDwARS provides better estimates for almost all cell types in the NL-4 evaluations and for at least one cell type with the Baron signatures