Fig. 4

Evaluation of cell type proportion and signature estimation from brain transcriptomic profiles. A, B Pearson correlation coefficient (PCC) computed between the estimated and true cell type proportions (A) or cell type signatures (B), averaged over cell types, when deconvolving 100 pseudo-bulk samples generated from Darmanis dataset. Category labels of bar charts indicate the reference signature used. BEDwARS and BayesPrism are similar and have higher PCC than the other methods in the estimation of cell type proportions for the IP signature and its noisy versions (NL-x), with the performance gap increasing as the noise level increases. For estimation of cell type signatures, RODEO provided with BEDwARS or BayesPrism-estimated proportions (RODEO/BEDwARS, RODEO/BayesPrism) outperform other methods including BEDwARS. C, D Average PCC between estimated and true cell type proportions (C) or signatures (D), using the IP signatures (same as in A, B) as well as the CA, NG, MM signatures. All methods perform comparably for proportion estimation when using the CA signature but BEDwARS exhibits better performance when the reference signature is more diverged from the true signature, such as NG (different region of human brain) and MM (mouse brain). All methods show comparable performance in signature estimation when provided the CA and IP references signatures, but RODEO provided with BEDwARS- or BayesPrism-estimated proportions exhibits superior performance for the more diverged reference signatures (NG and MM). E PCC for each cell type separately is compared between the two best methods (BEDwARS and BayesPrism) when using NG signatures. Both methods perform equally well evaluated by PCC criterion. F Estimated and true proportions in the 100 pseudo-bulk profiles are directly compared, for neurons (NEU) and astrocytes (ASTRO), for the two best methods when using the NG signatures. BEDwARS estimates are considerably more accurate in magnitude than BayesPrism estimates