Fig. 2

Evaluation of DCATS with multiple simulation datasets. A The UMAP plots of simulation data with 8 cell types. B The true proportions of different cell types in default setting (3 replicates in each condition across 30 runs; the 4 cell types with differential composition abundance have names started with “P”). C The boxplot of MCC, F1, AUC, and PRAUC of different methods in bootstrap sampling results with the default setting. D-G Comparing multiple methods with Splatter-simulated data by varying the number of replicates (D, E) and number of cell types (F, G). “bcANCOM-BC” indicates using bias corrected proportions estimated by DCATS as the input of ANCOM-BC [11]. The default number of cell types is 8, and the default number of replicates is 3. H The performance of each method in the special case. “DCATS” indicates the results of DCATS using total cell count as the normalization term. “dcats_autoRef” indicates using the reference group automatically detected by DCATS as the normalization term. “dcats_defRef” indicates using the know true reference group as the normalization term. I Detecting variable cell types with multiple covariates, for both categorical and continuous types. Data is simulated with Splatter. N.B., only DCATS and Milo support multiple covariates