Fig. 12.

scDesign2 guides the choice of cell number in rare cell type detection, in the case where the average sequencing depth is kept as fixed. scDesign2 generates synthetic 10x Genomics data with thirteen cell numbers. Two rare cell type detection methods—FiRE and GiniClust2—are applied to each synthetic dataset to detect rare cell types. a t-SNE visualization of six synthetic datasets and identification results—true positive (TP), false positive (FP), false negative (FN), and true negative (TN) cells—of FiRE in each dataset. b t-SNE visualization of the same six synthetic datasets and identification results of GiniClust2 in each dataset. c Four identification accuracy measures by FiRE (precision, recall, F1-score, and AUPRC) vs. cell number. d Three identification accuracy measures by GiniClust2 (precision, recall, and F1-score) vs. cell number. In c and d, the results of the six cell numbers in a and b are marked as dots and in the top, and the cell number of the real dataset [83] is marked as vertical dashed lines. Whenever there is no line for a cell number, FiRE or GiniClust2 does not detect any rare cells or fails