Fig. 3

Development and benchmarking of CellSIUS. a Schematic overview of CellSIUS. Starting from an initial assignment of N cells in M clusters (i), within each cluster, genes with a bimodal distribution are identified (ii) and only genes with cluster-specific expression are retained (iii). Among the candidate genes, sets with correlated expression patterns are identified by graph-based clustering (iv). Cells are assigned to subgroups based on their average expression of each gene set (v). b, c Performance comparison of CellSIUS to GiniClust2 and RaceID3 in detecting cells from sub-clusters and their signatures. b Recall, precision, and true negative rate (TNR) with regard to the detection of rare cells in synthetic data when varying the number of rare cells from 2 (0.2%) to 100 (10%) c Recall, precision, and true negative rate (TNR) with regard to the detection of outlier genes (gene signature) in synthetic data when varying and the number of signature genes from 2 to 100