Figure 2. Evaluation of clustering accuracy.

a, Expression PCA of a synthetic mixture cells from five HipSci cells lines (n=7073 cells) with 5% ambient RNA and 6% doublets colored by known genotypes. Because these samples only contain one cell type, the largest remaining source of variation in the expression profile comes from the genotype, although the signal is not sufficient for accurate genotype clustering. b, Elbow plot of the number of clusters versus the total log likelihood showing a clear preference for the correct number of clusters (k=5). c and d, PCA of the normalized cell-by-cluster log likelihood matrix from souporcell (n=7073 cells). As this is a synthetic mixture in which we know the ground truth, we color by genotype clusters and highlight errors in orange (false positive doublets) and pink (false negative doublets). e, Expression PCA of a single replicate (see Fig. S1 for reps) of the experimental mixtures (n=4925 cells) colored by genotype clusters from souporcell. f, Elbow plot of the total log likelihood versus different numbers of clusters showing a clear preference for the correct number of clusters. g and h, PCA showing the first four PCs of the normalized cell-by-cluster log likelihood matrix colored by cluster (n=4925 cells). i, ROC curve of the doublet calls made by souporcell and vireo and a point estimate for scSplit (blue dot) for a synthetic mixture with 6% doublets 451/7073 and 10% ambient RNA. We show both the curves and the threshold chosen (points) for each tool. scSplit did not give a score so we simply show the point estimate. Demuxlet’s doublet probabilities were all 1.0 until the solid line starts, so we show a theoretical dotted line up to that point. j, Doublet call percentages for all tools on synthetic mixtures for varying amounts of ambient RNA versus the actual doublet rate (dotted line). k, Adjusted Rand Index (ARI) versus the known ground truth of synthetic mixtures with 6% doublets and a varying amount of ambient RNA. For levels >=10% ambient RNA, scSplit identified one of the singleton clusters as the doublet cluster, which means that the ARI was not clearly interpretable. Right y-axis vs points shows the estimated ambient RNA percent by souporcell versus the simulated ambient RNA percent. l, ARI of each tool on a synthetic mixture with 8% ambient RNA and 6% doublet rate with 1,000 cells per cluster for the first four clusters and a variable number of cells in the minority cluster (25-800 cells in the minority cluster).