Fig. 2.
Benchmarking results for dataset AML-sim. a Performance metrics for dataset AML-sim, testing for differential abundance (DA) of cell populations. Panels show (i) receiver operating characteristic (ROC) curves and (ii) true positive rate (TPR) vs. false discovery rate (FDR) (also indicating observed TPR and FDR at FDR cutoffs 1%, 5%, and 10%). Representative results for one condition (CN vs. healthy) and abundance threshold (1%) are shown (complete results for this dataset are included in Supplementary Fig. 1). b Results for additional null simulations, where no true spike-in signal was included; p-value distributions are approximately uniform (additional replicates are included in Supplementary Fig. 2). c Heatmap displaying phenotypes (expression profiles) of detected and true differential clusters, along with the signal of interest (relative cluster abundances, by sample), for method diffcyt-DA-edgeR. Expression values represent median arcsinh-transformed expression per cluster across all samples (left panel). Rows (clusters) are grouped by hierarchical clustering with Euclidean distance and average linkage; the heatmap shows the top 20 most highly significant clusters. Vertical annotation indicates detected significant clusters at 10% FDR (red) and clusters containing >50% true spiked-in cells (black). (Additional heatmaps are included in Supplementary Fig. 3) d Results for varying clustering resolution (between 9 and 1600 clusters), showing partial area under ROC curves (pAUC) for false positive rates (FPR) <0.2 (additional figures are included in Supplementary Fig. 4). Performance metric plots generated using iCOBRA42; heatmaps generated using ComplexHeatmap43