Fig. 2

miloDE yields sensitive and precise detection in simulated data. A Schematic representing how detection power is estimated for each gene and cell type purity threshold. The top left panel illustrates single-cell data, embedded in UMAP space, with highlighted cell type, in which we will alter the counts. The top middle panel represents the “in silico” perturbation we introduce to the selected cells, and the top right panel represents neighborhood assignment, followed by per neighborhood quantification of whether the selected gene is identified as DE. The bottom left panel represents per neighborhood quantification of the cell type “purity.” The bottom middle panels represent how “ground truth” DE neighborhoods are selected based on cell type purity threshold. The bottom right panel illustrates the final quantification of DE detection (i.e., sensitivity + FDR). B Boxplots representing the distribution of cell type purity score across neighborhoods for each neighborhood assignment replicate and for k = 20, 25, and 30. Dashed lines correspond to selected cell type purity thresholds. C Boxplots representing how sensitivity and FDR change with estimated logFC, k, and cell type purity threshold. Each box represents data across 5 replicates for a single k