Fig. 2

CellWalker Correctly Labels Cells in Simulations and Validation Data. a Label-edge weight, defined as the ratio of label-to-cell edges versus cell-to-cell edges (x-axis), is tuned in order to optimize cell homogeneity, a measure of the separability of cells of different types (y-axis). When edge weight is low, the output is more similar to de novo cell clustering (purple area), and when it is high the output becomes more similar to directly assigning labels to each cell (red area). Higher values of cell homogeneity indicate improved ability to distinguish between cells of different types, while a cell homogeneity of 0 is equivalent to no difference between within-cell-type and between-cell-type influence (dashed line). Black dots indicate mean performance across ten simulations (gray lines) b As the percent of cells with labeling edges increases (x-axis), optimal cell homogeneity does as well, up to 30% labeled (blue dashed line). c The distribution of peak cell homogeneity scores across simulations when cell distinctness in scATAC-seq is low, medium, and high. As a higher percentage of labels is incorrect (x-axis), performance begins to decline, particularly when initial cell distinctness is low. d CellWalker correctly labels cells from the ATAC portion of SNARE-seq data (number of cells of each type in parenthesis) with no drop off for rare (max 500 cells) and very rare (max 100 cells) cell types