Stability of selected markers. (A) Cross-data set F1 scores of a logistic regression classifier. SepSolve selected 20 marker genes in the MaL (left) or MeL (right) human lung data set. Classification performance was evaluated either on the same data set (dark blue) or the other (light blue). (B) Stability of 50 marker genes computed by the different methods on random subsamples of cells (top) or perturbed counts (bottom). DE crashed on subsampled IPF data because an insufficient number of cells per cell type remained. (C) F1 scores of a logistic regression and a k-NN classifier on data set IPF when using 50 marker genes selected by SepSolve for varying separation constant c.