Figure 2.

scNym transfers cell identity annotations between young and aged rat cells. (A) Young and aged cells from a rat aging cell atlas displayed in a UMAP projection (Ma et al. 2020). Some cell types show a domain shift between young and aged cells. scNym models were trained on young cells in the atlas and used to predict labels for aged cells. (B) Ground truth cell type annotations for the aged cells of the Rat Aging Cell Atlas shown in a UMAP projection. (C) scNym predicted cell types in the target aged cells. scNym predictions match ground truth annotation in the majority (>90%) of cases. (D) Accuracy (left) and κ-scores (right) for scNym and other state of the art classification models. scNym yields significantly greater accuracy and κ-scores than baseline methods (P < 0.01, Wilcoxon rank sums). Note that multiple existing methods could not complete this large task. (E) Aged skeletal muscle cells labeled with ground truth annotations (left) and the relative accuracy of scNym and scmap-cell (right) projected with UMAP. scNym accurately predicts multiple cell types that are confused by scmap-cell (arrows).