Fig. 3: Applications of SciBet.

a Mean accuracy (across 50 repeats) for n = 6 cross-platform dataset pairs listed in Supplementary Table 2. b Cross-species classification with three human pancreas datasets projected to Tabula Muris dataset (Sankey diagram). The height of each linkage line reflects the number of cells. c Confusion matrix of the cross-validation result for 30 cell types in the “mock” human cell atlas (listed in Supplementary Table 3). d Single cell classification for a human liver dataset with integrated human dataset as reference, implemented by SciBet. e Confusion matrix for the case study of false positive control, with normalization for each row (origin label). Negative cells including malignant cells, CAF cells and endothelial cells were removed from the training set. Query cells with lowest classification confidence scores were labeled as unassigned. f False positive control evaluation with cell types not present in reference as negative cells, with n = 10 pairs of datasets (each point represents the mean accuracy score or FPR across 50 repeats). Box plot shows the center line for the median, hinges for the interquartile range and whiskers for 1.5 times the interquartile range. g Expression heatmap of the top 54 genes selected by E-test for the integrated immune dataset (Supplementary Table 6). h 2D-UMAP showing the dimensional reduction result based on the genes in g.