Fig. 3

Integration of human pancreas datasets. a Visualizations of iMAP batch effect removal results on the “panc_rm” dataset. Three kinds of colors are used to illustrate the cell type, batch, and evaluation information. b Quantitative assessments of different batch effect removal methods on the “panc_rm” dataset. See also Additional file 1: Fig. S2, 5, 6. c The training process of the “panc_rm” dataset. The changes of loss values versus epochs for all four loss functions of two stages are shown. d Running stability of iMAP. We re-ran the iMAP on the “panc_rm” dataset for 20 times, and computed the proportions of positive and true positive cells. The marked black points are the best results obtained by other methods. e Stage I increasing the number of MNN pairs. f The rwMNN procedure boosting the proportion of true positive cells. The “Smart-seq2” and “inDrop” sub-datasets of the “panc” dataset were used to compare the number of MNN pairs. See also Additional file 1: Fig. S11. g Importance scores of selected genes for building the representations of each cell in stage I. h Importance scores of selected genes for removing batch effects in stage II