Fig. 2

Performance of feature selection methods and impact of cell type numbers on gene selection for classifying cell type using Tabula Muris atlas. a F1 scores calculated from a k-nearest neighbor classifier (KNN; k = 7) using the union of top-10 cell type marker genes selected by each feature selection method for classifying cell types. Higher F1 scores indicate more accurate cell type classification. The number of cell types varies from 5, 10, 15 to 20 in the datasets and cell type classification was repeated 10 times, by random sampling from Tabula Muris altas, for each feature selection method in each setting to capture the classification variability. Statistical significance (* p < 0.05 based on Wilcox rank sum tests) was denoted if every deep learning-based method outperformed every traditional method. b Similar to a but quantifying KNN cell type classification performance by sensitivity and specificity. c For numbers of cell types set at 5, 10, 15, and 20, balloon plots summarizing the ranks of median F1 scores from KNN on datasets sampled from Tabula Muris and Tabula Sapiens atlases. The size of the balloon represents the rank of the method, the larger the better its performance. d mean F1 scores from a plotted against the increasing number of cell types, and e coefficients of slopes from least squares fitted lines to F1 scores in d