Fig. 6: Differentiate TNBC subtypes.

A Information of the datasets. The datasets comprise 3 patient groups, namely cold, mixed and compartmentalized, consisting of 6, 19 and 15 patients, respectively. B The cell type annotations and MENDER-predicted domains on representative patients from different groups are shown. PCA plots of the patients characterized by the proportions of cell type (fine), termed CT-fine repr (C), cell type (coarse), termed CT-coarse repr (D), and MENDER-identified domains termed MENDER repr (E), respectively. N = 6 independent samples for cold, N = 19 for mixed, and N = 15 for compartmentalized. Boxplot setting: the lower and upper hinges show the first and third quartiles (the 25th and 75th percentiles); the center lines correspond to the median; the upper whisker extends from the upper hinge to the largest value, which should be less than 1.5× the interquartile range and the lower whisker extends from the lower hinge to the smallest value, which is at most the 1.5× interquartile range. F The separability of patients by 3 different proportions is quantified by the classification accuracy of 2 supervised classifiers, i.e., KNN and SVM. The y-axis shows the classification accuracy (ACC) reported fivefold cross-validation, using either KNN or SVM as classifiers. The patient-level labels are cold, mixed and compartmentalized. The p-values (one-sided t-test) indicate the significance of the difference between CT-fine repr and MENDER repr, and between CT-coarse repr and MENDER repr, respectively. Error bars are based on mean and 95% confidence interval. G Similar to F, except that the features of these proportions are PCA-reduced to the same dimensions before classification. The number of PCs ranges from 2 to 17. The y-axis also shows the classification accuracy (ACC) reported by fivefold cross-validation. N = 5 independent experiments. Error bars are based on mean and 95% confidence interval. Source data are provided as a Source Data file.