Fig. 2

Characteristics of the considered datasets and approaches. A Class distributions of the considered scRNA-seq datasets. B Mean and standard deviations of the macro-averaged F1-scores of ACF (blue), KNN (orange), DBC (green) and conventional machine learning models (listwise deletion, red) on the three considered scRNA-seq datasets over 10 repetitions. For ACF and listwise deletion, we report the results for the baseline classifier maximizing the reported score. C Representative example of the approximated distributions of the correlations between samples from each pair of classes on the dataset by Xin et al. For the classes PP and delta (top and bottom), close similarities between inter- and intraclass distributions are observed. D Class distributions of the considered dataset from multiplexed proteomics. E Mean and standard deviations of the macro-averaged F1-scores of the considered classifiers on the considered proteomic dataset over 10 repetition. Scores are reported for raw and for batch-corrected (IRS) data. F Representative example of the improved prediction of proteomic subtypes on the dataset by Petralia et al. when considering histopathologic diagnoses. Scores refer to the baseline classifier maximizing the classification performance. G, H Importance of average class-wise correlations for the prediction of each class for the datasets by Xin et al. (G) and Petralia et al. (H). Variable importance was measured by the decrease in class-wise F1-score of an unoptimized support-vector-classifier (C = 100, kernel = rbf, balanced class weights) when removing the considered correlation