Fig. 1

Integration of multiple datasets enables robust extraction of informative gene sets. A, B ikarus workflow. ikarus is a two-step procedure for classifying cells. In the first step, integration of multiple expert labeled datasets enables the extraction of robust gene markers. The gene markers are then used in a composite classifier consisting of logistic regression and network propagation. C Comparison of cross validation accuracy for signature derivation and model selection. Minimal balanced accuracy on the validation set was chosen as the metric of choice (i.e., worse performance on the test set). Models trained on just one dataset achieved lower balanced accuracy than models trained on two datasets (p value given by the two sided Wilcoxon test is 0.063). The combination of colorectal cancer from Lee et al. [29] and lung cancer from Laughney et al. [27] achieved the highest minimal balanced accuracy of 0.97. D Comparison of gene signature scores in laser microdissected gastric cancer data [34]. The normal gene list shows lower signature scores in cancer samples (p value 0.052, N = 8, Mood’s median test), when compared to the cancer-associated normal tissue. The tumor gene signature is significantly higher for cancer samples than the normal tissue (p value 0.003, N = 8, Mood’s median test). E Primary cells and cancer cell lines have significantly different gene signature distributions. The normal-cell gene signature shows a gradual reduction in gene signature score distribution when compared in primary cells, cell lines, and tumor cell lines. The gene signature shows the complete opposite effect. Cancer cell lines have the higher gene signature score distribution, followed by cell lines, and primary cells. Distributions were compared using pairwise Wilcoxon tests with BH-FDR correction. All adjusted p values were lower than 0.01. F Patient-derived xenografts (PDX) show significantly higher tumor gene signature score, than the normal gene signature score. The same pattern is observed in multiple cancer types. Normal and tumor signature distributions were compared using Wilcoxon tests, for each cancer type, followed by BH-FDR correction. All adjusted p values were lower than 0.01