Fig. 3.

Network-based prediction quality of gene expression levels. The ten learned gene regulatory networks were analyzed for their performance to predict the expression levels of the 12,285 genes in independent tumor test sets (TCGA OD: 59 network-specific oligodendrogliomas left out from network inference, TCGA OA: 118 closely related oligoastrocytomas). Corresponding histograms of gene-specific median correlations between predicted and measured gene expression levels are shown. The strong shift of both histograms (red, blue) into the positive range shows that the prediction quality of oligodendroglioma-specific networks was significantly better than for random networks (grey) of same complexity (Wilcoxon rank sum tests: P<2.2·10⁻¹⁶)