Fig. 3

Comparison of prediction performances between the semi-supervised model (here called CNN–GNN) with the baseline model (CNN) where graph convolution was not used. Ten trainings were done for each model to make boxplots. A Comparison of shallow CNN and shallow CNN–GNN in a classification setting, in term of area under the roc curve (AUROC). B Comparison of shallow CNN and shallow CNN-GNN in a classification setting, in term of area under the precision recall curve (AUPR). C Comparison of shallow CNN and shallow CNN–GNN in a regression setting, in term of Pearson correlation. D Comparison of shallow CNN and shallow CNN–GNN in a classification setting for ATAC peaks, H3K4me3 peaks and POL2 peaks. E Comparison of AUROC between deep CNN and deep CNN–GNN in a classification setting. F Comparison of AUPR between deep CNN and deep CNN–GNN in a classification setting. G Comparison of Pearson correlation between deep CNN and deep CNN–GNN in a regression setting