Fig. 2.

Data preprocessing: downscaling and induced subgraph generation. Human protein–protein interaction (PPI) network (N100) with red colored edges to be deleted in the tenfold cross-validation stage to retrieve training dataset (N90) in one example fold (A). The networks constructed from truncated datasets (N90) generated with tenfold cross-validation are traversed with a modified version of the classical breadth-first search (BFS) to extract equal-sized induced subgraphs which serve as input for conditional generative adversarial network (cGAN) (see Fig. 1). Example induced subgraph node color intensity and node labels represent depth level of the modified BFS. In contrast to the classical BFS, the traversal was supplemented with size specifications and on the last depth level of modified BFS, nodes were randomly selected (possible nodes on last level: gray nodes; selected nodes on last level: pale green nodes with label) (B). Only network components covered by modified BFS are shown in these representations (giant component and appropriately sized isolated components). See Additional file 2: Fig. S5 for high-resolution image of the representative initial full network with node annotation