Fig. 2

Prediction performance and validation of network-based simulations. a ROC curves from our DNN model for the CRISPR model (upper) and RNAi model (lower) according to different cutoffs of the cell line screen data. b ROC curves derived from different prediction models when the most stringent cutoff (zGARP = − 4, CERES = − 1.5, and BAGLE = 4) was used for training. c ROC curves from the DNN model with different test networks. We chose genes that have > 100 downstream genes in the regulatory network and used their perturbed expression patterns as training input. As negative controls, the breast cancer network was randomized by shuffling the nodes while maintaining the network structure (shuffled network) or inverted by reversing the orientation of all links (inverted network). In addition, the same types of networks were constructed by using liver cancer data (liver network). Shown are the averages of the five best models