Fig. 2

Performance of DeepLinc in reconstructing cell interaction networks. A The levels of AUROC at different training epochs of DeepLinc. The original data or data with different levels of random noise in the gene expression profiles were used. Larger σ values would result in higher levels of random noise in general (see “Methods” for details). B The AUROCs between the randomly removed real edges and originally nonexistent edges. Different proportions of the edges in the original cell adjacency maps were randomly picked and discarded. Then, the remaining edges were used for training the DeepLinc model. This process was repeated 30 times for each proportion of the missing edges to draw a boxplot. C The AUROCs between mixed-in fake edges and pre-existing real edges. Different numbers (1- to 10-fold of the size of the original cell adjacency map) of randomly generated fake edges were added into the original cell adjacency map, which was then fed to DeepLinc. This process was repeated 30 times for each level of fake edges to draw a boxplot