Fig. 3

Technical benchmarks. a Left axis, amounts of detected correlations (sensitivity) decrease with the number of clusters. Right axis, mean and median of co-expression (specificity) increase with the number of clusters. b Distribution plots of the co-expression (Jaccard index) for 3 (top) and 48 (bottom) clusters. c Distribution plots showing all the correlations of the testing dataset (gray) compared with a selected subset (those predicted by the training dataset). d Pie chart, composition of the predicted correlations in the testing dataset (5000 cells). e Relationship between size of the set and resulting predicted correlations with ρ > 0.6 (mean ± S.E.M.). f Scatter plot of correlations inferred from the original counts against those inferred from imputed counts (scImpute). The two distributions are correlated with Pearson ρ = 0.983. g AUC of inferred correlations, simulated data, 10 independent repetitions for each sparsity (error bars S.E.M.). h ROC curves for 10 simulations with the highest sparsity of 97% are above the random ROC line. i Example of the validation of Eomes neighbors in the brain network. For each edge, we computed fold enrichment (proportional to edge width, highest Otx2 with 9.87, lowest Adcy7 with 0.83) and a p value (labels). In the case of Eomes, all edges but one (Adcy7) are validated with p < 0.05. j Overall distribution of edge-wise fold enrichment in the brain network was biased towards positive values, suggesting that neighboring genes are regulated together in perturbed systems. k Brain network, percentage of validated edges with p < 0.05 (y-axis) filtered by the number of MSigDB occurrences (x-axis). Higher percentages of validated (p < 0.05) edges are obtained by considering only edges with high occurrences