Figure 4.

Gene-gene correlation before and after imputation and its impact on the predicted interactions

(A) Gene-gene correlation distributions obtained in each cell type color coded by imputation method among top 500 most variable genes and significantly varying TFs.

(B) Paired density scatterplots before and after imputation with dca. GRNBoost2 reported the pairwise interactions between ASXL1, SNAPC3, and ZNF488 among the top-k network after imputation in hESC data.

(C) Change of edge ranks in TP interactions identified by unimputed model after imputation in hESC data. Dashed line indicates the rank threshold corresponding to the top-k network. Interactions below the dashed line represent TP within the respective model. Low edge ranks represent highly important interactions.

(D) Scatterplots comparing correlation values between genes before versus after imputation. Each scatterplot corresponds to one model in hESC data. TPs are red crosses, FPs are yellow dots, and FNs are blue dots. For each scatterplot, we fit one regression line for TPs, FPs, and FNs, respectively, with the corresponding color. For visualization purposes, we added a background color to the lines to better distinguish the line and the dots. Positively predicted interactions differ clearly from FN interactions.

(E) Counts of positively and negatively predicted network chain motifs in hESC data for each model. TP network chains agree both in prediction and evaluation networks (ChIP-seq-derived network). FP network chains are falsely positively predicted chains being actual feed-forward loops in the evaluation network. FN network chains are falsely predicted as being feed-forward loops when they are actually network chains in the evaluation network.

(F) TPR and FDR scores for network chain motifs obtained by statistics in (E). mDC dataset is not included as no motifs could be found among the top-k network. Ideally, TPR values should be close to 1, whereas FDR values should be close to 0.