Fig. 5. Inference of cell type-specific networks for human hematopoietic differentiation data.

a UMAP of LIGER cell clusters of the scATAC-seq and scRNA-seq data. b UMAP depicting the original cell types (samples) with scATAC-seq and scRNA-seq data. c The distribution of cell clusters in each sample. d The distribution of samples in each LIGER cluster. e Inferred lineage structure linking the eight cell clusters with scRNA-seq data. f Boxplots showing F-scores of n = 7 cell clusters (all cell clusters excluding C1) for top 1k edges in predicted networks from scMTNI, scMTNI+Prior, INDEP, INDEP+Prior, SCENIC and CellOracle compared to gold standard datasets (top). FDR-corrected t-test to compare the F-score of the row algorithm to the F-score of the column algorithm (bottom). The two-sided paired t-test is conducted on F-scores of n = 7 cell clusters for every pair of algorithms. A FDR < 0.05 was considered significantly better. The sign < or > specifies whether the row algorithm’s F-scores were worse or better than the column algorithm’s F-scores. The color scale is specified by − log(FDR), with the red color proportional to significance. Non-significance is colored in gray. In the boxplot, the horizontal middle line of each plot is the median. The bounds of the box are 0.25 quantile (Q₁) and 0.75 quantile (Q₃). The upper whisker is the minimum of the maximum value and Q₃ + 1.5*IQR, where IQR = Q₃ − Q₁. The lower whisker is the maximum of the minimum value and Q₁ − 1.5*IQR. g Pairwise similarity of networks from each cell cluster using F-score on the top 5k edges. Rows and columns ordered by hierarchical clustering using F-score as the similarity measure. Source data are provided as a Source Data file.