Fig. 2: Proposed metrics to unravel complex gene regulatory dynamics.

a SimiC’s inferred incidence matrices W⁽¹⁾ and W⁽²⁾ for phenotype 1 and 2, respectively, for the input data of Fig. 1a. The incidence matrices correspond to the inferred gene regulatory networks (GRNs). b For a given regulon, and hence transcription factor (TF), we compute the regulon activity score per cell as follows. First, for a given cell c, we sort the target genes by their expression values in an increasing order. Then, for each sorted target gene we show the contribution of its corresponding weight in the weighted regulon by increasing the y-axis by that amount. The regulon activity score corresponds to the area under the generated curve. c tSNE plot visualizing the input scRNA-Seq data, where each cell is colored by its regulon activity score for TF_j. d For a given TF, using the cells' individual regulon activity score, we can compute the empirical regulon activity score distribution for each cell phenotype. An example of the obtained distributions for the considered two phenotypes and a given TF are depicted, P(C₁) for phenotype 1 (yellow) and Q(C₂) for phenotype 2 (light blue). e Computing the distance between the obtained distributions for each TF provides a metric for the regulatory dissimilarity of each TF between the cells belonging to phenotype 1 (C₁) and those belonging to phenotype 2 (C₂). The heatmap shows an example of the obtained regulatory dissimilarity for several TFs when the total variation distance δ(P, Q) is used. f The dissimilarity score can be further computed for different cell-clusters across the phenotypes, as shown in the heatmap, which reveals TFs with high dissimilarity scores across cell clusters as well as TFs with similar ones.