(a) Visualization and inference of simulated scRNA-Seq data with the same parameters presented in Figure 2a but with 15 cell-types at variable frequencies corresponding to the proportions of cell-type clusters in the Hrvatin Et. Al visual cortex dataset. (b) The same as (a) but using a differential expression location parameter of 2.0 instead of 1.0, corresponding to GEPs that are more divergent from each other. The leftmost panels show the t-distributed stochastic neighbor embedding (tSNE) plot for the simulations, representing cell types with distinct marker colors, doublets as gray Xs, and cells expressing the activity gene expression program (GEP) with a black edge. In the heatmaps to the right, we display the Pearson correlation of the ground truth GEPs with the programs inferred by cNMF, cICA, and Louvain clustering. All correlations are computed considering only the 2000 most over-dispersed genes and on vectors normalized by the computed sample standard deviation of each gene. The GEPs are labeled by the type of GEP (activity, I for identity only, and I + A for cell-types that express the activity GEP) and with the frequency of the cell-type in the data.