Supplementary Figure 7: Clustering of random samples of L1 interneurons.

Investigating the smallest number of patched cells needed to identify the two L1 interneuron cell types with unbiased clustering of gene expression profiles. (a) Affinity propagation was used to cluster random subsets of L1 interneurons (ex vivo cells, n=46) at decreasing numbers of cells. After each subsampling and clustering, the cell assignments were compared to the assignment based on all cells (Fig. 2d) and the accuracy score calculated (i.e. the number of correctly assigned cells over all subsampled cells). Accuracy scores from the 250 iterations per step are shown as standard boxplots, with median and mean as a line and circle, respectively. Importantly, even with random samples of 31 cells we robustly captured the two cell types (median accuracy=90%). Sampling fewer cells, in particular below 26 cells, often failed to distinguish the two cell types in an unbiased manner. (b) Showing the percentage of iterations where the affinity propagation clustering resulted in one, or two or three clusters, with the subsampling of decreasing numbers of L1 interneurons. The decline in performance at lower cell numbers was mainly caused by an inability of the affinity propagation algorithm to find two clusters, and instead reporting only one cluster.