inflamed), epithelial cells and pericytes (healthy vs non-inflamed), while all cell types in non-inflamed vs inflamed were reported as significantly changing. In contrast to Dirichlet Regression, scCODA reported credible changes in inflammatory fibroblasts (IAFs) for the healthy vs. inflamed case. Similar to M cells in the epithelium, IAFs form a responding subgroup within the inflamed donor group: While the cell type was almost absent in the control group, 13 out of 24 UC patient samples had more than five cells, indicating that ANCOM and scCODA are more likely to detect lowly abundant or absent cell types, where changes manifest in a subset of samples. On the other hand, only 10 out of 24 samples in the non-inflamed group had more than five IAFs, which was not enough for both methods to detect a credible change. We tested scCODA’s performance to detect compositional changes when only a subset of samples exhibits a response (Supplementary Fig. 9). Consistent with the observations on IAFs, lowly abundant cell types show credible changes only when at least half of the samples change upon stimulation.
