In summary, using a comprehensive set of synthetic and scRNA-derived compositional datasets and application scenarios, we established scCODA’s excellent performance for identifying statistically credible changes in cell-type compositions, while controlling for the false discovery rate. scCODA compared favorably to commonly used models for single-cell and microbiome compositional analysis, particularly when only a low number of experimental replicates are available. We believe this is due to the Bayesian nature of the model as it adequately accounts for the uncertainty of observed cell counts, automatically performs model selection, and does not rely on asymptotic assumptions. scCODA not only correctly reproduced previously discovered and partially FACS-verified compositional changes in recent scRNA-seq studies, but also identified additional cell-type shifts that were confirmed by independent studies, including Treg cell enrichment in UC patients and neutrophils increase in severe COVID-19 cases. Using synthetic benchmarks, we confirmed that standard univariate tests, such as Poisson regression models, Beta-Binomial regression, or t tests are inadequate for cell-type analysis, since they do not account for the compositional nature of the data. While log-ratio transforms from compositional data analysis (such as
