To account for the inherent bias present in cell-type compositions, we drew inspiration from methods for compositional analysis of microbiome data8,9 and propose a Bayesian approach for cell-type composition differential abundance analysis to further address the low replicate issue. The single-cell compositional data analysis (scCODA) framework models cell-type counts with a hierarchical Dirichlet-Multinomial distribution that accounts for the uncertainty in cell-type proportions and the negative correlative bias via joint modeling of all measured cell-type proportions instead of individual ones (Fig. 1c, Methods—“Model description”). The model uses a Logit-normal spike-and-slab prior10 with a log-link function to estimate effects of binary (or continuous) covariates on cell-type proportions in a parsimonious fashion. Since compositional analysis always requires a reference to be able to identify compositional changes5, scCODA can automatically select an appropriate cell type as the reference (Methods—“Automatic reference selection”) or uses a pre-specified reference cell type11. This implies that credible changes detected by scCODA have to be interpreted in relation to the selected reference. On top, the framework offers access to other well-established compositional test statistics and is fully integrated into the Scanpy12 ecosystem.
