How would you test whether cell-type proportions differ between conditions in single-cell data?

Why

A differential abundance test asks whether a cell type is more common in one condition, across biological replicates. The Bioconductor multi-sample book counts cells per label per sample and uses edgeR, where the counts are cells instead of reads, so variability between replicates is modeled. The catch is composition.

Normalizing by total cells tests proportions, and the book notes that a large increase in one population creates decreases in proportion for every other population. That is technically correct, but misleading if read as those cells declining on their own. If most labels are assumed stable, normalization factors can correct for it, but that assumption is often too strong when a few cell types make up the sample. scCODA addresses this with a Bayesian compositional model and a reference cell type. propeller transforms proportions with arcsin square root or logit, because proportions near zero or one have small variance.

Milo avoids discrete clusters altogether and tests overlapping neighborhoods on a k-nearest neighbor graph with a spatial FDR. Cell capture itself can be biased, as dissociation favors some cell types, so even a correct test describes the captured cells.