Since extensive replication of scRNA-seq experiments is still costly and hence rare, yet essential for studying compositional changes, we also investigated the sample size dependency of effect size and rarity of affected cell type on scCODA’s performance (Supplementary Fig. 4d–f). We performed a power analysis fitting a quasibinomial model (R2 = [0.937, 0.9377, 0.936] for FDR = [0.05, 0.1, 0.2], Methods—“Power analysis”) on true positive rate values to infer the required sample size to reach a power of 0.8 with a fixed FDR for varying log-fold changes (Supplementary Fig. 4d–f). We estimated that a relative change of 1 (log2 scale) in abundant cell types (e.g., 1000 out of 5000 cells) can be determined with five samples, while the same relative change requires between 20 and 30 samples in a rare cell type (e.g., 125 out of 5000 cells) at an FDR level of 0.2. Notably, large relative changes (log-fold changes of 4) in rare cell types could be detected with less than ten samples. While this implies that for many situations only a few replicates are necessary, we would advise to increase the number of samples when detection of compositional changes in rare cell types is relevant.
