Since non-Bayesian methods are not able to produce any results for the case of one sample per group due to a lack of degrees of freedom, we assumed no discoveries on these datasets, resulting in MCC, TPR, and FDR of 0. In contrast, Bayesian models adjust prior assumptions by the evidence from the data. Therefore, tests on one-sample data are possible, albeit with a strong influence from the choice of priors. Because scCODA gives equal prior probability to exclusion and inclusion of an effect (Methods—“Model description”), the selection of credible effects is driven by the data, even when the sample size is small. Supplementary Fig. 3 shows that Bayesian models can still detect some very strong effects (increase = 2000), even in the one-sample case.
