While scCODA shows excellent performance in our simulation studies and applications, the current modeling framework possesses several limitations. In its present form, the scCODA framework requires pre-specified cell-type definitions which, in turn, hinge on statistically sound and biologically meaningful clustering assignments. In situations where crisp clustering boundaries are elusive, for instance, due to the presence of the transient developmental processes underlying the data, joint modeling of different resolution hierarchies26 or modeling compositional processes27,28 may help account for such continuities changes. Furthermore, scCODA assumes a log-linear relationship between covariates and cell abundance, which may be mis-specified in some cases. Thus, scCODA may benefit from incorporating appropriate transformation models for the covariate data to achieve approximately log-linear relations. In its current form, scCODA does not model or infer any dependency structure among the cell compositions beyond the ones induced by the compositional effects. While more complex dependencies could, in principle, be included via additional hyperpriors, this would considerably increase the computational complexity and would require more efficient inference algorithms. Finally, scCODA does not model the response variability within a condition and thus
