Assessing differences in cell type/state abundance: compositionality, heteroscedasticity and bias
Author(s): Koen Van den Berge,Alemu Takele Assefa,Bie Verbist
Affiliation(s): Janssen R&D
Social media: https://twitter.com/koenvdberge_be
Assessing differences in cellular composition between conditions and disease states is of principal interest in immunology and medicine, helping in unraveling disease and informing drug development. The data for such analyses typically consist of a count matrix, where each element of the matrix denotes the number of cells observed for a particular cell identity (be it cell type or state) in a sample. These data are compositional, as the count for each sample is constrained by an arbitrary total, making the counts negatively correlated between cell identities within a sample. While recent methods respect the compositional nature, often a naive analysis is performed where this is ignored, resulting in decreased performances of the analysis. In this work, we evaluate currently available approaches for testing differential cell type composition. We introduce a workflow that combines building blocks from state-of-the-art methodology by applying a compositional transformation to the cell count data while still respecting their mean-variance association in the modeling. We show, using simulations and a case study, the impact of correctly accounting for the compositional nature of the data.