Quantifying Differences in Cell Type Compositions

Across patient samples, cell type abundances may be different due to technical reasons or biological reasons. Some technical reasons include scRNA-Seq vs snRNA-Seq (e.g. comparing fresh and frozen tissue), scRNA-Seq using droplet (e.g. 10x) vs plate-based technologies (e.g. SMART-Seq2), or single cell vs spatial technologies. Biological reasons include differences in clinical covariates (e.g. age or treatment condition) or tissue sites sampled (peripheral blood vs tumor).

Because of the compositional nature of single cell data, we must apply the appropriate statistical models to identify whether a change in abundance of a cell type is significant. For example, an increase in proportion of one cell type in a sample will necessitate that other cell types in that same sample will have a decrease in their proportions, as the proportions must sum to one. Therefore the cell types cannot be considered independently of one another. We have applied two different statistical tests for compositional changes: a Dirichlet-multinomial regression model (applied to Ulcerative Colitis samples) and the Bayesian model scCODA (applied to COVID-19 lung autopsy tissue samples).