Here is a tweetorial for our recently published paper describing propeller, a method for testing differences in cell type proportions in single cell data. If you want to read a single cell paper without a single tSNE/UMAP, then this paper is for you!
I developed propeller as part of a computational challenge for @OzSingleCells in 2019. I have chipped away at writing the software, applying it to real data, developing simulations & benchmarking against commonly used statistical models for testing differences in proportions. 2/8
There are 5 main messages I would like to tell you about. 1) cell type props estimated from single cell data are highly heterogenous & methods that don’t account for sample-to-sample variability perform terribly (e.g. chi square test, logistic binomial & Poisson models). 3/8
2) the abundance of a cell type influences whether a method is able to detect a true difference. Propeller(asin) is conservative for rare cell types & negative binomial methods are conservative for abundant cell types. Propeller(logit) & beta binomial generally perfom well. 4/8
3) the sample size influences the performance of the methods, and as sample size increases, the methods show very similar performance. When n=3 and 5, propeller and the beta binomial distributions perform best. 5/8
4) the number of cell types detected in the experiment plays a role in performance. We simulated datasets with 2,5,7 & 20 cell types. When the no. of cell types increased to 20, the performance of the negative binomial methods improved to be comparable to the other methods. 6/8
We applied propeller to three different datasets & found significant differences in cell type composition that were not reported in the original papers. We noticed that propeller(logit) was sensitive to outliers in one dataset, while propeller(asin) was robust to outliers. 7/8
All of the analysis for the paper is available in the following worflowr website:
https://phipsonlab.github.io/propeller-paper-analysis/
https://github.com/phipsonlab/speckle
And I would like to give a big thank you to @AliciaOshlack and other co-authors @EvangelynSim @PorrelloER @AlexWHewitt @drjosephpowell for their lovely data.