Experimenting with zero-inflated models for single-cell differential gene expression analysis.

The analysis is built on a simulation-benchmark provided by Van-den-Berge, 2019, Genome Biology.

Aim: fast and reliable differential gene expression analysis method for scRNA-seq data

State-of-the-art:

edgeR with sample weights factored in has repeatedly been proposed as the best method for differential gene expression analysis. However, it is too slow slow with many samples (several hours with 20k cells, does not scale to 100k+ cells.)
MAST and limma-voom are scaleable, but have been shown to no always perform best.
diffxpy implements differential gene expression analysis in python/tensorflow. It uses the same fundamental model as edgeR but can run blazing-fast (on GPU)

Ideally, diffxpy could be extended to account for zero-inflation to achieve comparable performance as edgeR.

vanilla diffxpy: works fast and as well as vanilla edgeR
scqtl: implements zero inflated negative binomial model in tensorflow. not very fast and I failed to obtain the right model parameters to conduct a wald test.
fit a ZINB model from python statsmodels. Significantly better than NB model, but FDR inflation
run diffxpy with number of detected genes as co-factor. Failed because of matrix not being full rank (weird, as the same model works in edgeR)
run diffxpy with ZINB weights from zinber/zinb-WAVE: diffxpy currently does not support sample weights. Also zinb-WAVE/zingeR are quite slow and don't scale to the 100k+ cells.

check the performance of edgeR with #detected genes as covariate.
contact diffxpy developers
- why does it complain about the matrix not being full-rank (even though it works with edgeR)
- what's their take on ZINB-models?
Why does the FDR-inflation occor with the statsmodel ZINB model?