Co-expression analysis at single cell level

Question

[coldshot]

Hi, I am trying to find receptor coexpression at the single-cell level using scRNA-seq data generated using 10X chemistry. Given the dropout and doublet rate, what is your opinion on using scRNA-seq to perform receptor coexpression analysis? Any suggestions based on your experience?

Answer 1

[rpolicastro]

It's still somewhat of an emerging field of analysis, but I recently went down the rabbit hole msyelf. There is a clear problem with coexpression in scRNA-seq caused by sparse data and over-dispersion of counts. Because of this, measures of proportionality were found to be some of the best performing metrics of association with scRNA-seq (Skinnider, 2019). In the Skinnider paper they used the proprR package in R for both their phi and rho calculations, which I've had luck with myself. I want to note that although pearson and spearman did not perform as well as phi and rho, I've found them useful for initial exploratory analysis.

There has also been software developed specifically for proper coexpression in scRNA-seq data. This won't be a comprehensive list, but a mention of some of the more notable examples. One of the more promising methods in theory is COTAN, which uses the proportion of zero counts as a measure of coexpression. My one major gripe with the software right now is that it's not written as a proper R library and currently utilizes a dependency that is no longer on CRAN. These combined make the software difficult/impossible to use. Besides COTAN, there is now other software popping up in biocondcutor and elsewhere that are properly written and maintained, so may be more worthwhile to try out. Two examples include fcoex and scLink. I'll be testing these out myself soon and will report back here with anything notable.

Sticking with what works, trusty old WGCNA is perfectly fine to use with scRNA-seq data too. Depending on your settings and the amount of data, it may take a long time to run through.

[nanna]

Thanks for you answer, it was very helpful. Did you manage to try out fcoex and scLink?

[Quang]

rpolicastro nanna did you have any update on fcoex or scLink? There is also MAGIC followed by scran::correlatePairs() posted at Gene correlation with single cell data?