An update to edgeR

[SamGG]

Hi,

I posted edgeR questions for use with cytometry counts, but didn't get any answer yet.

There are 3 main points:

• how to set the filtering min.count when libraries are quite different (by a factor 10 or more)?
• how to set the prior count as an absolute CPM value?
• should we move from glmLRT to glmQLFTest?

If you have time to answer any of these points, I would appreciate.

In addition, I think an article comparing DA performed with edgeR vs Wilcoxon (or Student test in some cases) would be useful for the cytometry community in order to move it to the use of NegativeBinomial distribution, which is the "must" for the new chief editor of Cytometry part A, Bartek Rajwa.

Best,
Samuel

[timmocking]

Hi Samuel,

Totally agree with your last point. I see the appeal of performing differential testing at the "count" level rather than the "proportion" level. However, I do have my doubts that modeling uncertainty of counts is actually relevant for flow data, as most data that I've personally worked with almost always has more than 100K events/sample.

Whenever I used diffcyt-edgeR, more than 50% of clusters (typically >100) return significant (even after P-value adjustment). There is a massive discrepancy with Wilcoxon there, which is a lot more conservative.

What is your personal opinion? Should we indeed move to the NegativeBinomial?

[SamGG]

Hi Tim,

I think NegativeBinomial is the distribution to use for considering the real value of the counts and not only considering counts as a mean of ranking. In a 2 analyses, I tried to assess whether NB really fits the counts I had, and it did (far better than Poisson distribution). But I would like to have a way to report this for each analysis, taking into the biological variation.

Some authors perefer Wilcoxon for RNAseq (or scRNAseq?). I don't have any reference, just feedback from a reliable colleague.

Assuming that NB should be used, the inflation of significant p-values is a good news: we have more statistically significant results with the samples in hands. The bad news is that WE have to decide which clusters are the most interesting, ie with a practical/biological value, because the statistical p-values tell us to reject the null hypothesis. For this decision, we can't not rely any more on the p-value to make the decision for us.

I agree that with large counts, a small change might become statistically significant, in the same way a linear regression through >100 points often has a p-value < 1e-4, but a R2 that is far from 0.6. So, we should consider the Fold Change once p-value is OK.

Finally, I am also questionning the choice of the reference population(s) when comparing sample groups. If the cell count of one population is increasing in one group but not in the second group and none of the other cell counts change in both groups, then all the latter proportions will decrease in the first group in comparison to the second (reference) group of samples. So the single increase and all the decreases might be statistically significant although only one single population count changed. If we normalize so that the ratio of proportions is 1 between both groups, I think only the increased population will be reported. I hope I was clear enough. IIRC diffcyt is using edgeR but not the TMM library normalization, which is correct because the hypothesis of TMM are specific of RNAseq IMHO. I don't have any reference in mind about normalization between samples (aka libraries) in the field of cytometry, but I think it is important. Of course in true experiments, more than one cell population is expected to change, making the normalization not that trivial.

Hey @markrobinsonuzh and @lmweber, what are your views?

NB: I slightly edited my original post.