Hello,

I am pretty new to DEseq2 and am hoping someone can help me shed light on an issue. I've been experimenting with shrinkage with apeglm on a publicly available dataset and am finding what I think is maybe unusual? With the default null hypothesis, lfcThreshold=0, I am finding the same large(ish) number of genes before and after shrinkage: 2268 genes at an alpha of 0.05. When I then try to apply a very liberal lfcThreshold of 0.322, I see a total 13 genes without shrinkage (at the same alpha of 0.05) and 103 genes following shrinkage (s-value cutoff of 0.005). My understanding was that the shrinkage was supposed to be more conservative, so why might I be seeing the opposite in my data?

Is it the quality of the dataset or something else? Is it possible that, if a non-zero LFC threshold is used, the shrinkage towards this threshold pushes more genes over the significance threshold, especially for genes with LFC estimates close to the threshold? Any advice on how to proceed or what else to look at would be welcome!

I'm including the relevant figures below:

Default null, no shrinkage:

Default null, after shrinkage:

lfcTheshold=0.322, no shrinkage:

lfcTheshold=0.322, with shrinkage:

Dispersion plot for my data:

PCA plot: Not ideal, but I have already (to the best of my knowledge) removed the more problematic samples using various quality control methods.

The relevant code used to generate the four cases of results is below:

dds <- DESeq(dds)
res05 <- results(dds, alpha=0.05, name="treatment_vs_Control")
resLFC <- lfcShrink(dds, res = res05, coef="treatment_vs_Control", type="apeglm")
res05_withT <- results(dds, alpha=0.05, lfcThreshold = 0.322, altHypothesis = "greaterAbs") 
resApeT <- lfcShrink(dds, coef="treatment_vs_Control", type="apeglm", lfcThreshold=0.322)

Don't mix up the concepts of testing against a threshold and lfc shrinkage. Testing against a threshold is a more conservative version of the default Null hypothesis, asking whether there is evidence that fold changes are beyond a certain cutoff rather than different than 0. Your choice of 0.33 might "look" lenient, but depending on your power it might be quite strict. LFC shrinkage instead is not doing any testing. It tries to correct the LFC estimates (which can be noisy) given the observed variance of the data to make the LFCs more representative. That is, when noise is high and evidence is low it's shrunken towards zero, and only when there is good evidence that large LFCs are true then these will remain large after shrinkage. Essentially, they're penalized when evidence is low / variance is high.