Hi,

I tried to conduct threshold-based Wald tests by using the lfcThreshold and altHypothesis arguments to the DESeq2::results function. However, I noticed in the DESeq2 (2014 Genome Biology) paper, it stated in page 6 "DESeq2 offers tests for composite null hypotheses of the form |ßir| <= theta, where ßir is the shrunken LFC from the estimation procedure described above. (See Methods for details.)". So it appears one should first get the shrunken LFC values before doing the Wald test, but after version 1.16 the shrunken LFC values can only be obtained by using a separate function lfcShrink(), which can only be run after running the results() function. So it seems there is no way to do threshold-based Wald tests on the shrunken LFC values in recent versions of DESeq2. Is my understanding correct?

I did some test runs:

res_default <- DESeq2::results(object = dds);
res_composite_null <- DESeq2::results(object = dds, lfcThreshold = 0.5, altHypothesis = "greaterAbs");
res_default_lfcShrink <- DESeq2::lfcShrink(dds = dds, coef = 2, type = 'normal', res = res_default);
res_composite_null_lfcShrink <- DESeq2::lfcShrink(dds = dds, coef = 2, type = 'normal', res = res_composite_null);

It looks like the pvalues in 'res_composite_null_lfcShrink' is exactly the same as the pvalues in 'res_composite_null', so the lfcShrink() function only changes the log2FoldChange and lfcSE columns. Is there a way to re-run the results() function on the output of lfcShrink()?

Thanks!

Lei

hi Lei,

In version 1.16 and 1.18, the default is not to test on shrunken coefficients, but to leave shrinkage as a secondary step for ranking and visualization. lfcShrink() only moderates the effect sizes and does not change the p-values in the table.

So you test on |beta_ir| <= theta where beta_ir is the non-shrunken LFC. I want to point out that the test is perfectly valid on the non-shrunken LFC.

You are right that in these two versions, lfcShrink() does not offer a threshold test. It is one of my main todos in this devel cycle to implement lfcThreshold within lfcShrink() (see this note to myself) but I haven't done it yet. (And then, in future workflows, when we do want to replace p-values of MLE coefficients with inference on the shrunken LFCs, we will be outputting posterior tail areas and "s-values" instead of p-values and adjusted p-values.)