About DESeq2 logFC and pvalue

Question

[Dejian]

DESeq2 gives logFC values based on fitted model. We can also get the normalized counts for each sample based on the raw counts and normalization factors, thus we can calculate the logFC ourselves using normalized counts. The question is how to report the final result. I prefer the logFC+pval/FDR from the fitted model while some colleagues prefer the self-calculated logFC and model-based pval/FDR (usually these logFC values are more dramatic than that from the fitted model). I feel it is a bit weird to combine different sources of results. What do you usually do? Will you combine FC values from one method and p values from another? Just curious to know which is a common practice. Thanks.

[Martombo]

Why not both? Depending on the understanding of the methods of your collaborators, you can easily produce a table with both the the shrunken and unshrunken fold changes by adding addMLE=T in the results() function. and don't forget about the standard error!

Answer 1

[Devon Ryan]

Please ignore your colleagues, they're doing it wrong.

One of the benefits of DESeq2 is that it shrinks fold-changes (toward a Gaussian prior if I remember correctly), which should increase their reliability. This is discussed and tested in the preprint, which I advise everyone that uses DESeq2 to read (for a paper on a statistics package, it's surprisingly accessible).

So just report the shrunken fold-changes and adjusted p-values reported by DESeq2 and politely hand your colleagues a copy of the preprint.

[Dejian]

Thanks for making it clear.