Is excluding a whole chromosome a valid independent filter for DESeq2?

[Skiaphrene]

Hi everyone,



I’m currently working on some RNA-seq data where there is partial confounding between my main experimental factor of interest and the sex of the mice used, and I’m looking for a way to deal with the spurious DEGs that are consequently generated (in a more robust way that simply ignoring them).



Typically, in one experimental condition I have 2 samples from mice of one sex, and 1 sample of mice of the other sex, and in the control condition I have 1 sample of mice from the former sex and two for the latter.

I’m using DESeq2 to analyse this data and I have included Sex as a factor in the GLM model in order to account for it:

Design = ~ Sex + Condition + ...

Even though this drastically reduces the number of sex-related genes amongst the Differentially Expressed Genes (DEGs) between Conditions (instead allowing them to be recovered if I test for Sex differential expression), I’m still getting some Condition DEGs that are located on the Y chromosome. The biologists I'm working with don't like this at all.



DESeq2 offers the possibility of “independently filtering out” genes prior to DE testing,

“provided that the criterion for omission is independent of the test statistic under the null hypothesis” (DESeq2 paper)

Under the assumption that very few Y chromosome genes should be DE between Conditions, and that the few that might be (i.e. Sex and Condition do interact) do not interest us biologically, would it be a valid independent filtering strategy to simply exclude all Y chromosome genes prior to testing? That way the Sex factor in the model would account for other sex-related DEGs on non-Y-chromosomes, and Y chromosome genes would not be tested, thus increasing power for all other genes.



Please let me know what you think!

Thanks in advance,

-- Alex

[Michael Love]

Yes you can filter out Y chromosome genes prior to running DESeq().

The null hypothesis for a gene is that there is no differential expression across the condition. So the chromosome is definitely independent of the test statistic under the null.

An example of a non-independent filter is the within-group variance. Under the null, you can see a correlation between within-group variance and the test statistic. So this is an invalid statistic for filtering.

[Skiaphrene]

Thank you again Michael for a quick and clear answer!

Best regards,

-- Alex