Question: Multiple testing limma contrasts
0
2.7 years ago by
gregory.l.stone10 wrote:

I am using limma-voom to determine the change in the change of gene expression between men and women due to a treatment. I have previously received great help on this site, and am confident in my workflow up until choosing a method to do multiple testing correction int he decideTests method. I don't understand the minutia defining each option, and am hoping someone can help me choose the most appropriate method.

My contrast matrix is as follows: makeContrasts(M.PvP="fMale.post-fMale.pre", F.PvP="fFemale.post-fFemale.pre", sexDiff="(fFemale.post-fFemale.pre) - (fMale.post-fMale.pre)", levels = design)

I am interested in the contrast of the contrasts: sexDiff. From researching my issue I've come to appreciate that multiple testing correction across contrasts is complicated and way over my head. When I gather my results using decideTests(), should I specify for the correction method "separate", "global", "nestedF", or "hierarchical"? (https://bioconductor.org/packages/release/bioc/manuals/limma/man/limma.pdf pg.50)

Any help would be greatly appreciated! Thank you!

modified 2.7 years ago by James W. MacDonald51k • written 2.7 years ago by gregory.l.stone10

Additionally, when I designate method as "separate" I get no sig genes from decideTests, but when I designate method as "global" I get hundreds. However, when I use toptable to collect the gene names and logFC, I get no sig genes after BH correction. Am I doing something incorrectly?

0
2.7 years ago by
United States
James W. MacDonald51k wrote:

You should look instead at the limma User's guide, in particular section 13.3.

I have read that section, but am confused due to the fact that my contrast of interest is a contrast of contrasts. I understand that toptable corrects for the contrast of interest, and using "global" in decideTests would work for me because I am considering multiple contrasts, but there is no parameter in toptable to specify how contrasts will be combined, and I'm wondering if I need to consider the fact that my contrast of interest is a contrast of contrasts. I am confused because when I look at sexDiff in decideTests I get ~300 sig genes when I use "global", but I get none when using toptable

1

Ok, thank you for the clarification

Incidentally, the reason that you get more genes with method="global" is because there are many DE genes in the male/female-only contrasts. This relaxes the BH correction for the sexDiff contrast, resulting in some DE genes being detected. However, if you're only interested in the sexDiff contrast, method="separate" is the way to go as it will control the FDR across the set of DE genes from this contrast alone. The p-values from the sex-specific contrasts are irrelevant and should not be considered during correction. Indeed, if you use global correction but only take DE genes in the sexDiff contrast, the true FDR across this list will almost certainly be above the specified threshold. Obviously, you can still use the sex-specific log-fold changes as these are useful for interpretation.

Content
Help
Access

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.