DESeq GLM vs. Pairwise

[cburke11]

Hello all,

I have an experiment in which I have 3 time-points, (T=0, T=1, and T=2) and two strains of a moth species. The two species differ in a physiological response to a stimulus that is applied right before T=1, T=0 is before the stimulus, and T=2 is a few days after. One strain reacts right away and the other does not. My question is, what genes are different between the two strains at each time-point?

Here is the design:

condition strain
D E
D E
D E
D E
D E
D1 E
D1 E
D1 E
D1 E
D1 E
D7 E
D7 E
D7 E
D7 E
D7 E
D Z
D Z
D Z
D Z
D Z
D1 Z
D1 Z
D1 Z
D1 Z
D1 Z
D7 Z
D7 Z
D7 Z
D7 Z
D7 Z

My initial thought was to do pairwise comparisons of each of the 3 time-points between each strain but then a lab-mate brought up the problem of multiple-testing. My question is, will this be an issue? Or should I try and use a model? I will probably wind up subtracting the DE genes at T=0 from all DE gene lists because we are suggesting that these are due to population divergence. Any thoughts??

Thanks in advance for any help!

[Wolfgang Huber]

Dear cburke11

re multiple testing, if you compute, for each time point, an FDR-oriented p-value adjustment (e.g. Benjamini-Hochberg, what DESeq does by default), then you can combine gene lists from the different time points without further worring about multiple testing, since FDR is a rate (e.g. if you have list A with 10% false disocoveries, and list B with 10% false discoveries, then lists A+B and A\B will also have 10% false discoveries).

Doing the analysis time-point wise and doing arithmetic with gene lists is intuitive and easy, and may be OK in practice. More elegant (and more sensitive) would be to fit an ANOVA-like model with strain and time as factors, and with an interaction term, so you can find the genes which respond to the stimulus differently between strains E and Z. If unsure, consult one of the many textbooks on linear modelling in R.

Hope this helps
Wolfgang

[cburke11]

Thank you, your advice is very helpful! After also discussing it with some lab members I believe that modeling is the way to go.