DESeq2 LRT test with multifactorial design

Question

[lessismore]

Dear all,

I'm dealing with a multifactorial RNA-seq experiment. In this experiment, I have 4 factors and I'd like to perform a Likelihood ratio test (LRT) test.

In DESeq2 guidelines, it is indeed explained how to perform this test using 2 factors (batch + condition): it says you need "a full model with a certain number of terms and a reduced model, in which some of the terms of the full model are removed." Here I want to "determine if the increased likelihood of the data using the extra terms in the full model is more than expected if those extra terms are truly zero".

Of course I'm also interested in determining the importance of the interaction of the factors in my model which for the example in the DESeq2 guideline, for two factors would be like this:

full design: ~batch + condition + batch:condition
reduced: ~batch + condition

for testing the single factor will be:
full design: ~batch + condition
reduced: ~batch

My question is: how would this fuction be applied when you have 4 factors (e.g. batch + condition + genotype + time) ? I want to examine (i) all interactions and (i) each factor taken separately.

Thanks in advance for you help

Answer 1

[LChart]

A test requires a hypothesis; but you appear to want to do an exploratory analysis that will give rise to a hypothesis ("examine all interactions"). Ultimately if you want to apply the LRT framework as implemented in DESeq2, you need a full model and a reduced model; and if you're interested in specific terms, you have to drop them individually:

full:  ~ batch + condition + genotype + time + batch:condition + batch:genotype + batch:time + condition:genotype + condition:time + genotype:time
reduced: ~ batch + condition + genotype + time + batch:condition + batch:genotype + batch:time + condition:genotype + condition:time

(repeat ad nauseum).

That said "I want to examine all interactions and each factor taken separately" sounds a lot like you're interested in specific terms. In this case you really should fit the full model and use the regular old coefficient tests to reason about the data.

[lessismore]

Dear, thanks for your answer. There's indeed a double interest, one is purely exploratory and and one more specific and that's why I'd like to integrate LRT and Wald test based results. Do you agree, for the sake of a full exploratory analysis, I should start removing the 4-factor interaction term from the full model i.e. ~batc:genotype:condition:time, and then hierarchically, each of the 3-terms interactions (4 combinations), and then 2-terms interactions (6 combinations)., and 1 term ? Moreover, do you agree that when I remove a specific 2-term interaction e.g. condition:genotype, i should remove all interactions including these terms e.g. batch:condition:genotype. Thanks again for you time