Good afternoon,

I am aware that this questions is not only relevant to DESeq2 but is more general about Regression Models, but my application right now is in DESeq2.

I have design similar to the examples: 2 genotypes, 2 conditions. We wondered if the genotypes were behaving differently to the condition, so we modelled with an interaction term.

design = ~ condition + genotype + condition:genotype

Turns out that there is no difference, and plotting some transcripts confirms it: they behave the same way after treatment. Knowing that they behave the same way is important biological information for our research that we will use.

At this point, we are therefore interested in knowing the difference between the genotypes. However, is there any benefit in keeping the interaction term in the design? Can we exclude the interaction term from the design while maintaining the condition factor and treating the samples as biological replicates using a simplified design?

Alternatively, is there a way of getting the overall effect between genotypes as currently modelled? genotype_B_vs_A or contrast=c("genotype","B","A")) is the effect between genotypes untreated, and list(c("genotype_B_vs_A","genotypeB.conditionTrt")) is the effect between genotypes treated.

Is then the overall difference between both genotypes across all treatments accessible? If so, should it give the same result as genotype_B_vs_A in the following design without interaction?

design = ~ condition + genotype

Thank you very much for your support, and advice.

If you are pretty sure that the effect is consistent across genotype, I'd go with refitting the model with your second design. I don't worry too much about the extra exploration of the data (e.g. garden of forking paths), given the context of high-dimensional analysis.