Interpretation of interaction term in DESeq2

[fanli]

Hi all,

I'm having some difficulty understanding how to interpret the results of the interaction term in a DESeq2 setup. Simply put, I have an experiment with two variables - Condition (A or B), and Treatment (Untreated or Treated). I'd like to know a few things:
1) in Condition A, what is the Treatment effect?
2) in Condition B, what is the Treatment effect?
3) what is the difference in Treatment effect between Condition A and Condition B?

For the first two, I used the

factor(paste0(...))

method described in the vignette, no problem.

For the 3rd question, an interaction term seems appropriate, e.g.:

design = ~ Condition*Treatment

Looking at the results from the interaction term:

results(dds, name=c("ConditionB.TreatmentUntreated"))

I get something like 52 genes where the the Treatment effect is significantly different in Condition B vs Condition A. My intuition is that these 52 genes would show up in either of the first two comparisons above, since we're saying there is a significant difference in Treatment effects, doesn't that imply the effect should be significant in at least one of the two Conditions? But this isn't the case, as only 2 of the 52 genes are in either DEG list.

On the flip side, let's say ~450 genes were significant in comparison #1 and ~1000 genes in comparison #2 above. Why aren't the genes in the setdiff of these two lists, significant in the interaction term contrast? Am I thinking about this all wrong?

[dpryan]

A significant interaction in no way implies that there must be a significant effect within either condition. A setdiff in no way implies a significant interaction.

For your "setdiff" example, suppose you're using a p-value cutoff of 0.05. If you have a p-value of 0.049 in condition A and 0.05 in condition B then you're saying that there should be a significant interaction as well (there's normally no significant difference between such p-values). On the flip side, suppose condition A and condition B both have p-values of 0.2, but their direction of changes are in opposite directions. Then the interaction effect is quite likely to be significant.

[fanli]

Devon, thanks for the quick answer and clear explanation. I see what you're saying, I am just having trouble accepting that there would so little overlap between the two comparisons!

To follow up, I took a closer look at the genes that turned up as significant in the interaction and interestingly they all seem to have extremely large differences in log2FC with low baseMeans (see attached). That led me to wonder what statistical test is actually used to derive a p-value for the interaction terms..?

[dpryan]

I'm pretty sure it's doing a Wald test (that's the most straight-forward way do handle interactions). In recent versions of DESeq2, there's a function called plotCounts() or something like that. Use it to plot the actual counts and get a better idea about whether these are real.