Differential expressed genes - much more DE genes for one factor in the additive model than the interaction model . What may be the reason?

Question

[Pei]

Hello!

I applied DESeq2 to model my two-factor gene expression data: each factor has 2 levels - A0 and A1, B0 and B1.

I found no significant genes for the interactions term A:B (cutoff: FDR < 0.05).

However, the number of genes with significant changes by B1 relative to B0 was ~5,000 for the additive model (~ A + B) and ~230 for the interaction model (~ A + B + A:B)

(no significant genes observed for the A0 vs A1 in the additive model)

My question is that: if the number of DEG could decrease from 5,000 to 230 by adding an interaction term, I may expect many genes affected by interaction between A and B. Right? But I still found no significant genes due to interaction. So, does this suggest that interaction between A and B may affect many genes but the model, and data, have no enough power to reject the NULL model statistically?

I am trying to look at the data to see what could explain this difference. Could anyone please share your suggestion and comments? Let me know if anything else I could provide. Thanks!

P

[i.sudbery]

I'm a little confused by what you did here.

You said that you found no signficiant genes for the interaction model. But then in the next paragraph you say you found ~230 for the interaction model? When you say you found no/230 genes for the interaction model, so you mean testing for significance of the interaction term in the interaction model? Or for the significance main effects in the interaction model?

I.e. you need to tell us the contrasts as well as just the designs.

[Pei]

Thanks. It's mine. It should be:

When using ~ A + B,
I found 5,000 significant genes for B.
When using ~ A + B + A:B,
I found 230 significant genes for B but 0 significant genes for A:B.

Answer 1

[i.sudbery]

In general interaction models have much less power than purely additive models. Indeed, every term you add to a model, reduces your power, all other things being equal . Of course, all other things are not always equal, it could be that a main effect is being entirely masked by an interaction effect, and so you find no signficiant main effects if you leave out the interaction! But in general that is not the case.

The most likely explaination is simply that your experiment is underpowerd to detect significant interaction effects, and putting the interaction effect in the model is also reducing your power to detect the main effects.

One possiblity i've seen here is that taken by the stepR package. You conserve power by only testing the interaction term for those genes where one of the main effects is significant in a model without the interaction term. This might well miss situation like (lfcA = 5, lfcB=0, lfc(A+B)=0) because the interaction might mean the A main effect is not significant. But its better than nothing.

[Pei]

Thanks. Yes, I also guess the experiment was under-powerd to detect significant interaction effects.

Answer 2

[BioinfGuru]

"I may expect many genes affected by interaction between A and B. Right?"

For a gene to be returned as significant in the interaction results, it has to be significant in both comparisons. (i.e. A0vA1 has a significant effect, and B0vB1 has a significant effect). Then the interaction returns the result of subtracting one from the other.

You found 5000 significant DEGs in the comparison B0vB1 only. If you look at A0vA1 only, I think you will see that there is very few significant results. Thus the interaction A:B is returning very little.

[i.sudbery]

I disagree with this. Its not necceasry for the main effect to be significant for the interaction to be significant. Consider the situation below:

Treatment A    Treatment B    Expression
      FALSE          FALSE             1
       TRUE          FALSE             1
      FALSE           TRUE             1
       TRUE           TRUE            10

With sufficient replicates you'd find here that each of the main effects is insignificant, but the interaction is significant.

[Pei]

Thank you BioinfGuru. Yes, no significant genes observed for the A0 vs A1 in the additive model.
But I am not sure about your explanation. I will think about it carefully.

Currently, my main reference are page 16 on this and the "initial note" at DESeq2 bioconductor vignettes