DESeq2 LRT insights

[Asaf]

Hi,
We're working on a RNA binding protein and want to figure out if it binds different RNAs in 2 different conditions. We did an immunoprecipitation (IP) for the protein and prepared the RNA that was linked to the protein for sequencing. Eventually we have 4 libraries: total RNA in condition A, RNAs in IP in condition A, total in condition B and IP in condition B.
I want to know which RNAs are more (or less) in the IP in condition A vs. B normalized by the total RNA in these conditions. I used LRT of DESeq2 like this:
ddseq <- DESeq(ddseq, test="LRT", full=~group+condition, reduced=~group)
where group is IP and total and condition is condition A and B. This should tell me if the condition influence the ratio between IP and Total.
What I'm missing is the size of the effect. I can get the p-value from the results but I couldn't figure out how to see the effect of condition a vs B on the IP/Total. I'd appreciate your help.
Thanks

[Michael Love]

hi Asaf,

The effect size is the column log2FoldChange in the results table. You can get more information on columns in the results table by inspecting the metadata columns:

mcol(res, use.names=TRUE)

[Asaf]

Thanks Mike, it's much clearer now.
However, after doing some thinking since I wrote the first post I now think that the formula should be:
full=~group+condition+group:condition, reduced=~group+condition
if I really want to test if the condition has an effect on IP/Total. Am I correct?

[Michael Love]

Yes, sorry I skimmed the original post and just answered the question about the results table. You're right that you need to fit an interaction model and test the interaction term in order to test the ratio of ratios: IP/control / IP/control.

But it appears you don't have biological replicates to fit the model with an interaction. The model with an interaction has four terms:

intercept + group effect + condition effect + group:condition effect

And you have only 4 samples if I read correctly. So there are no residual degrees of freedom to estimate the within-group variance. Is it possible to perform more replicates of some groups in the experiment to observe the variability within group and condition? At the least you could replicate one or both of the IPs?

[Asaf]

Thanks again,
We have 2 replicates of all 4 libraries so we should have enough, I do get significant genes (13) and some of them even makes sense biologically .