search

williamritchie / IRFinder

Detecting intron retention from RNA-Seq experiments
53 stars 25 forks source link

using LRT in DESeq 2 #129

Closed jscaber closed 4 years ago

jscaber commented 4 years ago

Hi,

Thanks for this helpful tool. I don't much IR in my dataset(s), but what there is this seems to find pretty well, with fewer false positives that I am used from other splicing analyses.

I don't see any contracintications for adding a confounder in the DESeq part of the IRFinder analysis, like so: ~ grp + grp:confounder + grp:IRFinder I'm checking in case I am misunderstanding the IRFinder term in the model. Obviously, if unevenly distributed, the model matrix may need to be adjusted. Is there a possibility to specify a reduced model in the IRFinder DESeq analysis?

Bw, Jakub

dg520 commented 4 years ago

@jscaber , This is a good question. We did not include confounding factors in the regression equation because we assume they would have the same effect size on both exonic and intronic reads. As IR is measured as a ratio (i.e. intronic/(exonic + intronic)), this would cancel out the effect from confounders.

With that said, you can still estimate the confounding effect, for example:

design(dds) = ~grp + CONFOUNDER + grp:IRFinder

(I don't think you need grp:CONFOUNDER, unless there is a reason to believe the two terms have interactions) Then estimate each effect and run the test:

dds = DESeq(dds, test="LRT", reduced = ~ 1)
#Alternatively, you can do
# dds = DESeq(dds, test="LRT", reduced = ~Condition + CONFOUNDER)
# or
# dds = DESeq(dds, test="LRT", reduced = ~CONFOUNDER)
# based on which terms you care about in the downstream analysis
jscaber commented 4 years ago

Thanks @dg520 That's very helpful. I have inherited a dataset with subject variation in the dataset including a sex difference that is unbalanced (but not perfectly confounded). I have found GLMs to be poor at correcting such complex confounders, if not appropriately addressed at the experimental design stage, but I'll give it a try, if anything useful emerges I'll let you know here.

You're right about not needing grp:confounder. (I had actually wondered about the possible interaction between sex and intron retention, which is not unlikely, but does not need to be separately stated in the GLM.)