3' bias and DEXSeq

[unicornich]

Dear everyone,

Does anybody here know how DEXSeq handles 3'bias (or any other bias) of read coverage of genes?

Would it report differential usage of exons in the 3' part of the gene if libraries of one treatment have 3'biased coverage and the libraries of the other sample does not?
Also, would it report the same differential exon usage if 1 or 2 out of 3 replicate libraries of one treatment has 3'bias and the other treatment has normal replicate libraries?

Thanks!

[dpryan]

Realistically, you'd need a consistent 3' bias difference in one group to find differentially used exons due to it. Otherwise, you'd just get increased variance and non-significant findings. I imagine that you could add some sort of bias metric as a covariate to the model, but that's not something I've ever personally tried (hopefully Alejandro Reyes replies, since he would be the person to know if that'd work).

[unicornich]

I also thought that having one replicate being biased would only decrease power to detect DEU and not report false positives. But I am really not sure.
I am also hoping either Alejandro or Simon Anders will reply.

Making this biasness a covariate sounds like a good idea, thank you!
Thank you for the help!

[areyes]

Hi @unicornich,

Sorry for the late reply. How did the analysis go?

Are you expecting the treatment to cause a bias that is consistent for specific genes across replicates? Are you interested in detecting the genes that are affected by the 3' bias?

[unicornich]

Hi Alejandro!

So I did a few tests like how does differential exon usage look like only for the first 500bp or 1000bp of all transcripts (starting with 3' end) where the coverage is even and high, and compared it to my "normal" full transcript results, and it seems that I just lose power to detect differentially used exons due to the bias, but there is no increase in false positives.
I also noticed that if populations I'm comparing have the same type (or extent) of bias (there was a batch effect in my case) and if I compared 6 control replicates to each other (3 reps by 3 reps comparison) there were no false positives again.
In general my conclusion is if you compare samples with identical 3'bias, everything's pretty much fine, unless your sample replicates within each sample are variable. In this case your replicate variance is increased and consistency between replicates drops so your p-vals increase and you lose power to detect differential exon usage.

I don't know what would happen if the samples I compare would have different extent of 3'bias, though...

[unicornich]

And yes, I was thinking of adding a biasness covariate, but there are so many complications to it because I don't know how to quantify exactly different extents of bias. I was thinking fitting a spline to bias curves of replicates or something like that and taking spline points as measure, but I don't know... my test were satisfying enough, for me at least...