DESeq/DEXSeq and GC content

[leda]

Is there an easy way to add a per-gene covariate like GC content when I fit the GLMs in DESeq? I'm only able to find per-sample covariates in the vignettes.

A broader question is: how can I determine if I need to consider GC content when I do a differential expression analysis? I've plotted fold change vs. GC content and pvalues vs. GC, and there's no noticeable trend. I figured I would run with GC content as a covariate and see if things change drastically...

-------------------

The DESeq GLM formulas I'm using now have two per-sample confounders:
form0 <- count ~ dem.group + sex
form1 <- count ~ condition + dem.group + sex

I'm using DESeq_1.10.1.

[Wolfgang Huber]

Dear Leda,
this is not straightforward in DESeq, but the facility has been added to DESeq2, see
http://www.bioconductor.org/packages...ml/DESeq2.html and in particular the vignette Appendix H on "Sample-/gene-dependent normalization factors". Let us know if you have questions or feedback!
Best wishes
Wolfgang

[Wolfgang Huber]

Re your broader question:if the GC-dependent bias on the counts is the same across samples, then you are fine with an analysis that does not explicitly account for it, since it is already implicitly accounted for by the 'base mean'. In other words, DESeq never compares across different genes, only the same gene across different samples, and the bias 'cancels out' in differential analysis.

OTOH, if the bias differs across samples, you have a "batch effect" problem, which requires a lot more effort and consideration, including hopefully the approach mentioned in my previous post. However, for diagnostics, I would first try to visualise and understand that phenomen, and why it happened, before looking at downstream analysis results. One reason is that 'normalisation' can only do so much, and sometimes it is better to drop some of the samples or re-run an experiment.

Hope this helps -
Wolfgang

[leda]

Thank you Wolfgang! Visually, the GC bias looks pretty similar across my samples... but I haven't quite decided if I need a statistic to say that with greater confidence, or what that statistic might be. Thanks for your help though.