Entering edit mode
heikki.sarin
▴
10
@heikkisarin-13379
Last seen 3.9 years ago
Hi,
We have an experiment model: ~ condition + time + condition:subject.nested + condition:time . We're trying find any differentially expressed genes/exons between condition groups across any of the timepoints when accounting for the within subject variability. When we setup the model.matric we get 36 coefficients in total.
We can test it with glmLRT(fit, coef=35:36). Doing so glmLRTtest tests between two models: 1) full: ~ condition + time + condition:subject.nested + condition:time, and 2) reduced: ~ condition + time + condition:subject.nested.
However we would like to do this same also with glmQLFtest. How can we do this? How does one have to setup the coefficients or contrast?
Really appreciate all the help!
There is no problems or errors when running the glmQLFit and glmQLFTest instead of LRT approach. The point of interest on the matter is that is the statistical background of both approaches similar in such way that if I also set the coef=35:36 in glmQLFtest the test will produce an answer to the same question as with glmLRT?
Yes, it will.
Thank you very much!
Little follow-up on the matter. Is it an issue if on plotQLDisp() the Average Log2 CPM value axis and the values plotted start from -5 ? Otherwise the plot seems symmetrical and somewhat similar to the ones in the user guide.
Average abundances of -5 suggest that the lowest-abundance genes have, on average, one read per 32 million reads in each library. This seems pretty small for typical library sizes - have you forgotten to filter out low-abundance genes?
Yeah, I've done pre filtering before the analysis in two steps:
1) remove.zeros=TRUE (when creating the DGEList object)
2) A <- rowSums(dge$counts)
dge <- dge[A>10, ,keep.lib.sizes=FALSE]
Should I do more aggressive pre-filtering or is it possible that the plots are fine? Also on the Dispersion estimate plot the Average Log CPM values start from the -5 value.
This seems okay; I can only assume that you must have very large libraries (>100 million reads, perhaps?).
Well, that's good to hear. Yeah, on average around 100 million reads and highest ones a bit over 200 million.