What is the best threshold for FDR value

[ademcan]

Hi Owen,
I was surprised with some of my recent TaDa-seq data analysis. I used the polii.gene.call script to identify "expressed" genes in a Dam-Pol2 experiment but didn't find some genes that were supposed to be expressed. Then, I was wondering if one could "play" a bit with the FDR threshold. By default a gene is considered as expressed if log2>0 and FDR<0.01. However, FDR is different than p-value and I thought that one could allow a bit more flexibility with FDR compared to p-value. They explain it briefly here http://www.cbil.upenn.edu/PaGE/fdr.html.
What do you think about it? How did you decide on the 0.01 threshold?
Thank you for your help.
A.

[owenjm]

You can certainly use a less stringent FDR if you like, although obviously you'll get more false positive calls. However, I agree that what you're seeing doesn't sound "right" -- for most cell types you should see at least several thousand genes come up with the default parameters. Noisy data will impact on the FDR, though, and so will an incorrect normalisation: is the signal very weak or noisy? Do you know that the driver line is OK?

[ademcan]

Thank you for your reply Owen.
I do indeed identify a few thousands of genes as expressed but some key genes are missing.
In terms of data, what do you mean by noisy or weak exactly ( in this particular scenario) ?
I didn't perform the lab experiments, but I can check/ask for the driver line.

[ademcan]

I performed some detailed analyses about the gene I am interested in (ninaE) in control flies.
The log2 ratio is 0.102149532177245, the number of GATC sites is 16 and the FDR is 0.398215110436791.
Based on the FDR, this gene is not considered as expressed. As you can see in the "log2 scores" boxplot (the red bar corresponds to the ninaE log2 ratio) I guess we can consider it as not weak.

In addition I ploted all the max scores under the peak for the identified peaks and again the red bar represent ninaE.

I am having a look at the reads distribution on IGV to see if there is too much noise, I don't see anything weird or wrong so far.