Hello all,

I have been working with the "fdrtool" lately, but am having trouble understanding how it works on an intuitive level. Specifically, I have been using it with non-uniform p-value distributions in order to estimate the variance of the model using the z-scores from wald tests (DESeq2). 

Could someone provide me with an easy explanation on how fdrtool works?

Thanks,

-R

If by 'fdrtool' you mean the CRAN package by that name, you are not in the right place. This support site is for Bioconductor packages, not CRAN. Korbidinian Strimmer has an explanation page and three additional documents that you could read. Or you could take a shot on r-help or stackoverflow.com.

For anyone else who is looking, I found this post helpful: DeSeq2 - hill–shaped histogram of pvalues

From my own analysis to gain an intuitive understanding of what fdrtool is doing when it is re-estimating the null model for a right-skewed p-value distribution (overestimation of dispersion), I compared the ranks of genes using raw p-value before and after correction. I found that when comparing this ranking in the first 1000 genes before and after using fdrtool, the ranks of genes remain exactly the same. So I understand the way fdrtool works to make the p-value distribution uniform by adjusting shoving over all the p-values while keeping the same rank.

In addition, this was backed up by looking at the plot of counts for genes that are now significant under the chosen alpha after using fdrtool. Indeed, these genes have large effect sizes and small variance within groups which corresponds to the p-value.