Remove 1 sample and all pvalue adj bump to 1.

[ericloud]

trafficstars

Hi DEP's authors,

I have some trouble with the p-value adj. compute by fdrtool.

In my design, I have 7 samples of cond1 and 19 samples of cond2. If I use all these samples, DEP works with 1844 proteins:

• 154 proteins are significant
• the protein x have:
  • pvalue: 2.03691168653476E-06
  • pvalue adj: 3.32E-13

If I remove 1 sample in my cond2 (7 vs 18), DEP works with 1893:

• no proteins are significant, all pvalue adj bump to 1.
• the protein x have:
  • pvalue: 4.14508026818263E-06
  • pvalue adj: 1

It doesn't seem to be a bug, but I'm quite surprise that removing one sample change the p-value adj so much! So I'm wondering:

1. Why you have choose this method compare to p.adjust() and the qvalue package ?
2. If the fdrtool package is appropriate in all case, to calculate a p-value adj ?

If the answer of 2) is not: 3) Have you an idea why my case doesn't work ?

Bonus: is It interesting to offer the possibility to change the method ? (I can work on it).

Thanks for your work and your help, Eric.

[ericloud]

Here some fdrtool plots, with all samples:

Here some fdrtool plots, after removing 1 sample :

The null component doesn't fit at all the mixture.

In that case, fdrtool (with default param) can't ajust the pvalue. Which is why all pvalue adj bump to 1.

[JieWu2012]

Hi ericloud, I got similar problem as yours when running DEP. I got very few differentially expressed proteins if I use p.adj to filter since most of them are very high. For example, the p value of one protein is 4.8e-5, but the p.adj=0.2. It seems that fdrtool use t-statistic-derived FDRs. I thought maybe I can adjust the p value using BH method by myself using the p values output from DEP. But I am not sure if it is correct way to do. The DEP authors hasn't answer the issues for a long time. Does anyone have any idea? Thx a lot.