Considering only a handful of genes of a priori interest (RNA seq), is it legitimate to not perform FDR adjustment of p-values?

Question: Considering only a handful of genes of a priori interest (RNA seq), is it legitimate to not perform FDR adjustment of p-values?

11 months ago by

marcos.georgiades.18 • 10

marcos.georgiades.18 • 10 wrote:

Hi all,

I have generated a database of normalized counts for transcripts produced from an RNA sequencing experiment. I have an a priori interest in only 5 of them and was wondering whether it is legitimate to simply perform tests of significance for those 5 without considering the rest, and thus get away without FDR-adjusting the p-values. Is that ever acceptable?

Thanks!

rna-seq p values false discovery rate fdr • 296 views

ADD COMMENT • link •

written 11 months ago by marcos.georgiades.18 • 10

11 months ago by

ashish • 570

ashish • 570 wrote:

If you were interested in only 5 transcripts why did you go for RNA-seq. A simple real-time PCR would have been a lot cheaper. Moreover, DE tools borrow information across genes for modelling the counts and estimation of parameters so I don't think its a good idea to just focus on 5 of them for downstream analysis. If you have some other independent ways to verify those 5 transcripts then it may be acceptable.

ADD COMMENT • link modified 11 months ago • written 11 months ago by ashish • 570

Thanks for the reply! The initial goal was to take the non-hypothesis driven path but then my boss, based on some other experiments, decided to focus her research on a few genes only.

ADD REPLY • link written 11 months ago by marcos.georgiades.18 • 10

If you have enough samples, I'd suggest do correlation analysis and then create a network. Then identify which clusters contain those 5 genes and then examine those clusters. This way you can show something to your boss and backed by experimental data you can generate some hypothesis from it as well.

Also, as German suggested you don't really need to do FDR adjustment. It just takes care of avoiding false positives when doing multiple comparisons. If you take a really low pvalue threshold (not 0.05), FDR can be avoided.

ADD REPLY • link written 11 months ago by ashish • 570

11 months ago by

German.M.Demidov • 1.9k

Tübingen

German.M.Demidov • 1.9k wrote:

It is not necessary to do FDR adjustment at all - just be sure that you know the FDR of your data without adjustment (it is not 0.05 when you use 0.05 threshold, for further details: https://www.frontiersin.org/articles/10.3389/fphy.2016.00006/full ). FDR correction makes things "easier" due to magic - you don't need to bother with prevalence and power anymore, but if you are ready to do this analysis - you may go without FDR correction.

If you want to incorporate prior knowledge, I believe you have to proceed with Bayesian analysis.

ADD COMMENT • link written 11 months ago by German.M.Demidov • 1.9k

Thanks for the advice and the link!:)

ADD REPLY • link written 11 months ago by marcos.georgiades.18 • 10

Please log in to add an answer.

Similar posts • Search »