RNA-SEQ where only a subset of genes is of interest

Question

[Aspire]

I am performing an analysis on RNA-Seq data, where only genes relating to specific pathways are of interest to the researcher.

1) One option is normalizing, estimating the dispersion and performing DE with DESeq2 as usual (since DESeq's assumption that most genes are not differentially expressed pertains to the whole set of genes, not to the subset).

Following that, it would be possible to manually select only the relevant subset of genes, and apply FDR only to this specific subset, ( based on the p-values calculated when taking all genes into account).

This is somewhat analogous imho to what independent-filtering does. Independent-filtering after calculating p-values for all genes, subsets the list for only those with mean higher than a certain cuttoff, maximizing that cutoff. The explicitly stated goal is to increase the number of significantly DE genes, the rationale being that genes with genes with low expression are not interesting in the first place.

Here, what defines which genes are interesting is not the mean level, but the inclusion in a specific set.

Would the described process be suitable, and is there another one if not?

Answer 1

[andres.firrincieli]

manually select only the relevant subset of genes

Just my humble opinion, but that sounds like p-hacking. What is the selection criteria applied to get such relevant subset of genes?

[Aspire]

Biological criteria, not statistical criteria. This is the pathway(s) of most interest to the PI in the first place.

[andres.firrincieli]

Now I understand. I am sorry but I never heard something like that. By doing so you are basically getting rid of p-value ranking which is used for the FDR calculation. That does not sound right to me

Answer 2

[ATpoint]

See the answer in a somewhat similar context from the DESeq2 developer: https://support.bioconductor.org/p/133932/#133938

I advise users to not try too many choices when desiring a certain outcome. See: “garden of forking paths” and issues with replication.

Answer 3

[swbarnes2]

Are you trying this because doing the DESeq pipeline the normal way didn't give you what you wanted?

Unless you have an advanced degree in statistics, you should not presume to know more than the people who wrote your software. Don't get creative with statistical methods. You need to stay on the path in order to be sure that your results mean what you think they mean.