Is it possible to weight previously validated genes in DE analysis?

[scottdaniel]

Hi,

This is somewhat of an open-ended question and is related to this post from over a year ago: RNAseq analysis by DESeq : can't find a gene previously published as important

The basic question is this: If Gene X has been shown by quantitative PCR to have more copies of mRNA in condition A when compared to condition B, is there a way to weight the gene in RNA-seq analysis if it is not initially showing as differentially expressed? (let's assume the more copies are due to expression and not because of a lack of degradation)

FWIW, I've used the tuxedo suite and STARaligner > htseq-count > edgeR as my two pipelines for DE and have not found anything in those tools to 'mark' a gene as something that should come out as significant.

What I can imagine is this: let's say my p-value cutoff is .05 and gene X is showing up as .07, i.e. not significant. Is the solution to simply set .08 as my 'true' cutoff, thus making gene X significant? Or is that just way too simplistic? Is there a more sophisticated way to do this with one of the software tools available (either the ones above or not).

Thanks for any suggestions / comments.

[dpryan]

Why not just use a more bayesian approach and use the previous study as a prior probability (n.b., you probably don't want to use the actual p-value from the previous study)? Then you just get an updated probability for that gene. Note that this would probably make sense on the unadjusted rather than adjusted p-values (you'd then adjust the results of the applicable genes for multiple testing).

Anyway, that's the first thing that comes to my mind.

[scottdaniel]

Why not just use a more bayesian approach and use the previous study as a prior probability (n.b., you probably don't want to use the actual p-value from the previous study)? Then you just get an updated probability for that gene. Note that this would probably make sense on the unadjusted rather than adjusted p-values (you'd then adjust the results of the applicable genes for multiple testing).

Anyway, that's the first thing that comes to my mind.

That sounds good. Can you elaborate? I understand the basic idea of Bayesian analysis (make no assumptions on the distribution, everything is based on one probability at a time) but have never really applied it. I think it's used in limma? I really wouldn't no where to begin but I could take a stab at it with R. Was there a software package you would suggest to attempt this with?

[Gordon Smyth]

A simple and good way to handle genes of a priori interest is by limiting the multiple testing adjustments. Normally you will run topTags(fit) on the whole dataset, where fit is output from either glmLRT or exactTest. When you have genes a prior interest, possibly because they were DE in a previous study, then you can run topTags on the subset:

i <- fit$genes$ID %in% PriorInterest
topTags(fit[i,])

The FDR adjustments for these genes will now be more lenient than when adjusting overall all genes. This is statistically valid when validating the a priori genes.

[scottdaniel]

A simple and good way to handle genes of a priori interest is by limiting the multiple testing adjustments. Normally you will run topTags(fit) on the whole dataset, where fit is output from either glmLRT or exactTest. When you have genes a prior interest, possibly because they were DE in a previous study, then you can run topTags on the subset:

i <- fit$genes$ID %in% PriorInterest
topTags(fit[i,])

The FDR adjustments for these genes will now be more lenient than when adjusting overall all genes. This is statistically valid when validating the a priori genes.

I will try this, thanks.