This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Finding significantly differentiall expressed genes in RNA-Seq with one sample

Hi friends

If in a RNA-seq study we have just one sample per condition,

what should we do for finding those genes which expressed differentially? (I mean those significant)

How much it'll be reliable to use one sample in DE study?

Last question, Which statistical test is proper for finding P-Value in this situation (1 sample)?

Tnx e laat ,

rna-seq p-value

You can maybe (I strongly emphasize "maybe". I do not recommend this) try to get a common biological variance by only looking at the subset of your data that shouldn't be DE, IE. house keeping genes. But how you define house keeping genes is another whole set of problems.

Like Devon said, you need more replicates.

I like Devon's answer, "There is no valid statistical test for your case, at least unless you want to assume no biological variance." And in case you decide to assume no biological variance, I would add: There is no valid statistical test for your case. Nonetheless, sometimes one has to work with the data at hand. Have a look at the edgeR users guide, specifically the section titled: "What to do if you have no replicates."

1 answer

There is no valid statistical test for your case, at least unless you want to assume no biological variance. The best you can do is rank genes by fold-change, likely using something like GFOLD. Before spending too much time on that, of course, you might just rethink the experiment.

I know about problems with experiment design but i have result which produce by sequencing company and they gave me fold change with P-value and FDR , but how ? i don't know. !1!1

They're likely assuming Poisson variance, so the p-values are meaningless.

Sometimes it just isn't possible to get proper replicates, but RNA-Seq data can still be useful. Our group has several rare disease projects for example where we may only be able to get RNA-Seq data on one patient's relevant tissues for a host of real-world conditions that we can't do anything about. But in that case you need a bunch of other data as well and yes, the p-values are essentially meaningless.

Well, in those cases you can use multiple control samples and at least get an idea about the background variance. There actually are cases where it's completely reasonable to use an unreplicated design, such as prenatal diagnostics of heritable disorders from maternal blood, but those are more the exception than the rule.

Anyway, your comment regarding multiple experiments is pivotally important.

Log in to answer this question.