Fisher's exact test in RNASeq analysis

Question

[sibrunne]

In RNA-Seq analysis it is common to use tests analogous to Fisher's exact test to evaluate whether a gene is differentially expressed in two measured conditions. Fisher's exact test relies on compiling a 2x2 (or greater) table of outcomes x conditions. When applied to RNA-Seq, I was wondering what the 2x2 table consists? I would assume that the two different genes are the two columns, but what then are the rows? The actual data and the average read count in each condition, to test the gene versus a null hypothesis of random sampling of read counts? I would be happy for help clarifying this issue.

Answer 1

[Devon Ryan]

Kind of, typically one uses a negative binomial, which allows for replicates. If one wanted to do use a Fisher's test (don't do this) then the second row would be the total counts in the sample excluding those for the given gene being examined. But really, you need biological replicates, so just use one of the many (Bioconductor) packages that already allow for this.

[sibrunne]

Indeed the algorithms such as DESeq and EdgeR use the negative binomial. However, they also use tests that they describe as 'analogous to the Fisher's exact test'. So I'm not sure that the use of the Fisher's exact test is impossible when using the negative binomial distribution.

[Devon Ryan]

DESeq1 allowed that, but that's since been disabled in DESeq2. That works by ignoring groups when determining the dispersion. There isn't really a table ever made, but the distributions are similarish.

[sibrunne]

Thanks for your help. Could you please be a bit more specific? As far as I know: - DESeq uses the NB-distribution, whether or not it uses the Fisher's exact test. - DESeq says it uses a test analogous to Fisher's exact test. But if they do, you say they do not construct a 2x2 table? So is that how it is analogous but not identical?