View Single Post
Old 11-25-2010, 03:26 AM   #7
Simon Anders
Senior Member
 
Location: Heidelberg, Germany

Join Date: Feb 2010
Posts: 994
Default

I've explained this in a number of posts before, so I just repeat the core points.

- If you use FIsher's exact test or something similar, you don't need any replicates because it cannot accommodate for them. The results, though, will be wrong, especially for strongly expressed genes.

This is because Fisher's exact test tests whether two samples differ in the concentration of a given transcript. This is, however, not the question you want to ask. What you want to know is whether the difference between two samples with different treatment is stronger than what you expect to see between two samples that are replicates, because otherwise, you cannot attribute the difference to the treatment.

This criticism also applies to cuffdiff, at least to the version described in the paper. (There is a new version of cuffdiff that allows for biological replicates but there is no documentation on its method yet, and hence it is unclear whether it now asks the relevant question.)

- If you have many replicates, use a t test.

- With only two or three replicates, you need to pool across genes, i.e., assume that similar genes have similar variance. Our DESeq package assumes that genes with similar expression strength have similar variance, and so pools information from these in order to get a reasonable estimate of biological variability, which is then used for the test.

Simon
Simon Anders is offline   Reply With Quote