Hello,

the edgeR manual (i.e. user guide that was last revised 25 September 2014) explains very nicely, with nice examples, what to do when there are no replicates.

For example, for at section 2.10, sub-section 2 (that is picking a dispersion value, based on your experience with similar data, and use that for exactTest or glmFit) it is given this as example :

> bcv <- 0.2
> counts <- matrix( rnbinom(40,size=1/bcv^2,mu=10), 20,2)
> y <- DGEList(counts=counts, group=1:2)
> et <- exactTest(y, dispersion=bcv^2)

How this example can be expanded to be used with mroast? Is it even advisable to "force" mroast for this case?

The same questions could be asked about the entire section 2.10.

Cheers,

Daniel

Having no replicates, the only gene set test you can use is geneSetTest. For example:

  et <- exactTest(y, dispersion=bcv^2, prior.count=5)
  geneSetTest(index, et$table$logFC, alternative="up")

where index is the index of your gene set. Repeat for each gene set.

The ROAST procedure performs permutations by rotating points throughout the residual space. If you weren't able to estimate the dispersion in the first place, this means that you don't have any residual degrees of freedom in your experimental design. So, there'd be no residual space for the rotations to occur. Moreover, ROAST requires variance estimates to compute the (moderated) t-statistics, and this won't be available when there are no replicates.

In short, it's probably not advisable (or even possible) to force ROAST to handle this case.