I'm running CAMERA on one of my datasets with gene set collections imported from MSigDB and graphite, and I am getting NaN results for a large fraction of gene sets that I'm testing (around 50% of pathways for some collections). This is odd because I don't remember that happening a few versions ago the last time I used CAMERA on the same gene sets. I've prepared a test case that demonstrates the problem. You can download the script and data file from here: https://www.dropbox.com/sh/dp5aheqzvf2rij5/AADuA68Hk5BdRIJpL4H_knbHa?dl=0 Simply put them in the same folder and run the script, which is just this:

library(edgeR)
testcase <- readRDS("camera-nan-testcase.RDS")
result <- with(testcase, camera(dge, pathway, design, makeContrasts(contrasts=contrast, levels=design)))
print(result)

The result should be:

Warning message:
Zero sample variances detected, have been offset
data.frame with 1 row and 4 columns
        NGenes Correlation   Direction    PValue
     <numeric>   <numeric> <character> <numeric>
set1        40         NaN          Up       NaN

I have no idea what's causing this. Is there something about this gene set that makes it incompatible with CAMERA, or is there an issue in my data?

As you have figured out, this occurs for gene sets that contain at least one gene with no counts. A gene with no counts is essentially missing. It has undefined differential expression, undefined dispersion and undefined correlation with other genes. Hence the differential status of the whole gene set is also undefined, hence the camera output.

In principle, we could put a check in the camera method for DGEList objects to remove genes with no counts, and that would give you non-missing camera results. But it would be much better for you to filter out genes with very low counts before you do the dispersion estimation step prior to running camera. As you know, we recommend that edgeR users always filter genes with very low counts.