Context: I've been doing some in silico research involving large-scale simulations with various DGE packages. I detected an error (detailed below) in my simulations that seemed new to edgeR 4.2.1 and did not occur in earlier versions.

The following example might involve some non-sensical data, but I was able to reduce the original, meaningful, large-scale simulation dataset to this 3 by 3 one for debugging convenience.

count <- data.frame( c1=c(2,0,0), c2=c(1,2,4), c3=c(2,0,1))
norms <- c(0.65001, 1, 1.7)
groups <- c('case', 'control', 'case')
genes <- c('G1', 'G2', 'G3')
libsize <- c(1,1,1)
library(edgeR)
dglist <- DGEList(count, libsize, norms, group=groups, genes=genes )
library(stats)
dm <- model.matrix( ~ groups )
dglist <- estimateGLMCommonDisp(dglist, design=dm)
dglist <- estimateGLMTrendedDisp(dglist, design=dm)
dglist <- estimateGLMTagwiseDisp(dglist, design=dm)
fit <- glmQLFit(dglist, dm)

And this leads to the error:

Error in .compressOffsets(y, lib.size = lib.size, offset = offset) :
  offsets must be finite values

Some further observations:

• Weird enough, if "0.65001" was changed into "0.65", there wouldn't be any such error and the GLM fit would go through.
• If "4" in the count matrix was changed into "3", there wouldn't be this error.
• A cursory dive into the source code tells me that in .compressOffsets(), the offset is obtained from getOffset(dglist), which is completely specified by my inputs and should be equal to log(norms). Indeed, if I enter getOffset(dglist), I get the correct values -0.4307814 0.0000000 0.5306283. The error here, however, is only supposed to be raised when these offsets are infinite.
• This happens with edgeR version 4.2.1 on both my local PC (Windows) and server (Linux). But this did not happen with earlier version 4.0.x, as I remember.

Regards,

Hongjian

hjnwang [at] cmu [dot] edu

Thanks very much for the reproducible example. The bug has been fixed now in edgeR 4.3.17 on the Bioconductor developmental repository and in edgeR 4.2.2 on the release repository.