gene-level effects in DEXSeq

[bswhite]

Hi,

I've just updated to DEXSeq 1.10.8. The manual claims that DEXSeq
fits a GLM for each gene (section 3.4 Testing for differential exon
usage), which I would expect. However, the formulae presented in that
section do not include any gene-level effects:

e.g., full model ~ sample + exon + condition:exon

In previous versions of DEXSeq incorporating the gene effect would
have been accomplished by selecting out the (exon) count rows
corresponding to a gene and performing the test individual on those
rows. i.e., with testGeneForDEU. But I don't see any similar partitioning
of the data by gene in DEXSeq 1.10.8. Can someone verify that it
is indeed fitting the GLM on a per-gene basis and how it is doing so?

[dpryan]

It'll fit whatever design you give it with a GLM. You typically don't need to explicitly add "condition" since it's a linear sum of "sample", which would normally make the resulting model matrix rank insufficient.

[bswhite]

Thanks. But I wasn't understanding how the gene level effect was accounted
for given that I thought each row of the _original_ matrix of exon counts
was processed independently. That last assumption was incorrect
and lead to my confusion.

Previous versions of DEXSeq aggregated rows sharing a gene in testForDEU. In this version that aggregation into "this" and "other" exons happens in DEXSeqDataSet. So that each row of the augmented matrix has the
original counts for the exon and the summed counts for all other exons.

This is started pretty clearly in the manual, I just missed it (e.g., section 3.1 Loading and inspecting the sample data). Hence, the gene level effect is
absorbed into the 'sample' term (as stated in Section 3.4 Testing for differential expression).