Hi,

I am using DESeq2 for analyzing an RNASeq Experiment. The details of the setup are as follows:

I have two treatments for which I want to test for differential expression. However, when inspecting the data, the variability of the counts in one treatment group is considerably higher than the variability in the other group (for same mean).

If I understand the DESeq2 documentation correctly, the DESeq function fits one dispersion parameter per gene, not allowing for sample groups with different dispersion. 

a) Is this understanding correct?

b) Is there a way to allow for sample-group specific dispersions? In a linear model setting, this would usually be achieved by setting different weights for the samples in different groups.

Thanks!

Holger

Not sure if it has hit your radar, but the limma/voom folks have recently published a paper that describes the voomWithQualityWeights function, which allows you to achieve this behavior to analyze rna-seq data with sample level weights:

• Paper: Why weight? Modelling sample and observational level variability improves power in RNA-seq analyses
• Supplementary Information

hi Holger,

a) yes, you're right. we fit one dispersion estimate per gene, so locking the relationship Var = mean + dispersion * mean^2.

b) There is not a way to have group-specific dispersion values with DESeq2. Off the top of my head, I think baySeq offers different dispersion estimates per group: getPriors.NB(..., equalDispersions=FALSE)