Hi,

I recently did RNAseq using a phosphoTRAP protocol to determine which cell types are activated in the brain following exposure to a stimulus. So, i have paired samples where each pair is the baseline RNA expression for the brain (input) and then the RNA that was attached to phosphorylated ribosomes being actively translated (ip). My intention is to distinguish differences in my ip samples across my three different stimulus conditions. To do this, i used edgeR and modeled just the ip gene counts with a design matrix that includes batch and stimulus conditions and set the offset as my input gene counts. However, I am worried that because i manually set the offset to my input gene counts, other typical normalization factors are omitted from the analysis and this may be problematic. The ultimate goal is to compare ip gene counts across stimulus conditions, while taking into account that samples are paired and that for each ip sample there is a corresponding input baseline value. Is the method i proposed a good way to do this, or is there a better way to address the question i am posing? Thank you!

Did you convert the input counts to log scale before setting the offset? Setting linear counts as an offset wouldn't make sense. However, I feel uncomfortable with the idea of using input counts as offset for many reasons. It ignores the variability in the input, assumes that ip counts are have slope 1 vs input and, mostly importantly, assumes that ip library sizes are exactly proportional to input library sizes, which I assume is not true.

I'm not familiar with this sort of data and I can't recommend a definitive analysis. But I would be inclined to simply do a regular paired analysis in edgeR with the input and ip as the two samples in each pair. Or alternatively you might also try edgeR's normalizeChIPtoInput() function.