Design Matrix with EdgeR or DESeq2

Question

[reui879]

Hello,

I'm trying to create a design matrix using edgeR or DESeq2. I have a time course experiment with includes different does of a drug and different cell lines. For each cell line, I'd like to get DEGs for treated vs untreated but I am unsure of how to do this while taking the dose and time into account.

This is the metadata:

cell_line drug dose time    
A untreated 0 0
A treated 1 2
A treated 1 6
A treated 1 24
A treated 1 48
A treated 3 2
A treated 3 6
A treated 3 24
A treated 3 48
B untreated 0 0
B treated 1 2
B treated 1 6
B treated 1 24
B treated 1 48
B treated 3 2
B treated 3 6
B treated 3 24
B treated 3 48

[Kevin Blighe]

Cross-posted: https://support.bioconductor.org/p/p134163/

[i.sudbery]

You are likely to get a better reply to this on bioconductor.

[rpolicastro]

Your design matrix will depend on what questions you are trying to answer and the presence of confounders in those comparisons. Using all of your factor levels, can you describe some questions you want answered?

Answer 1

[i.sudbery]

You can't directly compare treated to untreated while taking time and dose into account with this data because drug, time are confounded with drug. In-fact untreated is just a pseudonym for time=0, dose=0.

If you looking for genes that are in anyway affected by drug treatment in each of the cell lines, I'd probably start by dividing this into two studies, some for cell line A, and one for cell line B. I'd then probably use an LRT-type test comparing the design ~time + dose to the design ~1, or perhaps ~time + dose + time:dose to ~1.

It gets more complicated if you want to compare the effects in the two cell lines, but possibly I'd be looking at: ~ time + dose + cellline + time:dose + time:cellline to ~time + dose + cellline + time:dose. Although I suspect this design is sufficiently complex that unless you've got lots of replicates of each of those conditions, you won't have suffient power to detect differences.

[reui879]

Thanks! How would I compare the design ~time + doe + time:dose to ~1?

[i.sudbery]

See the LRT section of the DESeq2 manual: http://bioconductor.org/packages/devel/bioc/vignettes/DESeq2/inst/doc/DESeq2.html#likelihood-ratio-test