Search
0
3.1 years ago by
Silva0
Finland
Silva0 wrote:

Hi,

in my data, I have two groups: a treatment group and a control group. Both groups have two biological replicates (i.e., two replicated populations): "1" and "2". I would like to nest the replicates within the main groups OR set the replicates as a random effect when running the DESeq2. I am not particularly interested in differences between the replicates (we would like to assume that there are none) but I have to take replicate into account in the model. However, I don't know how to run a model with a random effect in it and I just set the group and replicate as main effects. When doing that, can I simply have two columns in my design: one for treatment ("treated","control") and one for replicate ("1","2")? Can I be sure then that in my model (counts ~ replicate + treatment) the replicates are nested within the treatment? Or should I code my replicates as "1", "2", "3", "4" ("1" and "2" within "treated" and "3" and "4" within "control")?

This is probably a silly question but I would really like to be sure that my design is correct.

modified 3.1 years ago by Michael Love18k • written 3.1 years ago by Silva0
3
3.1 years ago by
Michael Love18k
United States
Michael Love18k wrote:

DESeq2 only allows for fixed effects. Because you are interested in comparing *across* replicates, you can't simultaneously fit the replicate and condition effects (e.g., the treatment effect is collinear with replicate 3 + 4). Your choices are then: (i) you can assume that the treatment effect is (rep3+rep4)/2 - (rep1+rep2)/2, which seems reasonable, or (ii) you can just fit the condition effect (where variance across replicates within condition will increase the dispersion estimates).

For (i) you would use a design of ~replicate and then:

results(dds, contrast=list(c("replicate3","replicate4"),c("replicate1","replicate2")), listValues=c(0.5, -0.5))

Also, I point out when people have experimental designs like this that with limma you can inform the model of correlations among samples (e.g. samples from replicate 1 are correlated) but without including a fixed effect in the model. See the duplicateCorrelation function in the limma package.