Hi,

I am trying to create another design matrix with 4 groups in mind:

1. DiseaseA+DiseaseB, 2. DiseaseA, 3. DiseaseB, 4. Control.

I could create appropriate design matrix to basically perform several analyses of interest such as finding differentially expressed genes between any two groups, However, what I want to find is the genes that are differentially expressed between Disease A and Disease B, but I would like to incorporate DiseaseA+DiseaseB group into this comparison. I have tried various ways such as incorporating a table such as below into creating a design matrix:

DiseaseA     |     DiseaseB

         1         |           1

         1         |           0

         0         |           1

         0         |           0

but it is not doing what I thought it would do. Any suggestions will be very helpful.

Thank you.

Well, you need to make up your mind as to what you want to do. If you want to compare between any two groups, then use a one-way layout where each disease status (control, disease A only, B only, or A + B) forms its own group. As an example, here's how I would do it if the target information were present in has.A and has.B:

has.A <- rep(c("Y", "N"), each=4)
has.B <- rep(rep(c("Y", "N"), each=2), 2)
groups <- factor(paste0("A", has.A, ".B", has.B))
design <- model.matrix(~0 + groups)
colnames(design) <- levels(groups)

You can then compare between any set of groups, e.g., between disease A and B using makeContrasts:

con <- makeContrasts(AY.BN - AN.BY, levels=design)

If you want to use the A + B information in the comparison between A and B, then you need an additive model:

design <- model.matrix(~ has.A + has.B)
con <- makeContrasts(has.AY - has.BY, levels=design)

However, this assumes that the effects of A and B are additive in the group with both diseases. One can easily imagine that this will not be the case in reality, e.g., due to epistatic effects or other interactions between the disease states. If the assumption is violated, you will probably get a decrease in power because the dispersion estimate gets inflated by the poor fit. This may end up being more damaging than just leaving the A + B samples out of the comparison in the first place.

In short, I would go for the one-way layout. It's more flexible in terms of the comparisons you want to do, and also in terms of avoiding the additivity assumptions in the other model. That said, I hope the design matrix in your original post is just an example, and that you actually have replicates within each group, otherwise you'll end up with other problems.