Question

DESEQ2 design matrix for double knockout interaction with condition

0

Entering edit mode

user9635643 • 0

@user9635643-20829

Last seen 5.7 years ago

Hi,

I have data with a wt control, knock outs of two genes (geneA, geneB), and a double knockout (KO of both genes). From what I can understand of design matrices, the following should give me the interaction between geneA and geneB: model.matrix(~ geneA + geneB + geneA:geneB)

However, I have the same data under two different conditions, and I want to find genes which interact (i.e. change more than expected if they are assumed to be additive) upon changes in condition. More specifically, I want to see how geneA and geneB interact when the cells are in a given condition (if any). I feel knowing which genes only change/change more than expected when both genes are KO whilst in conditionB vs conditionA would give me the best chance at answering this.

model.matrix(~ genotype + condition + genotype:condition) gives me genes which are specific to that genotype interacting with the condition. But in this model, double knockout is counted as a genotype of its own, as opposed to geneA knockout + geneB knockout. In simple 1's and 0's, I would like to represent the genotypes as follows, where the columns are my samples, and the rows indicate presence (1) and absence (0) of the gene:

        geneA_KO.   geneB_KO.   double_KO.
geneA.   0          1          0     
geneB.   1          0          0

The closest I got to was model.matrix(~0 + (geneA*geneB)^2 + condition + (geneA*geneB)^2:condition)) but I'm not sure if this would correctly model the data - there are a lot of terms and not all of them are relevant. The idea was (geneA*geneB)^2 would model the interactions between the two genes, condition would take care of the effect of going from conditionA to conditionB, and the last term (geneA*geneB)^2:condition would model how each genotype interacts with the change in condition (and I'd be mostly focussed on the one with the double knockout).

Can anyone help me model this, please?

deseq2 model.matrix design matrix • 1.2k views

ADD COMMENT • link updated 5.7 years ago by Michael Love 43k • written 5.7 years ago by user9635643 • 0

0

Entering edit mode

I'm not in favor of support site posts that trigger emails to multiple package authors (unnecessarily). Here you've tagged 'limma' so that those package authors are automatically emailed, although the title of your post is DESeq2 design matrix. The approaches for the different packages are sometimes different, so then you're just asking many people to reply to your request at the same time.

ADD REPLY • link 5.7 years ago Michael Love 43k

0

Entering edit mode

Thanks for letting me know. I'll be more careful in the future not to tag more than one package!

ADD REPLY • link 5.7 years ago user9635643 • 0

0

Entering edit mode

Thanks for letting me know. I'll be more careful in the future not to tag more than one package!

ADD REPLY • link 5.7 years ago user9635643 • 0

0

Entering edit mode

I've removed the limma tag because this appears to be a DESeq2 question.

I will say that squaring the formula term (geneA*geneB)^2 has no meaning for factors. It is identical to geneA*geneB. The formula you've written is just a very complicated way of writing geneA*geneB*condition.

ADD REPLY • link 5.7 years ago Gordon Smyth 52k

score 1 · Accepted Answer · 2019-05-20

If you want to find where the interaction term changes across condition, you distribute condition throughout your first design:

~condition + geneA + geneB + condition:geneA + condition:geneB + geneA:geneB + condition:geneA:geneB

There may be condition-specific differences in the level of gene A or B, which are covered by the 4th and 5th terms in the design, and the final term is the one telling you if the interaction changes across condition more than expected by additive differences.