DESeq2: Difference between condition+type vs. 3 conditions

[mevers]

Dear all.

I am unsure about how to use DESeq2 in the case of 3 conditions vs. 2 conditions + 2 types. Assuming I have the following design table

          condition    type
sample1   A            T1
sample2   A            T1
sample3   B            T2
sample4   B            T2
sample5   A            T2
sample6   A            T2

I am unsure about how this would be treated differently from

          condition
sample1   A:T1
sample2   A:T1
sample3   B:T2
sample4   B:T2
sample5   A:T2
sample6   A:T2

The second design table describes a 3-condition scenario.

Now, obviously one would be interested in a detailed analysis of the counts for

1. A:T2 vs. B:T2 (since they have the same type but a different conditions), and potentially
2. A:T2 vs. A:T1 (since they have the same condition but different types).

Question 1: If I reduce the problem to that of a 3-condition no-type design table, is this correctly taken into account?

I know I would have to re-factor the columns of the 2nd matrix to reflect the correct order of fold changes that I want to calculate. So for example following re-factoring the levels as

levels=c("A:T2","B:T2","A:T1")

and performing a DESeq2 analysis

dds<-DESeqDataSetFromMatrix(countData = countData, colData = design, design = ~ condition + type);
dds<-DESeq(dds);

Question 2: I could calculate the fold changes of B:T2 wrt A:T2 and A:T1 wrt A:T2, is this correct?
I do get some issues with non-convergent dispersion fits, which I can get around if I call estimateDispersions manually with fitType="local".

Question 3: But what happens in the case of the 1st condition+type table? I am confused as to the output of DESeq2. What role does the type play in the differential expression analysis and/or the dispersion fitting?

Any help on this issue would be greatly appreciated.

Regards,
Maurits

[Simon Anders]

In your first table, the type is always the same. Is this a typo? If not, I'm not sure I understand your question.

[mevers]

Hi Simon.

Yes, that was a silly mistake, you are absolutely right. I've changed it now in the original post. It should have read

type=c("T1","T1","T2","T2","T2","T2")

Cheers,
Maurits

[Michael Love]

Question 1:

You can technically represent it either way, although I would recommend to keep the variables separate for the following reason: if you combined the variables (as in "A:T1"), then you cannot make a clean B vs A comparison. Instead you have a B:T2 vs A:T1 comparison which mixes the effect of B vs A and T2 vs T1.

Question 2:

Note that fitType is also an argument for DESeq()

Question 3:

Both variables are used for finding fitted means (mu in the GLM formula given in the reference manual and vignette). And then the fitted means mu is used to estimate the dispersion. Dispersion is a measure of how far the counts deviate from the mu for that sample. Both variables will have fitted coefficients (betas in the GLM formula) and you can extract tests for each variable of the null hypothesis that the coefficients are equal to zero. By default the results for the last variable is provided by results(). For more, see the section in the vignette on "Multi-factor designs" and the man page for results().