Dear all, 

I am analyzing my RNA-seq results with DESeq2 and I have a question about model matrix design and interpretation of the results. I have two cell types in two conditions, that is 4 groups in total. I would like to know the difference if I use factor design or pair-wise group comparison, and the interpretation of it. 

Factor design: 

data.group$cell <- c(“A”, “B”)

data.group$condition <- c(“cond1”, “cond2”)

design = condition + cell + condition:cell

…..

And I am able to get the differential gene lists from: 

1. results(DESeq2data, contrast = c(“cell”, “B”, “A”))

2. results(DESeq2data, contrast = c(“condition”, “cond2”, “cond1”)

3. results(DESeq2data, contrast = list(c(“cell_B_vs_A”, “conditioncond2.cellB” 

If I do it in Pairwise comparison: 

data.group <- c(“cellA.cond1”, “cellB.cond1”, “cellA.cond2”, “cellB.cond2”)

….

4. results(DESeq2data, contrast = c(“data.group", “cellB.cond1”, “cellA.cond1”))

5. esults(DESeq2data, contrast = c(“data.group", “cellA.cond2”, “cellA.cond1”))

6. results(DESeq2data, contrast = c(“data.group", “cellB.cond2”, “cellB.cond1”))

My question is: 

What does the log2FoldChange mean from result 1.? Is it comparing expression changes “B”/“A", irrespective of “condition”? Or is it comparing expression changes “B”/“A" under cond1 (essentially the same as 4.)? 

Using attr(DESeq2data, “modelMatrix”) to analyze the contrast, it seems the latter. But the results from 1. and 4. do not agree 100%. I can find genes that are differentially expressed in 1. but not in 4., and vice versa. Any explanation why? And similar differences apply to 2. vs 5. , 3. vs 6.. 

I have turned off Cook’s cutoff and IndependentFiltering, so I am sure the difference is due to the model itself.  

Thanks a lot for your help! 

Best, 

Shin

"What does the log2FoldChange mean from result 1.? Is it comparing expression changes “B”/“A", irrespective of “condition”? Or is it comparing expression changes “B”/“A" under cond1 (essentially the same as 4.)? "

It is comparing B vs A for condition = 1. It would be exactly the same as 4 if betaPrior=TRUE. But the shrinkage is used differently in the two settings. With the interaction design, the shrinkage is applied to only the interaction term. This is a necessary modification, or else shrinkage of main effect terms would increase the interaction terms (we discuss this in the vignette and in the paper). For the grouped design, the shrinkage applies to all groups equally. I recommend if investigators are interested in the interaction term, to use the first design, but if investigators are interested in making pairwise comparisons of all groups to use the second design.

Hi Mike, 

Thanks so much for your quick reply, and your explanation is very clear. I am using default for betaPrior, so I assume it is set to TRUE. 

Best, 

Xin