Comparing two conditions DESeq2

[rubbertjes]

I would like to compare several conditions to a reference condition using DESeq2. Currently I am using the contrast option in the "results"-function as follows:

dds$condition <- relevel(dds$condition, reference)
#Do the calculation
dds <- DESeq(dds, betaPrior=FALSE)
res <- results(dds, contrast = c('condition',selected_condition,reference))

However in a manual I found that they actually take a subset. http://www.ebi.ac.uk/training/sites/...d-s.anders.pdf I found

#First we subset the relevant columns from the full dataset:
> dds <- ddsFull[ , colData(ddsFull)$treatment %in% c("Control","DPN") &
+ colData(ddsFull)$time == "48h" ]

My question is: Are both approaches still valid? And, what is the benefit of one over the other?

Thanks,

Rob

[blancha]

I don't know if there is a difference in the results.

I would however stick to the most recent manual, which was written in 2014, whereas the manual you link to was written in 2013.

http://master.bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.pdf

DESeq2 has been updated many times, and the contrast option is currently the recommended method.
I suspect the new implementation is just a more elegant way of accomplishing the same objective, but I could be wrong.

[Michael Love]

"Are both approaches still valid? And, what is the benefit of one over the other?"

Yes both approaches are valid.

I typically recommend running DESeq() on an object with samples from all the groups. The more samples, the better the estimates of dispersion (under the model assumptions of a dispersion parameter for each gene for all samples).

One situation where it might be useful to subset is if the groups have very different within-group variance. You can see this in a PCA plot (see vignette for how to). Consider an experiment with groups A,B,C. If groups A and B have small within-group variance, and group C has much larger within-group variance, then subsetting to only A and B for a B vs A comparison might be more powerful, than including C, which would drive up the genes' dispersion estimates.

[blancha]

@rubbertjes, Michael's answer is obviously much better than mine.

@Michael, that is an interesting point. I do have several analyses where the variance varies greatly from one group to another. It's always good to remember that one should take into account that the estimates of dispersion are computed from all the samples, and not just the samples from the groups being compared when using the contrast option.

[rubbertjes]

"Are both approaches still valid? And, what is the benefit of one over the other?"

Yes both approaches are valid.

I typically recommend running DESeq() on an object with samples from all the groups. The more samples, the better the estimates of dispersion (under the model assumptions of a dispersion parameter for each gene for all samples).

Thanks! That was something I was also thinking.

One situation where it might be useful to subset is if the groups have very different within-group variance. You can see this in a PCA plot (see vignette for how to). Consider an experiment with groups A,B,C. If groups A and B have small within-group variance, and group C has much larger within-group variance, then subsetting to only A and B for a B vs A comparison might be more powerful, than including C, which would drive up the genes' dispersion estimates.

There no reason for me to a priori assume this for this particular experiment. I guess this is something I should always check...
In this case I have 6 groups in my study, where I am only comparing 2 at a time.