In DESeq, should one estimate sizeFactors and Dispersion in a pairwise fashion?

[carmeyeii]

When using DESeq, does one also need to import two conditions each time to do the estimatesizefactor and estimatedispersions?

Can one import all conditions (>=3) together, and estimante sizefactors and dispersion together, and then do the negative binomial test in a pairwise manner?

Thanks!

[amcloon]

I'm not sure if you worked out an answer to this yourself, but my observation is that you can set up the count table with either a subset of conditions (just A and B) or all conditions (A, B, C) with all suitable replicates and do the size estimates, dispersion estimates off of either size of count table. Then, for the actual differential expression calling, you just specify which you are interested in...
>res=nbinomTest(cds,"A","B")

Don't ask me, though, which size count table you -should- use for dispersion estimation, though.

[sdriscoll]

If your conditions are very different, having more conditions in the count matrix when estimating dispersions than you'll use in a pair wise test can severely impact the dispersion estimates relative to the tested pair's actual dispersion. So, yes, separate out the paired conditions prior to estimating dispersions.

[chadn737]

If you have sufficient biological replicates you can estimate the dispersions for each condition separately by method="per-condition"

[sdriscoll]

That's right. I should add that what I noted is true if you're using the "pooled" or "blind" methods for dispersion estimation.