[BioC] DEseq2 metagenomic analysis without replicates

Simon Anders anders at embl.de
Tue Jan 14 18:18:02 CET 2014

On 14/01/14 18:03, Kristina M Fontanez wrote:
> I agree that for the treatment effect, combining the dispersions across
> depths is the way to proceed with these data. I think your reasoning for
> depth as an ordered factor is sound and likely applies to the depth
> profile I am working with. I attempted to follow your suggestion
> (snipped below) to create a design that allows me to use depth as an
> ordered factor. However, I found it confusing that the design is
> ~treatment + depth. 

> With depth as the second variable, doesn’t that mean
> that the dispersions will be combined across treatments? I thought the
> goal was the combine dispersions across depths so that the treatments
> could be compared.

Oh, I was just one step ahead.

To see the effect of treatment, just leave out the depth factor, i.e.,
use "~ treatment". Only once you want to see the effect of depth, use "~
treatment + depth", with depth as a numeric variable (not an ordered
factor, I may have been a bit unclear there).

> In the code below I used a ~TREATMENT+DEPTH design with depth as
> numerical vectors and the “None” (seawater) treatment as the base level
> of the treatment factor. My main concern is that the output results in
> hundreds of differentially abundant taxa (344 OTUs for the Dead vs
> Seawater comparison with p-adjusted values <1e-5). These results include
> both independent Filtering and the cooksCutoff (by default). Given the
> really high numbers I am inclined to think that a large number of these
> are false positives and we have modeled the dispersion incorrectly. What
> do you think of this assessment?

Maybe try again with ignoring "depth" altogether, even though, in
principle, including ti should not be wrong. Also have a look at the
estimated fold changes. Maybe they are significant but very small. An MA
plot would help here, and also allow to check the normalization.

I'm still puzzled what kind of treatment "live" and "dead" might be. I
find it hard to come up with a guess what kind of treatment might have
no effect on most taxa, which nevertheless has such a drastic name as
"dead". Can you explain a bit more about the biology? (By private mail,
in case you don't want to expose unpublished work on a public mailing list.)


More information about the Bioconductor mailing list