[BioC] DESeq2 - precisions about interaction and fold change calculation

Michael Love michaelisaiahlove at gmail.com
Mon May 26 14:21:53 CEST 2014


hi Samuel,

On Mon, May 26, 2014 at 12:20 AM, samuel collombet
<samuelcollombet at gmail.com> wrote:
> Dear all,
>
> I am trying do performed some analysis with DESeq2, using a model with
> interaction therms, but I have a hard time understanding what DESeq2 is
> doing exactly (I am also not very familiar with R notation and have a
> hard time understanding it too, sorry).
>
> What I want to do correspond to the what is described in section "3.3
> Interactions" of the vignette " Differential analysis of count data -
> the DESeq2 package", i.e. one model with 3groups-2treatments, and one
> with 2groups-2treatments (2 different analysis).
>

I respond below with direct answers to you questions. We have a
slightly different implementation in DESeq2 using "expanded model
matrices", where each group gets its own effect rather than having
non-base levels compared to a base level. The log fold changes (LFC)
end up very similar compared to the standard analysis with a base
level, except that the shrinkage of LFC is symmetric with our
implementation. An exception is the 2x2 model with interaction term,
where we do not use expanded model matrices for simplicity and because
the LFC shrinkage happens to be symmetric in this case.

> - in the first example (p21, comparing "patient4.treatmentOHT" to
> "patient4.treatmentControl"), what will be exactly the numerator and
> denominators? How is the log2FoldChange calculated in this case?

The numerator is the OHT interaction term in patient 4 and the
denominator is the Control interaction term in patient 4. Subtracting
the later from the former gives the OHT vs Control interaction effect
for patient 4. If this is significant, it means the OHT vs Control
effect for patient 4 is not explained by the main OHT vs Control
effect.

> (I have
> read section 4.2 and the pre-submission on bioArxiv, but I am still not
> sure of what is calculated in this case ; I also do nto really
> understand the sentence " Note that the log2 fold change for treatment
> of OHT over control for patient 4 is the interaction effect above in
> addition to the main effect of treatment OHT over control", p21).
>

This sentence just reminds the user that the interaction terms are
added to the main effect terms: treatment effect for patient 4 = main
treatment effect + patient4 treatment interaction term. If the
interaction term is 0, then the main effect term is sufficient to
explain the fold changes due to treatment for that group. If the
interaction term is nonzero, the above formula gives the fold change
due to treatment for patient 4.

I will make a note to clear up this sentence in the vignette.

> - in the second example p22, with only 2groups and 2 treatment, to what
> contrast does "patient2.treatmentOHT" correspond to? The specific effect
> of treatment OHT on patient 2, taking into account the general effect of
> OHT?
>

Yes.

Mike



> Many thanks in advance!
> Samuel
>
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> Bioconductor mailing list
> Bioconductor at r-project.org
> https://stat.ethz.ch/mailman/listinfo/bioconductor
> Search the archives: http://news.gmane.org/gmane.science.biology.informatics.conductor



More information about the Bioconductor mailing list