[BioC] Block x Treatment interaction test with DeSeq

Simon Anders anders at embl.de
Mon Nov 28 14:05:43 CET 2011

Hi Miguel

On 11/28/2011 10:16 AM, Miguel Gallach wrote:
> I am interested in testing Genetic x Environment interaction for RNA-Seq
> data. However, if I understood correctly, DeSeq do not test for
> interaction, right?

Of course, DESeq can test for interaction.

> My experimental design consists in comparing expression of two different
> populations, two replicates per population (Pop A1, Pop A2 vs. Pop B1 and
> Pop B2), at different temperatures (T1 and T2). The next table represents
> my data frame and my procedure:

You table got a bis messed up, so I rather write down how I understood 
your design data to look like.

library	population	temperature
A1.T1	A		T1
A2.T1	A		T1
B1.T1	B		T1
B2.T1	B		T1
A1.T2	A		T2
A2.T2	A		T2
B1.T2	B		T2
B2.T2	B		T2

This is now assuming that by different "population", you mean something 
like different genotype or ecotype or colony, and that of either of 
these, you have four aliquots/cultures/etc, two of which you let grow at 
temperature T1 and the other two at T2.

If you give DESeq the columns 'population' and 'temperature' as a 
'design' data frame and the call 'estimateDispersions' with 
'method="pooled"', it will consider those sample pairs which have the 
same population and temperature values as replicates and estimate 
dispersions accordingly. Then, you can fit models like

fit0 <- nbinomFitGLM(cds, count ~ population)
fit1 <- nbinomFitGLM(cds, count ~ population + treatment )
fit2 <- nbinomFitGLM(cds, count ~ population + treatment + 
population:treatment )

The last one is equivalent to

fit2 <- nbinomFitGLM(cds, count ~ population * treatment )

Now depending on what you want to test, you may for example do

pvals <- nbinomGLMTest( fit1, fit0 )

if you consider 'population' as a blocking factor (i.e., a nuisance 
covariate whose influence you want to get rid of), and

pvals <- nbinomGLMTest( fit2, fit1 )

if you consider both factors as biologically interesting and want to see 
their interactions.


More information about the Bioconductor mailing list