GLM in DESeq

[SMcTaggart]

Hello;

I am using the nbinomFitGLM and nbinomGLMTest in DeSeq, which have worked well. However, I would like to know if both of the factors in my model (in my case genotype and treatment), as well as their interaction, explain a significant proportion of the variance in the model. Does anyone know the commands that would get at this information (i.e the type of output that one could access with summary.glm)?

Many thanks,

Seanna

[Emanuel Heitlinger]

Hi Seanna,
Not sure wether this is really answering you questions, but I have fun at the moment playing around with the R-package vegan:
You can use principal component analysis (e.g. in the rda function in vegan) on you set of DE-genes. Plot the results to see if one of the PCs is somehow corresponding to the original factors you used to identify the DE-genes.
You can also use rda or cca to constrain a component to one of your factors. This gives you a variance explained by these factors and the variance explained by additional PCs.

Emanuel

P.S. drop me a mail if you want code-examples. I just wanted to post this here as I am not sure, that what I am saying/doing makes sense... Maybe somebody more competent will to come by and debunk it as nonsense ;-)

[Simon Anders]

For an analysis of deviance, you do a number of fits, progressively including more factors:

fit0a <- nbinomGLMTest( cds, count ~ genotype )
fit0b <- nbinomGLMTest( cds, count ~ treatment )
fit1 <- nbinomGLMTest( cds, count ~ genotype + treatment )
fit2 <- nbinomGLMTest( cds, count ~ genotype + treatment + genotype:treatment )

Now, you can compare to fits in order to see whether the factor in which the two differ is significant. For example, to see whether treatment has an effect, you do

pvalT <- nbinomGLMTest( fit1, fit0a )

Similarly, comparing fit1 with fit0b shows whether genotype has an effect and to test for the interaction, compare fit2 with fit1.

Remember to adjust for multiple testing. Also be aware that once you have significant interaction, it can become debatable whether the the concept of significant main effect for the genes with interaction is still meaningful.

[Emanuel Heitlinger]

Did I misunderstand "if both of the factors in my model [...] explain a significant proportion of the variance in the model" ?

I thought Seanna wanted to know how much variance the significant factors (or factor combination) explain in the respective significant subset of the data in total (all genes).

Emanuel