[BioC] Multifactorial analysis with RMA and LIMMA of Affymetrix microarrays

[Gordon Smyth]

At 02:48 AM 18/03/2004, Jordi Altirriba Gutiérrez wrote:
>Thank you very much Gordon for your quick answer!
>My phenoData is:
>>pData(eset)
>     DIABETES TREATMENT
>DNT1     TRUE       FALSE
>DNT2     TRUE       FALSE
>DNT3     TRUE       FALSE
>DT1      TRUE        TRUE
>DT2      TRUE        TRUE
>DT3      TRUE        TRUE
>SNT1    FALSE       FALSE
>SNT2    FALSE       FALSE
>SNT3    FALSE       FALSE
>ST1     FALSE        TRUE
>ST2     FALSE        TRUE
>ST3     FALSE        TRUE
>
>(DNT=Diabetic untreated, DT=Diabetic treated, SNT=Health treated, 
>ST=Health untreated)
>
>I want to know the genes characteristics of the diabetes, the treatment 
>and the treatment + diabetes. Moreover when I analyse my data with SAM and 
>I compare Health treated vs the Health untreated I don't see many 
>differences, but when I compare the Diabetic treated vs the Diabetic 
>treated I see a lot of differences, so is correct to apply a 2 x 2 
>factorial design?

You simply need to fit a model which contains four coefficient which 
distinguish your four groups. The classical 2x2 model is just one 
particular parametrization you can use:

design <- model.matrix( ~ DIABETES*TREATMENT, data=pData(eset))
fit <- lmFit(eset, design)

>Is LIMMA the correct tool to answer my questions? If it is the correct 
>tool, how can I do a factorial design matrix (if to do a factorial design 
>is correct)? (Robert Gentleman has suggested me to use the factDesign).

You're just fitting a linear model, so the above calculation is exactly 
equivalent to what factDesign does, although probably a bit faster. I would 
use limma myself because it allows you go on to do empirical Bayes 
moderation of the residual standard deviations etc, which I think it 
important, but Robert may be able to make a further case for factDesign.

Cheers
Gordon

>Thank you very much for your time, patience and your suggestions.
>Yours sincerely,