Dear Simon thanks for your reply.

Your suggestion at the end of the mail, was my starting point (MTsD vs
MTcD). But in this case I was loosing the effect that the drug could add to
the model.

The only problem here is to remove the effect of the drug (That I can infer
from the comparison between WTsD vs WTcD. But this also add a new variable
... the mice).

So I don't know how to deal with it neither in edgeR nor in DESeq

may be :

A design without taking into account the mouseType (that is assuming, the
WT and the MT are more or less very similar), but using a overexpression
covariate:


data<-read.table(file="ALL_MTcD_9.5_NO_MTsD1",header=TRUE,row.names=1)
#
#Design for WTcD + MTsD + MTcD
#
Design<-data.frame(
  row.names=colnames(data[4:11]),
  treatment
=c("treated","treated","treated","untreated","untreated","treated","treated","treated"),
  OverExpression=c("NO","NO","NO","NO","NO","YES","YES","YES")
)

Design$treatment<-relevel(Design$treatment,ref="untreated")

cdsFull<-newCountDataSet(data[4:11],Design)

cdsFull<-estimateSizeFactors(cdsFull)

cdsFull<-estimateDispersions(cdsFull,sharingMode="maximum")


All_variables<-fitNbinomGLMs(cdsFull,count ~ treatment + OverExpression)

drug<-fitNbinomGLMs(cdsFull,count ~ treatment)


pvalsGLM<-nbinomGLMTest(All_variables,drug)
padjGLM<-p.adjust(pvalsGLM,method="BH")

experimento$pval<-pvalsGLM
experimento$adj.pval<-padjGLM

What do you think ?

Thanks in advance





2013/6/3 Simon Anders <anders@embl.de>

> Hi Eduardo
>
>
>  The problem here is that there is only 5 common miRNAs between DESeq
>> and edgeR.
>>
>
> This probably has little to do with the differences between edgeR and
> DESeq and much more with the fact that you ask the two tools two
> different questions.
>
> With DESeq, you did this:
>
>  #fit with the mouse and the drug
>> experimento<-fitNbinomGLMs(**cdsFull,count ~ mouseType + treatment)
>> #fit with the mouse + drug + interaction from mouse:drug
>> experimento_todos_factores<-**fitNbinomGLMs(cdsFull,count ~ mouseType +
>> treatment + mouseType:treatment)
>>
>> pvalsGLM<-nbinomGLMTest(**experimento_todos_factores,**experimento)
>>
>
> So, you compare a reduced model that accounts for the main effects that
> treatment and mouseType have with a full model that also includes an
> interaction between the two factors. So your null hypothesis is: "Both
> mouseType and treatment might influence the genes' expression strengths,
> but independently so, i.e. the effect of drug treatment on a gene's
> expression does not depend on the mouseType (and likely, the effect of
> the mouseType does not depend on whether the mouse was treated." Or, on
> other words: You are looking for gene, where the strength of the effect
> of the drug treatment is _different_ between the two mouse types.
>
> With edgeR, you do:
>
>  design<-model.matrix(~Mouse+**Treatment) fit<-glmFit(y,design)
>> lrt<-glmLRT(fit)
>>
>
> If I remember correctly, edgeR drops the _last_ factor to get a reduced
> model. So you are comparing with the model "~ Mouse" and if you do the
> same with DESeq, you will get similar results. This time, you simply ask
> which genes' expression is affected by the treatment (whilst controlling
> for mouseType), but you do not ask whether the strength of the drug's
> effect depends on mouse type.
>
>
>  The mutant has an insert to over express a gene of interest. and It's
>> a different mouse from the wild type because the wild type with the
>> insert did't born.
>>
>
> This is a very unfortunate experimental design. As your two mice differ in
> more than just the gene of interest, you will not be able to argue that the
> differences you find between the two mice is due to this gene and not one
> of the other genes in which they differ!
>
> If the mutation was lethal in the strain you tried first but not in the
> second strain, why did you keep using the first strain as control rather
> than using wild-type mice of the second strain as controls?
>
>   Simon
>
>
> ______________________________**_________________
> Bioconductor mailing list
> Bioconductor@r-project.org
> https://stat.ethz.ch/mailman/**listinfo/bioconductor<https://stat.ethz.ch/mailman/listinfo/bioconductor>
> Search the archives: http://news.gmane.org/gmane.**
> science.biology.informatics.**conductor<http://news.gmane.org/gmane.science.biology.informatics.conductor>
>



-- 
===================================================
Eduardo Andrés León
Tlfn: (+34) 91 732 80 00 / 91 224 69 00 (ext 5054)
e-mail: eandres@cnio.es        Fax: (+34) 91 224 69 76
Unidad de Bioinformática       Bioinformatics Unit
Centro Nacional de Investigaciones Oncológicas
C.P.: 28029                Zip Code: 28029
C/. Melchor Fernández Almagro, 3    Madrid (Spain)
http://bioinfo.cnio.es    http://bioinfo.cnio.es/people/eandres
===================================================

	[[alternative HTML version deleted]]