[BioC] Limma lmFit function and spot quality weights

[Benoit]

Dear Gordon,

Thank you for your answer.
I understand the interest to make comparisons even with one sample per 
group, it could be informative but in my analysis I try to identify 
robust differential expressed genes.
Concerning the weights, I am a very new user in R and Limma and it's not 
easy to generate appropriate weights. Curently, I'm using 0 and 1 flags. 
I know that it's possible to use intermediate weight values but I don't 
know on which probe allocate these values.
Another question is, is it judicious to filter before or after fitting 
linear model ?
On affy data, I apply filtering on the PMA call before analysing with 
limma and keep probe with only a minimal number of "P" calls per group. 
Do you think it is pertinent or is it a bad method ?

Thanks,

Benoit

Gordon K Smyth a écrit :
> Dear Benoit,
>
> It doesn't seem to me to be desirable to place restrictions on the 
> weights that people can specify to lmFit. In some cases it is 
> desirable to be able to make comparisons for probes with only one 
> available sample per group.
>
> On the other hand, this does means that you are responsible for the 
> weights you create, and you may get poor results if you input weights 
> that are innappropriate for the data.
>
> Best wishes
> Gordon
>
>> Date: Mon, 27 Jul 2009 13:41:06 +0200
>> From: Benoit <benoit.loup at jouy.inra.fr>
>> Subject: [BioC] Limma lmFit function and spot quality weights
>> To: bioconductor at stat.math.ethz.ch
>>
>> Hello,
>> I'm using Limma to assess differential expression on double colour
>> microarray data and have a question about the lmFit function.
>> When I fit linear model using lmFit, as I understood, the function uses
>> the weights extracted from the MA object when present and/or specified.
>> Thus, I tried fitting with and without the spot quality weights and I
>> found different results (not very surprising in fact).
>> In fact, when I used weights, zero weighted spots seemed to be removed
>> from the analysis and it's here that I have a problem.
>>
>> For my experiment, I compare two groups (control vs treated) in a
>> classical design experiment "Two Groups: Common Reference" as describe
>> in the Limma documentation.
>>
>> design=modelMatrix(targets,ref="ref")
>> design
>> fit=lmFit(MA,design,weights=MA$weights)
>> /alternative without weights : fit=lmFit(MA,design,weights=NULL)/
>> cont.matrix=makeContrasts(pollutedVScontrol=polluted-control,polluted,control,levels=design) 
>>
>> cont.matrix
>> fit2=contrasts.fit(fit,cont.matrix)
>> fit2=eBayes(fit2)
>> res=toptable(coef=1,number=15744,fit=fit2,genelist=fit2$genes,adjust.method="BH",A=fit2$Amean,eb=fit2,p.value=0.01) 
>>
>>
>> The difference between the analysis with and without weights is that
>> when I use weights new genes highly differentially expressed appeared.
>> When I control these genes, in fact they correspond to spots that are
>> flagged (0) on the majority of the arrays (i.e. only one weight at 1 for
>> the control and one weight at 1 for the treated). Thus for these genes
>> the comparison is performed only one "control array" versus one "treated
>> array".
>> So is it possible to specify to lmFit that there must be a minimum of
>> "1" weights or a maximum "0" weights per groups of array ?
>>
>> Thank you for any help you can bring me.
>>
>> Benoit
>>
>> -- 
>> Benoit Loup, PhD
>> UMR Biologie du D?veloppement et Reproduction
>> Diff?renciation des Gonades et Perturbations
>> INRA ? Domaine de Vilvert
>> B?timent Jacques Poly
>> 78350 Jouy en Josas
>> France
>>
>> Tel: 33 1 34 65 25 38
>> Fax: 33 1 34 65 22 41
>> E-mail: benoit.loup at jouy.inra.fr
>
>

-- 
Benoit Loup, PhD
UMR Biologie du Développement et Reproduction
Différenciation des Gonades et Perturbations
INRA – Domaine de Vilvert
Bâtiment Jacques Poly
78350 Jouy en Josas
France

Tel: 33 1 34 65 25 38
Fax: 33 1 34 65 22 41
E-mail: benoit.loup at jouy.inra.fr