[BioC] RMA hist

[Rafael A. Irizarry]

i dont expect the distribution of exprs(eset) to be normal. why should 
RNA expression log-scale levels in any tissue follow a normal 
distribution? some physical 
models (and SAGE data) suggest expression should follow a power law 
(very different from a normal distribution) and 
the results of exprs(eset) are sometimes in agreement with this.

what has a chance of being normal is the distribution of expression for 
the same gene across technical or biological reps. im not sure if the 
procedure you follow requires these assumptions to hold, but if they do 
you should 
look at significance assessments with caution as they are sometimes driven 
by these assumptions (especially with small amounts of data). permutation 
tests are an alternative when one cant make these assumptions althought 
they too can be misleading with small amounts of data.

i find looking at volcano plots helpful in these situations.

hope this helps,
rafael


 On Tue, 26 Aug 2003, Delphine Fleury 
wrote:

> Dear all,
> I analyse affymetrix data to compare transcripts between the wild-type
> and 3 mutants. After normalization using the RMA method, the histogram
> of distribution of exprs(eset) isn t normal (cf histesetDataRMA.ps). It
> seems there two distributions in each array hist. If I continue the
> analysis using lm.series, all genes are significant. I guess that it
> remains a variation source which is not normalized. Do you have any idea
> of what it could be and how I can remove this variation ?
> 
> The procedure I did:
> > Data<-ReadAffy()
> > eset<-rma(Data)
> > fit<-lm.series(exprs(eset), d1)
> > eb<-ebayes(fit)
> > d1
>              WT A B C
> Exp143hyb246   1   0   0   0
> Exp143hyb247   1   1   0   0
> Exp143hyb248   1   0   1   0
> Exp143hyb249   1   0   0   1
> Exp143hyb410   1   0   0   0
> Exp143hyb411   1   0   0   0
> Exp143hyb412   1   1   0   0
> Exp143hyb413   1   1   0   0
> Exp143hyb414   1   0   1   0
> Exp143hyb415   1   0   1   0
> Exp143hyb416   1   0   0   1
> Exp143hyb417   1   0   0   1
> > toptable(coeff=2, number=22810, genelist=gnames,
> fit=fit,eb=eb,adjust="fdr")
> 
> I tried also the MAS5 procedure to normalize but the histograms of
> exprs(eset) is more far from the gaussian and all genes are also
> signficant with the same linear model.
> 
> Thanks for your answers.
> Delphine
> 
> --
> ==================================================================
> Delphine Fleury
> DEPARTMENT OF PLANT SYSTEMS BIOLOGY           Fax:32 (0)9 3313809
> GHENT UNIVERSITY/VIB,    Technology Park 927, B-9052 Gent, Belgium
> Vlaams Interuniversitair Instituut voor Biotechnologie         VIB
> mailto:defle at psb.ugent.be                  http://www.psb.ugent.be
> ==================================================================
> 
>