[BioC] Limma or something else ? plus Normalization question

[Gordon Smyth]

At 02:47 AM 29/09/2003, Phguardiol at aol.com wrote:
>Hi,
>let say I have 9 chips (Affy U133A)
>A1, A2, A3  triplicates = group A
>B1, B2, B3  triplicates = group B
>C1, C2, C3  triplicates = group C
>
>I d like to know what are the genes differentially expressed between A & 
>B, A & C, B & C
>I see two options (1) comparing 2 by 2 these groups using for instance LPE 
>or another test like this one,
>  option (2) using Limma.

LPT is the "local pooled error test" proposed in a papery by Jain et al, 
http://hesweb1.med.virginia.edu/bioinformatics/research/index.html, which 
is apparently to appear in Bioinformatics.

Limma and LPT address the same problem but from different points of view. 
Scanning the paper by Jain et al, they don't compare their method with any 
of its natural competitors (a common complaint with the Bioinformatics 
literature at the moment), so there are no objective grounds for choosing 
between the different methods which moderate, pool or smooth the variances. 
The LPT uses intensity as a predictor of variability and, given intensity, 
doesn't seem to allow genes to have individual variances. One might guess 
that LPT would do well if you are using a method like MAS for affy data or 
local background correction for cDNA data which produces a strong 
relationship between intensity and variability, and if the genes are not 
otherwise very different re variability. One might guess that limma would 
do better if you use BioC style normalization for affy or smoothed 
background for cDNA or if the genes are greatly different re variability 
not related to intensity.

One can't really know though without tests on good data where some form of 
truth is available.

>What would be the best option ? I like using LPE since it is well designed 
>for low replicates. If I  use option (1) should I normalize all the chips 
>during the same process or first normalize group A with group B and run 
>the comparison..etc ? My concern here is that I will use a relatively low 
>number of chips with RMA / Quantile normalization, in the other way 
>(normalizing everything at the same time) I could introduce "a bias" that 
>is not needed.

Don't really understand your question here. It's almost always best to 
normalize your data all at once, and the same for the analysis. Is the 
problem that LPT is applicable only to pairwise comparisons?

Gordon

>Any comments is apppreciated as usual.Thanks
>Philippe


Gordon