[BioC] vsn vs quantiles, bimodal data

[Ghai, Rohit]

Dear all

I have a few questions on comparing vsn and quantile normalization. It would
be really great to have some thoughts from the mailing list.

To begin with, are there any special criteria on how to deal with expression
data that is somewhat bimodally distributed ? We have
single-channel codelink chips and nearly all show a bimodal 
distribution (logged histograms). There seems to be a peak in low intensity
region and another in high intensity region with a furrow in between.

How does the deviation from the normal distribution affect quantile
and vsn normalization ?

we have extracted the raw intensity expression values using codelink's own
software. we want to use bioconductor for normalization. 

we have 10 arrays in one experiment and have repeated the experiment
independently once again.

problem in experiment 1:
->normalization using VSN did not yield satisfactory results. there are
still marked differences between chips in experiment 1 even though
all replicates show excellent correlation (>0.99).
->normalization using quantiles yielded near perfect boxplots.

problem in comparing experiment 1 with experiment 2(repeat):
->expression values on the whole seem higher in experiment 2.
VSN normalization does not really do anything drastic as before. so
differences in intensities remain between experiment 1 and 2 and also within
experiment 1.
->Quantiles again yield excellent boxplots.

What can one conclude ? 

Does the problem really lie with the data or does quantile normalization
overfit the data ? What can be then thrown away in this case? The data  from
experiment 1 or the vsn normalization?

Are there other normalization methods that can be used with such data ?

any comments would be helpful.

regards
Rohit