[BioC] Normalisation method worries

[J.delasHeras at ed.ac.uk]

Hi,

here I come again about normalisation... :)

I have a set of 4 cDNA 2-colour arrays, including dye-swap.
00 & 01 are in one direction, 02 and 03 in the opposite.

Each hybridisation corresponds to untreated sample vs. transfected 
sample (experiments on cell lines). The treatment appears to affect 
MANY genes.

I plotted the raw data log2(ratio) vs log2(product)... and this is what i get:

http://mcnach.com/MISC/raw_RI_plots_M1.png

No background correction was applied.

The print-tip loess curves are shown here for the first slide (00):

http://mcnach.com/MISC/print-tip_loess_plot_for_slide_140000_M1.png

By the nature of this experiment (transfection of a fusion gene, made 
up of a DNA-binding section with unknown specificity, but we expect 
it'll pick up a lot of genes, and a transactivator domain) we are not 
surprised to find many genes are activated. These are mostly the ones I 
want, not so much the ones that have enhanced expression, but the ones 
that go from not detectable expression to detectable levels.

The plots show there is a strong dependence between log2(ratio) and 
intensity, mostly on the left... the higher intensity spots show a less 
strong dependency. This is also seen nicely when looking at the loess 
curves.
Now, I feel that a lot of the genes I might be interested in will 
probably be in teh first half, since I'm looking at genes where the 
signal before transfection is just about background... and it's this 
region where the curves are steeper.
If I apply loess, I'll flatten the whole thing.
In fact, it looks like this:

http://mcnach.com/MISC/MAplots2.png

this image is from a different experiment (loess normalised, and 
processed with limma), but the plot looks pretty much the same, and you 
get that higher density diagonal (around 45 degrees) where I find that 
these genes that are activated by the treatment tend to cluster.

So... on the one hand, I feel I may not be doing the best sort of 
normalisation, yet I do get meaningful results that verify okay by RT, 
so it's not all too bad. And I have no idea what other method would be 
better suited anyway... I have no set of known invariant control genes 
I could use, which I guess would be the best... although I might be 
able to figure something out in the future.
What is the feeling of more experienced people here? Have you worked 
with experiments like this? Am I worrying too much just to try to get a 
few % more reliable data?

Any comments from anyone greatly appreciated!

Jose

-- 
Dr. Jose I. de las Heras                      Email: J.delasHeras at ed.ac.uk
The Wellcome Trust Centre for Cell Biology    Phone: +44 (0)131 6513374
Institute for Cell & Molecular Biology        Fax:   +44 (0)131 6507360
Swann Building, Mayfield Road
University of Edinburgh
Edinburgh EH9 3JR
UK