[BioC] Question about normalization of microarray data

[Sean Davis]

On Nov 18, 2004, at 6:42 AM, STKH ((Steen Krogsgaard)) wrote:

> Hi Johan,
>
> since you don't seem to have a suitable common reference, how about 
> using a balanced block design instead?
>
> cheers
> Steen
>

Yep.  The true power of the two-color design is in comparing within 
slide two samples of interest.  If one is going to use a common 
reference, then the common reference should probably have SOME 
semblance to the test sample in terms of gene expression.  Even when 
using a commercially available reference, one presumably has some 
variation in expression that mimics that in the test sample better than 
using a non-biologic reference like PCR products.  It will be 
interesting to see what you end up doing here, but I do agree that sing 
within array contrasts whenever possible is a good idea.

Sean

> -----Original Message-----
> From: bioconductor-bounces at stat.math.ethz.ch 
> [mailto:bioconductor-bounces at stat.math.ethz.ch] On Behalf Of Johan 
> Lindberg
> Sent: 18. november 2004 09:15
> To: bioconductor at stat.math.ethz.ch
> Subject: [BioC] Question about normalization of microarray data
>
>
> Hi all. I have a question about normalization of microarray data.
>
> In our lab we use in-house spotted cDNA arrays. We have so far used 
> commercial reference when doing reference design. We are now trying a 
> new approach but we have problems with normalization. What we have 
> done is to pool product from every spot on the chip and done in vitro 
> transcription on the PCR product. So we have RNA corresponding to 
> every spot on the chip. Then this is used as a reference. It is much 
> cheaper and we get signal from every spot on the chip instead of 
> having spots with no signal in both channels. But when one looks at an 
> MA-plot the plot will be skewed towards the reference. There are about 
> (in this pilot case) 2000 spots that only give signal in the reference 
> channel (which will skew the MA-plot). This will make many assumptions 
> not correct when normalizing the data, e.g. using lowess normalization 
> assuming that the ratio R/G should be 1 for most spots.
> Since the case for this kind of data is that one channel should be 
> much stronger than the other, and we want to keep the normalization 
> within slide (to be able to correct for spatial biases and intensity 
> dependent) the only way I could think of is by spotting a lot of 
> control spots (not present in the tested RNA or the reference RNA) and 
> use these to normalize the data.