On Jun 30, 2008, at 11:01 AM, Sean Davis wrote: > On Mon, Jun 30, 2008 at 1:54 PM, Kasper Daniel Hansen > <khansen at="" stat.berkeley.edu=""> wrote: >> Hi >> >> I have gotten my hands on data from the single color Agilent >> platform using >> a custom array design and I would like to hear what people are >> usually doing >> when it comes to preprocessing. >> >> I have previously analyzed some two color arrays from Agilent and >> found that >> the data I had was pretty standard when it comes to normalization. >> Even >> though I preferred doing my own preprocessing the Agilent supplied >> gProcessedSignal and rProcessedSignal columns were decent (this was >> from a >> much earlier version of their software - Feature Extractor). >> >> But for the one color arrays I find that gProcessedSignal performs >> horrible >> - flat out horrible, the raw data looks much better. Furthermore, >> when I >> normalize between I arrays I see relatively little effect of >> normalization, >> sometimes the normalization even increases the spread on MA plots >> where I >> would not expect it to do anything. Of course this may be related >> to the >> hybridizations done or the array design I have in hand, but I still >> find it >> somewhat surprising. >> >> I have tried vsn2 from vsn, quantile normalization and quantile >> normalization following normexp (offset 25 and 50) background >> correction >> from Limma. All 3 (4 if you count the 2 offsets) combinations have >> also been >> done with and without subtracting the local background estimate >> from Feature >> Extractor (the gBGMeanSignal column). >> >> Anyway, I am curious as to what other people's experience using this >> platform are. > > What type of array is it? In particular, is it miRNA? No, it is a custom splice junction design using (I believe) a standard mRNA protocol (expect that we are hybing at 70 degrees instead of 65 degrees based on some assessment). We are doing a pilot study so we do not have too much experience with this platform. But the raw data looks very nice and interpretable - it is more the fact that normalization seems to have little effect (we can always argue about how much I want to normalize - but that is not really my concern here) coupled with the fact that the processed signal looks crappy based on comparing two replicate arrays. I am not really saying that the platform sucks, in fact one could interpret the fact that normalization have little effect to mean that the raw data is super good. I am just wondering what other peoples experience is. The only normalization that really seems to have an (big) effect is normexp with an offset of 50, which certainly shrinks the M values towards zero. Kasper