[BioC] Three questions about processing Agilent Arrays

[Gordon K Smyth]

Dear David,

Since no one else has responded, I've interpolated some brief responses 
below.

>[BioC] Three questions about processing Agilent Arrays
>David Garfield dag23 at duke.edu
>Mon Mar 2 21:32:51 CET 2009
>
>Hi all,
>
>I am new to micro-array processing and to Bioconductor.  The 
>documentation provided by the authors of packages like limma and marray 
>have made it very easy to get a toehold on how to pre-process, normalize, 
>and analyze microarrays.
>
>The Agilent Feature Extraction Software Manuals, alas, are not as 
>helpful, and I would be very appreciative of some guidance on two issues.
>
>First, The Data: I have inherited an Agilent two-color microarray 
>dataset.  This dataset consists of eight microarrays.  To each array was 
>hybridized both a common reference sample and a tissue specific sample. 
>In 6 of the arrays, the reference is Cy3.  In the other 2, the reference 
>is Cy5 labeled.  Included in our probeset are a set of Agilent designed 
>spike-in and negative controls, along with ~600 probes we have designed 
>that we hope act as negative controls for this species (sea urchin).
>
>The questions:
>
>1) Loess normalization when many genes are expected to be differentially 
>expressed: The first pre-processing step I plan to take is a within array 
>global normalization using loess.  However, I am concerned that because 
>the data consist of comparisons between different tissues, even with a 
>common reference, there may be a large number of genes that really are 
>differentially expressed. Can anyone provide insight as to the limits of 
>loess normalization in the face of an expectation that many genes will be 
>differentially regulated? Can anyone suggest alternatives for array 
>normalization based on our experiment?

See Oshlack et al, Genome Biology, 2007.

I doubt you can do much better than loess though.

>2) Spatial Detrending of the Background Signals: Agilent's Feature 
>Extraction manual discusses something called spatial de-trending. The 
>goal of the algorithm is to apply an offset to the spot intensities that 
>reflect spatial specific variation on the array. Unlike other 
>normalization modules I've seen in bioconductor, this algorithm is 
>applied only to the negative controls or low intensity spots and is 
>applied to each channel separately (though they still refer to the 
>process as using 2D loess normalization).  I have not seen similar steps 
>references in the literature.  Can anyone speak to the importance of this 
>step or bioconductor packages that facilitate doing this.  I'm not 
>against programming my own, if this step is needed, by why re-invent the 
>wheel.

Not sure what you're asking for here.  This is likely to be a proprietry 
Agilent algorithm.  There aren't any Bioconductor packages which do it, 
i.e., we find we can live without out.  But if you want it, it would seem 
that Agilent already provides it for you.

>3) The best ways of incorporating our dye-swaps:  The most commonly 
>referenced between array normalization strategy I have seen is 
>quantile-normalization.

No, this is seldom used for two-colour data.

>However, given our spoke-and-hub sort of design 
>(with dye-swaps), it seems that this method would lose some potential 
>information from the common reference and the dye-swaps. Can one suggest 
>a (ideally already implemented) normalization strategy in this case?

I can't see any reason for not using loess as usual.

>I've asked many questions.  If you could provide any insight into any of 
>them, that would be appreciated.  Hopefully I will be able to contribute 
>information from this experience to the list at a future date.
>
>Cheers,
>David

I don't see any reason why a standard two-colour analysis, e.g., limma 
package style, would not be suitable for your data.  Jump in and give it a 
go.

Best wishes
Gordon