[BioC] duplicateCorrelation function and custom array design

[Gordon Smyth]

>Bela Tiwari btiwari at ceh.ac.uk
>Fri Oct 29 16:45:17 CEST 2004

...

>Hmmm, as I write this, I just realised that I could have just done this
>all on the command line like:
>
>dupcorr <- mydupcorr(myMAList[indices,], design = design1, ndups = 2,
>spacing = 520)

There is no need to make your own mydupcorr() function. Just

   dupcorr <- duplicateCorrelation(myMAList[indices,], design = design1, 
ndups = 2, spacing = 520)
   fit <- lmFit(myMAList[indices,], design = design1, ndups = 2, spacing = 
520, correlation=dupcorr$consensus)

is the way to do it. All you need is that

  myMAList[indices, ]

is a valid MAList object with genuinely regularly spaced duplicates. For 
example, if your blocks contain 520 genes printed twice, followed by a 
number of unwanted spots, then you could use

pr <- printorder(myMAList$printer)
indices <- pr$printerorder <= 2*520

and proceed from there.

Do check that the estimated correlation is reasonably large though -- with 
dups at this spacing you'd expect to get a correlation around 0.6 or 
higher. If you don't get a correlation which is at least 0.3 say, then 
something may be wrong.

>but my essential question remains the same - is this sensible?

>My second question is due to my lack of experience with the functions
>involved - if I try to use the correlation consensus value generated via
>the above function as input into the lmFit function, will it matter if I
>include only the MAList elements for my genes and species-specific
>controls?  I.e. does it matter if I give   myMAList[indices,]    as the
>object parameter to the lmFit function, rather than the whole MAList
>object. I don't think lmFit needs to refer back to the array layout as
>stored within  myMAList$printer, but I'm not well versed enough to know
>if there are downsteam effects of entering only a subset of the MAList
>object to lmFit or not.

No, there's no problem. Removing blanks and unwanted negative control spots 
should actually improve the process.

Best
Gordon