Hi Mick,
You're right, the model described in Kooperberg et al (JCB, Vol 9 No
1,
2002) is supposed to avoid NA's, and if implemented as described in
the
paper it will. My implementation has a line which sets the adjusted
foreground values to missing if they are estimated to be off the scale
(ie larger than 2^16).
This is possible when the denominator of equation (2) in the
Kooperberg
paper is very small, which occurs when the GenePix background is much
larger than the foreground. This causes the model estimate to blow
up,
giving a value which I'd argue is as useful as a negative value (hence
the decision to set it to NA).
If you send me the full details for the spot in question (eg fg, bg,
fgSD, bgSD, and the number of foreground and background pixels), I can
work out the model estimate for the foreground for you (which may be
in
millions!)
The kooperberg() code has been modified to return an RGList object
instead a list in the next update of limma. It should be on the
website
in the next few days.
As for the hump-backed phenomena you mentioned in your earlier email,
the MA plot will look slightly different depending on the amount of
background you subtract. If you take off a bit too much, you'll see a
lot of fanning at low intensities, and not much of a RG bias. If you
don't subtract background, you tend to see a 'hump-backed' bias (ie
the
low intensity spots have slightly negative log-ratios). This means
that
the low intensity spots have higher signal in the green channel than
in
the red, which is due in part to the substrate (glass slides tend to
have an underlying 'greeness.') The plotDensities(RG,
log.transform=TRUE) function will give you a different view of this.
I
guess the kooperberg() function will tend to give results more like
the
'no background' case than the 'too much background' which would
explain
what you observe. The important point is that intensity based
normalization should remove such biases.
Best wishes,
Matt Ritchie
michael watson (IAH-C) wrote:
>Hi
>
>I have read in some data using the limma examples and everything is
>fine, I have used the kooperberg algorithm and everything works,
APART
>from one single spot where I have an NA for some reason:
>
>>RGmodel$R[2483,]
>>
>[1] 255 919 371 315
>
>>RGmodel$G[2483,]
>>
>[1] 152 929 116 NA
>
>Now, the spot in question has a Cy3 foreground mean of 2256 and a Cy3
>background median of 4782.
>
>OK, so it is a bad spot. But GenePix hasn't flagged it, perhaps
because
>background is only high in the Cy3 channel - whatever.
>
>Anyway, I didn't realise that Kooperberg could introduce NAs into the
>data set - I thought it was this type of problem that kooperberg was
>meant to address? (ie negative spots).
>
>Thanks in advance for your help
>
>Mick
>
[[alternative HTML version deleted]]
