[BioC] GCRMA/RMA bimodal distribution

[Matthew Hannah]

Hi,

Sorry for including the developers, but I guess you are the only ones
that will be able to answer this, (and I'm not sure BioC accepts .docs).
I saw a comment from Jean addressing the same question but couldn't find
the reply he referred to. 

https://www.stat.math.ethz.ch/pipermail/bioconductor/2004-August/005769.
html

It seems the mouse chip exprs values have a double peak after gcrma
(looking at a density plot).

As I'd received no response I've been doing some investigating (see
attached). Basically gcrma gives a single peaked distribution only for
U95 human chips (optimised with these?). Double peaks for exprs
estimates appear in the following - U133A(least) - Drosgenome1 - ATH1
(worst).

To a lesser extent this also occurs with RMA. U133A has a single wide
peak, and then they get worse in the order Dros1 - U95 - ATH1 (The last
two have obvious double peaks).

>From what has been said this is likely to be a problem of BG correction.
I don't know if there are opportunities to change this for RMA, but in
GCRMA there are tuning factors and I don't know if the ad-hoc estimate
(rather than full model) is causing this to happen. Turning of optical
correct had no effect.

I wanted to play about with GCRMA to see if the distribution changed
with the tuning factors but currently I seem to have an error (see
below) with gcrma and justGCRMA not finding gcrma.bg.transformation, and
I'm not sure how k should be expressed. 

I know people should look more at their data but with the ease of
just(GC)RMA and RMAexpress I know a lot of people just computing
expression measures for different chip types without looking at density
of the returned expression. Clearly these people are going to be working
with data that may be skewed in some way.

I guess that each chip type will need its BG correction optimising for
RMA and GCRMA to allow for a better estimate of true expression levels
and changes. I really hope this can be fixed as RMA and GCRMA seem to be
really useful expression measures and it would be a shame to have to
find alternative methods just because they are not optimised for your
chip type.

Thanks in advance,
Matt

R devel 2.0, win2k
affy 1.5.2 (I know it's not the latest but getBioC is not working for me
at the moment)
gcrma 1.1.0
 <<Exprs_meas_comp.doc>> 

> esetgcrma_slow <- gcrma(raw,fast=FALSE)
Computing affinities.Done.
Adjusting for optical effect.........Done.
Adjusting for non-specific binding.Error in bg.adjust.fullmodel(pms[,
i], mms[, i], pm.affinities, mm.affinities,  : 
        couldn't find function "gcrma.bg.transformation"
> esetgcrma_slow <- justGCRMA(fast=FALSE)
Computing affinities..Done.
Adjusting for optical effect..........Done.
Adjusting for non-specific binding.Error in bg.adjust.fullmodel(pms[,
i], mms[, i], pm.affinities, mm.affinities,  : 
        couldn't find function "gcrma.bg.transformation"
> esetgcrma_k4 <- justGCRMA(k=4*fast+0.5*(1-fast))
Computing affinities..Done.
Adjusting for optical effect..........Done.
Adjusting for non-specific binding.Error in
gcrma.bg.transformation.fast(pms, bhat, var.y, k = k) : 
        Object "fast" not found













Hi,

This has been mentioned before in the context of rma and that it was an
artifact of BG correction.

http://files.protsuggest.org/biocond/html/5066.html

I was very suprised to see that gcrma also gave a very pronouned bimodal
distribution. When comparing samples, obviously the relative positions
of the 2 peaks may influence observed expression changes. Would such
peak shifts be more likely in divergent samples, and if anyone wants to
comment on those.... ;-)

This example is using 12 chips (biological reps). But I initially
noticed it using 3 and 6 chips in rma.

Hope attachment works.

Cheers,
Matt


-------------- next part --------------
A non-text attachment was scrubbed...
Name: gcrma_dist.png
Type: image/png
Size: 6633 bytes
Desc: gcrma_dist.png
Url :
https://stat.ethz.ch/pipermail/bioconductor/attachments/20040825/083e56a
5/gcrma_dist.png