[BioC] remove probes - GCRMA 1.1.3 weird behavior

[Ariel Chernomoretz]

Dear list,

1) I wrote a couple of functions  to remove a list of probes and/or 
probesets from
a given CDF and Probe environments (I did not find that anywhere). The 
idea is to be able to remove
  'offending' probes from rma and gcrma calculations. These functions 
are included in the attached file
'removeProbes.R',  and an example is included in 'removeProbesEx.R'.
The code is there to anyone that want to give it a try.

2) The function 'RemoveProbes' modifies the specified CDF and PROBE 
environments
      in a consistent way.  IMHO this should be  gcrma-compliant (i am 
using gcrma 1.1.3) .
However I found a strange behavior I would like to discuss. If just a 
single probe is removed (e.g '1000_at1' from
Dilution dataset, see removeProbesEx.R) then only the RMA expression 
value of the modified probeset changes  when compared with the 
full-probe result  :-)  . The same is true for GCRMA 1.1.0.
However, the expression values  of many different probe-sets (and not 
only the modified one) calculated by  GCRMA 1.1.3 change (see figure 
attached where expression values calculated with and without 1000_at1 
are plot for both, RMA and GCRMA ).

After some analysis I think  the problem is in the function 
'bg.parameters.ns' of gcrma 1.1.3.
The sampling of 5000 affinity values performed in order to fit the 
MM-affinities dependency produce slightly different results when the 
number of sampled affinities changes  (even in one unit!). This results 
in different loess fitting curves, and different mean and sigma 
estimates for PM nonspecific signals, etc, etc.
On the contrary the method to estimate bg parameters implemented in 
version 1.1.0  (based on the analysis of PM distribution, group by 
affinities) seems to be far most robust. ( of course I am talking here 
of implementation, not on the different working hypothesis) .
I would like to here opinions regarding how this behavior affect gcrma 
estimates?
Should one for instance conclude that this instability precludes 
accuracy in version 1.1.3?

Regards,


-------------- next part --------------
A non-text attachment was scrubbed...
Name: oneout.png
Type: image/png
Size: 5236 bytes
Desc: not available
Url : https://stat.ethz.ch/pipermail/bioconductor/attachments/20050420/2a5bc8c1/oneout.png
-------------- next part --------------


Ariel Chernomoretz, Ph.D.
Centre de recherche du CHUL
2705 Blv Laurier, bloc T-367
Sainte-Foy, Qc
G1V 4G2
(418)-525-4444 ext 46339