[BioC] normexp (limma) failure?

[Gordon Smyth]

>Date: Wed, 30 Nov 2005 04:13:03 -0800 (PST)
>From: bregje wertheim <bregjewertheim at yahoo.com>
>Subject: [BioC] normexp (limma) failure?
>To: bioconductor at stat.math.ethz.ch
>
>   Dear all,
>
>since upgrading to the latest version of R (2.2.0) and the latest version 
>of limma (2.4.4), backgroundCorrect(method="normexp") seems to fail. I've 
>used this method before on a dataset of 3 dual-color chips without 
>problems, but now, exactly the same dataset gives result like this:
>
>   An object of class "RGList"
>$R
>                s1       s2         s3
>[1,] 1.046122e+53 4.402358 2.50177231
>[2,] 1.046122e+53 2.356062 4.36288331
>[3,] 1.046122e+53 3.226432 3.52955031
>[4,] 1.046122e+53 2.254210 4.41843931
>[5,] 1.046122e+53 2.217173 0.03891324
>697 more rows ...
>   $G
>                s1    s2           s3
>[1,] 3.412653e+59 1e-06 6.411355e+49
>[2,] 3.412653e+59 1e-06 6.411355e+49
>[3,] 3.412653e+59 1e-06 6.411355e+49
>[4,] 3.412653e+59 1e-06 6.411355e+49
>[5,] 3.412653e+59 1e-06 6.411355e+49
>697 more rows ...
>
>
>   For reference purposes, the same data would result in the following 
> RGlist under older versions of R and limma:
>
>   An object of class "RGList"
>$R
>               s1       s2          s3
>[1,] 3.040345497 4.389520 2.465760255
>[2,] 3.419975497 2.343224 4.326871255
>[3,] 4.364419497 3.213594 3.493538255
>[4,] 0.003308759 2.241372 4.382427255
>[5,] 4.003308497 2.204335 0.002797883
>697 more rows ...
>   $G
>               s1        s2        s3
>[1,]  9.48290819  4.130162  3.416828
>[2,]  0.01069039 11.593125 12.148309
>[3,] 12.21438919  3.056088  8.889050
>[4,]  1.03846319 13.435718 16.352013
>[5,]  6.11253719 13.028310 12.546458
>697 more rows ...
>
>   Could anyone advice me what is happening, and if I can do anything 
> about it?
>
>   Best regards,
>
>   Bregje Wertheim
>
>   Dept of Biology
>   University College London
>   bregjewertheim@ yahoo.com

Dear Bregje,

The normexp background correction method requires an interative fit of a 
model to the data. We have had trouble getting this numeric process to 
converge reliably for all possible data sets. It appears that the changes I 
made before the BioC 1.7 release have made the convergence problems worse 
rather than better, and this seems to have affected you.

Because of the difficulty of obtaining maximum likelihood estimates of the 
normexp model for all possible datasets, I am moving towards fitting the 
normexp model using the density-smoothing method used for Affymetrix data 
in the affy package. In limma 2.4.4 (available from CRAN or from the BioC 
developmental repository) you may specify instead

  backgroundCorrect(RG, method="rma")

and this will use the bg.adjust() function from affy. This is equivalent to 
the normexp method except for the estimation of the model parameters.

Best wishes
Gordon