[BioC] Re: L-BFGS-B needs finite values of fn error in vsn package

[Wolfgang Huber]

Hi Isaac,

vsn uses a maximum-likelihood method for the parameter estimation, and 
since the log-likelihood is not quadratic, it uses the numeric optimizer 
"L-BFGS-B". As far as I can see, the log-likelihood is concave, i.e. has 
only one maximum, for practically all data sets. But for odd data sets 
and with bad start values, the numeric optimizer can run off into 
nirwana before finding the maximum. That seems to happen in your case.

Can you try the following:

1. Look at scatterplots of the probe intensities between all (or at 
least many) pairs of chips, and see if there's anything odd:
for(i in 2:24)
   for (j in 1:(i-1)) {
     plot(exprs(data)[, c(i,j)], pch=".", main=paste(i,j))
     locator(n=1)
   }

2. Try to use different (non-default) start parameters, e.g.

pstart <- array(c(rep(0, 24), rep(1,24)), dim=c(1,24,2))
vsn.data <- expresso(data,
              normalize.method="vsn",
              normalize.param(list(pstart=pstart)),
              bg.correct=FALSE,
              pmcorrect.method="pmonly",
              summary.method="medianpolish")

3. For a better diagnosis, I would have to lok at your data. I will be 
off email for the next two weeks, but please let me know after that 
(July) - I would be interested in making vsn more robust e.g. by having 
a smarter guess at the start parameters.

Best wishes
  Wolfgang





Isaac Neuhaus wrote:
> I have installed R version 1.9 (64 bit compiled) and the Biocunductor 
> packages release 1.4 which include the vsn package 1.4.13 and I am 
> getting the following error:
> 
> data <- ReadAffy(filenames=cels, sampleNames=names, verbose=F)
>  > vsn.data <- expresso(data, normalize.method="vsn", bg.correct=F, 
> pmcorrect.method="pmonly", summary.method="medianpolish")
> normalization: vsn
> PM/MM correction : pmonly
> expression values: medianpolish
> normalizing...vsn: 20000 x 24 matrix (1 stratum). Please wait for 4 
> dots: Error: L-BFGS-B needs finite values of fn
> Execution halted
> 
> Isaac
> 


-- 
-------------------------------------
Wolfgang Huber
Division of Molecular Genome Analysis
German Cancer Research Center
Heidelberg, Germany
Phone: +49 6221 424709
Fax:   +49 6221 42524709
Http:  www.dkfz.de/abt0840/whuber