[BioC] illumina --> limma?

[Mark Dunning]

Hi Wolfgang,

Thanks for the advice about the logarithm transformation. I think the data
in question here is the output from the Illumina software, in which case it
will have already been background corrected at the individual bead level and
summarised in a single value for each gene. Can the vsn transformation still
be safely applied to this data?

btw I would recommend using only the non-normalised output from the Illumina
software if you plan to analyse on the log2 scale. All of the Illumina
normalisation methods use a "background normalisation" step first which
seems to create a lot of negative values. Illumina like to analyse all their
data on the un-logged scale despite a very obvious relationship between the
mean and variance. In fact, they take this relationship into account in
their models for differential expression. If anyone is interested, the
following paper discusses their methods:

A Model of Technical Variation of Microarray Signals - Chudin et al. Journal
of Computational Biology, May 2006

Regards,

Mark

-----Original Message-----
From: bioconductor-bounces at stat.math.ethz.ch
[mailto:bioconductor-bounces at stat.math.ethz.ch]On Behalf Of Wolfgang
Huber
Sent: 24 August 2006 22:30
To: rasche at molgen.mpg.de
Cc: Bioconductor
Subject: Re: [BioC] illumina --> limma?


Hi Axel,

> Maybe it is to mention that for some normalisations, the log-values can
> get negative (i.e. for expression values lower than 1). To proceed you
> can convert the negative values to the smallest positive value. So add
> this line:
>
> exprs(newExprSet)[ exprs(newExprSet) < 0 ] = min(exprs(newExprSet)[
> exprs(newExprSet) > 0 ])

But that is a really bad idea. There has been a lot of research into
this, and if you use a decent normalization method you will not have
this problem. Data transformation (e.g. log) and the preceding
background correction step have to be considered together, and depending
on your taste, you can either use a method that

1. biases the background correction in such a manner that the values
stay away well above zero, so that you then can safely apply the
logarithm, or

2. use an unbiased background correction, but then apply a
variance-stabilizing transformation function (glog) that is meaningful
even with negative values.

There are many parallels between the two approaches, the bottom line
results can be quite similar. The 'vsn' package offers software for doing 2.


  Best wishes
  Wolfgang

_______________________________________________
Bioconductor mailing list
Bioconductor at stat.math.ethz.ch
https://stat.ethz.ch/mailman/listinfo/bioconductor
Search the archives:
http://news.gmane.org/gmane.science.biology.informatics.conductor