[BioC] negative expression values
michael watson (IAH-C)
michael.watson at bbsrc.ac.uk
Thu Jun 10 12:19:53 CEST 2004
I quite often use the "set negative to one" approach, however, for the
record I must say there is a problem with this, as setting one value of
a ratio to 1 can result in very skewed ratios at low intensity levels
e.g. if Cy5=24 and Cy3=1, then my ratio is 24, which suggests a SERIOUS
amount of upregulation, when in effect the gene is probably "switched
off" in both channels.
It is important to be aware of this problem
From: Arne.Muller at aventis.com [mailto:Arne.Muller at aventis.com]
Sent: 10 June 2004 11:15
To: michael watson (IAH-C); bioconductor at stat.math.ethz.ch
Subject: RE: [BioC] negative expression values
thanks for your reply, please see my comments below.
Arne Muller, Ph.D.
Toxicogenomics, Aventis Pharma
arne dot muller domain=aventis com
> -----Original Message-----
> From: michael watson (IAH-C) [mailto:michael.watson at bbsrc.ac.uk]
> Sent: 10 June 2004 11:46
> To: Muller, Arne PH/FR; bioconductor at stat.math.ethz.ch
> Subject: RE: [BioC] negative expression values
> Hi Arne
> There are several approaches:
> 1) create a positive lower threshold below which you do not believe
> anything is being expressed. Some people use 1, others use 50, but
> essentially you do not "shift" your distribtion, you simply re-set all
> values below a threshold to that threshold
I think I'll go for this one, and choose a cutoff of 1.
> 2) perform a linear shift of the distribution - essentially, one takes
> the lowest value and adds this to every value in the distribution.
This was actually meant with "shift" intensities.
> 3) perform a slightly more intelligent shift - I believe
> there are some
> functions in the R library vsn, but I have never used them. These
> result in an avoidance of zero or negative values (I think)
> 4) perform kooperberg background correction, which is available in the
> limma package. This is a published method which estimates low
> expression values from negative numbers using some kind of Bayesian
> magic ;-)
3 and 4 are probably approrpiate, but I'd like to stay as close to the
original Resolver data as possible (and Resolver already performs a
de-trending based on the variance, i.e. something like vsn does). One
purpose of my analysis is to reproduce some anova results from Resolver.
> -----Original Message-----
> From: Arne.Muller at aventis.com [mailto:Arne.Muller at aventis.com]
> Sent: 10 June 2004 10:25
> To: bioconductor at stat.math.ethz.ch
> Subject: [BioC] negative expression values
> I'm analysing a set of expression data that were processed
> in Rosetta Resolver version 4. I've exported the data into R to run
> linear model + anova with 3 factors.
> In Resolver expression values can be zero or negative - about
> 4 percent
> of the probesets in my data set.
> I'd like to work with log2 transformed intensities, so I was wondering
> if there's anything speaking against shifting intensities, so that the
> minimal intensity is e.g. 1. The actual minimmal value is -106 and the
> maximum is about 14000.
> I'm happy to receive suggestions or comments,
> Arne Muller, Ph.D.
> Toxicogenomics, Aventis Pharma
> arne dot muller domain=aventis com
> Bioconductor mailing list
> Bioconductor at stat.math.ethz.ch
More information about the Bioconductor