[BioC] Illumina BeadChips and beadarray
Wolfgang Huber
huber at ebi.ac.uk
Tue Aug 26 11:12:56 CEST 2008
25/08/2008 18:21 Ina Hoeschele scripsit
> Matt and others,
> one additional question regarding Illumina BeadChip expression data.
> If one uses a background correction method that produces negative
> intensities, then what is the best course of action (apart from using
> normexp!) in your experience? For the joint analysis of multiple
> beadchips (each with 6 samples), if one would delete all genes that
> have a negative intensity in some sample, then there would soon be no
> data left. So set negative values to a small positive value (which)?
> Thanks, Ina
>
Dear Ina,
just replacing negative values by a made-up positive value is a bad
idea. For example, it will distort your statistical inference.
There are two main types of background correction methods:
(1) ones that are guaranteed to result in positive intensities; these
methods are necessarily biased, since there are some genes which are not
expressed. An example is normexp.
(2) ones that try to be unbiased, and then because of noise in the
estimator sometimes produce negative, and sometimes positive, values for
those genes that are not or only negligibly expressed.
The motivation for (1) is that one can directly proceed with taking the
log-transformation. With (2) one needs to spend a bit more thought on
what is the appropriate transformation, and one answer is the
glog-transformation, as provided in the vsn package; see also its
vignette, and
library("vsn"); citation("vsn")
vsn introduces a bias towards 0 for those (g)log-ratios in which
numerator and/or denominator are small.
Interestingly enough, the bottom-line results of the two methods are
often surprisingly similar - the biases are introduced in different
places, but (when appropriately tuned) have similar regularizing effects.
Best wishes
Wolfgang
------------------------------------------------------------------
Wolfgang Huber EBI/EMBL Cambridge UK http://www.ebi.ac.uk/huber
More information about the Bioconductor
mailing list