[BioC] dependence of log2-ratios on scanning sensitivity?

[Henrik Bengtsson]

Hi.

On Feb 3, 2008 10:31 AM, Kamila Naxerova <knaxerova at chip.org> wrote:
> Dear Wolfgang,
> thanks for your reply. To obtain the intra-array normalized values, I
> followed the simplest possible approach, i.e. I did the following:
>
> targets = readTargets("TargetExiqon.txt")
> RG = read.maimages(targets, source="genepix")
> MA = normalizeWithinArrays(RG,method="loess")
>
> Can you recommend some alternatives (vsn, of course, and anything else)?
> I have tried the normalization provided in the marray package, but that
> seems to be largely similar, only maNorm(raw,norm="loess") returns
> normalized data with many more missing values than
> normalizeWithinArrays, 4608 missing values vs. 23 for limma- I don't
> understand what the difference in loess implementation here is.
>
> I have noticed that there are relatively large differences in PMT gain
> across my chips. Could that be an explanation? I suppose if that is the
> case, I will not be able to fix the problem computationally?

Having different PMT settings will affect the scale of the pixel
intensities in the scanned images.  If the PMT is set too hight, pixel
values will get censored to the maximum value of the scanner,
typically 65535.  Since the spot/feature signal is an average of all
pixels, this "saturation/censoring" effect comes into play at a
slightly higher PMT at the feature level compared to the pixel level,
especially when a robust average is used.   If you set the PMT too
low, the signal might drown in the "background" noise.  However, from
a study on Axon and Agilent two-color scanners, we found that after
controlling for different scale factors (and offset), the exact PMT
setting is not that important, as long as it is not extremely low or
extremely high (which is rarely the case).  In other words, you should
not worry too much about your PMT settings.  For more details about
our study, see

H. Bengtsson, G. Jönsson and J. Vallon-Christersson, Calibration and
assessment of channel-specific biases in microarray data with extended
dynamical range, BMC Bioinformatics, 2004, 5:177.

That paper also show how to estimate the offset added by the scanner,
given that you scan more than one time at different PMTs.  If you
control for the scanner offset, the main effect of the PMT setting
goes a way and only a scale factor remains which is very is to
estimate and adjust for.

Curve-fit normalization, e.g. lowess, loess, smooth spline and so on,
will *not* correct for scanner offset or other types of offset (the
"background" that Wolfgang refers to).  Offset is the main cause of
observing curvature in log-ratio log-intensity plots.  I would like to
suggest to use a normalization method that explicitly tries to control
for differences in offset and scale.  The variance stabilizing
normalization (vsn) method is one and affine normalization
(aroma.light) is another.

Hope this helps

/Henrik

>
> Thanks a lot,
> Kamila
>
>
> Wolfgang Huber wrote:
> > Dear Kamila,
> >
> > I don't know what exactly your normalisation did to obtain your "real,
> > intra-array normalized probe log2-ratios", but it is a common
> > challenge of two-color arrays to do good background correction, and if
> > the background intensities are quite different between different
> > arrays and the background correction is not adequate, this might
> > indeed lead to the kind of problem you describe.
> >
> > In general it is good practice to not vary anything that might change
> > the distribution of the data (such as scanner settings, labeling
> > efficiency, amount of CIP'ed RNA) between arrays, because post hoc
> > data normalisation will at best imperfectly remove these variations.
> >
> > Best wishes
> >  Wolfgang
> >
> > ------------------------------------------------------------------
> > Wolfgang Huber  EBI/EMBL  Cambridge UK  http://www.ebi.ac.uk/huber
> >
> >
> >  Naxerova wrote:
> >> Hi all,
> >>
> >> I am analyzing a bunch of miRNA arrays (Exiqon, dual channel,
> >> Hy3/Hy5). I
> >> am confused about the following issue.... Please apologize potential
> >> naivete, I have practically no experience with two-color designs.
> >>
> >> The chips have so called "Hy3 landing" lights, i.e. Hy3-labeled capture
> >> probes that help position the array for scanning. The mean intensity of
> >> these landing lights is very different between arrays (I assume that
> >> means
> >> slides were scanned with different sensitivity).
> >>
> >> First I thought that I don't have to worry about variable brightness
> >> among
> >> the arrays - I hybridized the exact same reference to all of them. But
> >> then I computed the correlation of all "real", intra-array normalized
> >> probe log2-ratios with the Hy3 landing light brightness... and the
> >> distribution has a disconcerting peak around 0.5. Am I missing something
> >> obvious?
> >>
> >> Thanks a lot.
> >> Kamila
> >>
> >> _______________________________________________
> >> 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
> >
> >
> >
>
> _______________________________________________
> 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
>