[BioC] double violation of normalization assumptions?

[Matthew Hannah]

>I would never normalize within treatment, as the normalization could 
>cause the apparent differential expression.
>
>In the absence of any suitable between array method, I would use 
>MAS5.0.  Of course, the MM binding could also be affected if the 
>virus affects RNAs that bind to the MM probes, but this still seems 
>preferable to normalization within groups.
>
>--Naomi

I would agree that it would be a bad idea to normalise within
treatments, however, I also do not see that MAS5 normalisation would
necessarily be any better as the primary problem is that the amount of
mRNA is changing and the median scaling to a target intensity as
performed by MAS5 on the expression estimates will lead to an equalising
of the up and down regulation in a similar way as RMA. However, it would
not change the distribution as much as a quantile based method obviously
would but on the down-side may have more noise.
 
The real problem here is knowing how much the mRNA content is affected
by the treatment. Maybe this could be assayed independently, but note
that this it is not the same as measuring total RNA and could
additionally be complicated by the presence of the viral RNA in high
amounts. I would suggest 2 possibilities, one experimental and one
bioinformatic, both far from ideal but better than ANY normalisation
that assumes equal mRNA.
1. You could try to find the 'real' expression of a number of genes
throughout the distribution of all genes. This could be assayed by Q-PCR
(obviously with high replication), however you will hit a problem of how
to normalise your results, is the standard housekeeping method ok -
probably not, and so then you need to use additional spike-ins, but what
to spike to? Total RNA has the problem discussed above, and you come
back to the problem of measuring mRNA per cell.
 
2. Select a set of 'house-keeping' genes on the array and assume that
these are present at a constant amount per cell. You could either select
these as accepted literature ones eg: actin, UBQ etc... or find publicly
available data from the same cells treated in ways that are unlikely to
affect total mRNA and select genes with the lowest variability and then
make the assumption they should also be similarly expressed between your
treatments and normalise to them. The drawback is that are there really
HK genes or genes that are constant across treatments? The other major
factor here is how the cell is highjacked and if this will affect HK
gene expression (I'm sure it will but maybe different genes will have
different likelyhood.
 
As for getting past review, then any method will probably have trouble
with a good reviewer, but unfortunately the amount of biologists that
know that microarray expression values can be massively affected if
there is a change in total mRNA is limited.

If you want to understand (and worry) more about the problem with this
kind of data then these papers by Frank C.P. Holstege are essential
reading.
EMBO reports 4, 4, 387-393 (2003)
Monitoring global messenger RNA changes in externally controlled
microarray experiments

EMBO reports 5, 10, 964-969 (2004)
In control: systematic assessment of microarray performance

As a final suggestion and the only positive - the rank of the genes
expression is not going to be affected and so the most up-regulated
genes will be the most up-regulated (or least down-regulated!!) and vice
versa, so if the biggest changes are what are of interest then maybe
some information can be salvaged. If you want the global picture then
the current data are unlikely to be sufficient.

HTH,
Matt

At 11:26 AM 1/20/2006, Jenny Drnevich wrote:
>Hello all,
>
>I'm analyzing a set of data that turns out to be a little unusual, but
>related to the recent discussions on what to do if you have a large
(>40%)
>proportion of genes changing .  I'd like some advice on my approach,
>particularly from the point of view of a manuscript reviewer...
>
>Here's the scenario: I get a set of 6 affymetrix chips to analyze, 2
>treatments, 3 independent reps each. The QC on the chips is
outstanding,
>the distributions of intensities within each set of reps are very
similar,
>but the "Inf" treatment has slightly lower expression values overall
than
>the "Non" treatment, based on boxplot() and hist(). I use GCRMA for
>preprocessing, and limma functions for the two-group comparison.
Results:
>about half of the genes are differentially expressed at FDR=0.05, and
twice
>as many are downregulated as upregulated. I am now worried about the
>normalization, because quantile normalization (and just about every
other
>normalization method) assumes that only a small proportion of genes
(~20% -
>40% at most) are changing. So I ask the researcher if she would expect
a
>large number of genes to be changing, and if most of them would be
>decreasing, and she says "yes, of course". Turns out her treatments on
the
>cell line are mock-infected (control) and infected with a virus that
takes
>over the cell completely to produce viral RNA and eventually kills the
>cell. The infected treatment was harvested right when the first cells
>started dying, so there should be broad-scale down-regulation of host
mRNAs
>due to infection. This corresponds to the lower overall intensities in
the
>"Inf" group; extraction efficiencies were equivalent for all the
samples,
>and equal volumes of labeled RNA were hybridized to each chip, so I
assume
>the remainder of the RNA in the "Inf" samples was viral. The viral RNA
did
>not appear to have much effect on non-specific binding because MM
>distributions were extremely similar across all arrays, although again
>slightly lower for "Inf" replicates.
>
>What is the best way to normalize these data? Suggestions in the
>Bioconductor Archives for dealing with disparate groups mostly involved
>samples from different tissue types, and the consensus seemed to be to
>normalize within each group separately. However, there were cautions
that
>the values across tissue types may not be comparable, and that scaling
each
>array to the same mean/median intensity might be a good solution.
However,
>in this case I don't think scaling is appropriate because there is
reason
>to believe that the mean/median intensity is not the same between the
>treatments. I remember a paper discussing normalization assumptions
that
>mentioned a case where programmed cell death was being assayed, and so
most
>transcripts were going way down. However, I can't remember what they
>advised to do in this case, nor which paper it was - anyone know?
>
>This situation also turns out to be very similar to the spike-in
experiment
>of Choe et al. (Genome Biology 2005, 6:R16) where they spiked in ~2500
RNA
>species at the same concentration for two groups(C and S), and another
>~1300 RNA species at various concentrations, all higher in the S group;
to
>make up for the difference in overall RNA concentration, they added an
>appropriate amount of unlabeled ploy(C) RNA to the C group. So in
total,
>~3800 RNA species were present of the ~14,000 probe sets on the Affy
>DrosGenome1 chip. Even though less than 10% of all the probe sets were
>changed, because they were all "up-regulated", the typical
normalization
>routines resulted in apparent "down-regulation" of many probe sets that
>were spiked-in at the same level. Their solution was to normalize to
the
>probe sets corresponding to the RNAs not changed, so they could
evaluate
>variants of other pre-processing steps and analysis methods. Obviously,
we
>cannot do this. There are only 4 external spike in controls, so I am
>hesitant to normalize to them as well.
>
>Here is what I propose to do to account for both a large proportion of
>genes changing, and most of them changing in one direction, along with
>justification that I hope is acceptable:
>
>Background correction was performed based on GC content of the probes
(Wu
>et al. 2004). Because infection is expected to cause a large proportion
of
>genes to change, normalization across all arrays could not be performed
>because most normalization methods assume that only a small fraction of
>genes are changing (refs). Instead, quantile normalization was
performed
>separately for treatment group, as has been suggested for disparate
samples
>such as different tissue types. Additionally, the amount of host RNA in
the
>infected cells is expected to decrease, so both sets of arrays were not
>scaled to the same median but instead were left alone; in this
experiment,
>the extremely high correlation and consistency of arrays values
suggests
>that the arrays can be directly compared.
>
>
>What do you think? Would this past muster with you if you were the
reviewer?
>
>Thanks,
>Jenny
>
>
>
>Jenny Drnevich, Ph.D.
>
>Functional Genomics Bioinformatics Specialist
>W.M. Keck Center for Comparative and Functional Genomics
>Roy J. Carver Biotechnology Center
>University of Illinois, Urbana-Champaign
>
>330 ERML
>1201 W. Gregory Dr.
>Urbana, IL 61801
>USA
>
>ph: 217-244-7355
>fax: 217-265-5066
>e-mail: drnevich at uiuc.edu
<https://stat.ethz.ch/mailman/listinfo/bioconductor> 
>