[BioC] how deal with multiplicate affy probes?

[Jeff Gentry]

Laurent was having trouble mailing this, so I forwarded it for him:
--------------------------------------------------------------

Jonathan,

You bring up several issues (which I see as almost separated ones). 

One is the probe intensities:
Up to a very recent date, there have been apparently a consensus
in the Affy world. The individual probe intensities in a probe set
were 'summarized' by one value. The trimmed average of probe intensities
first suggested by the manufacturer was soon followed
more sophisticated approaches, trying to provide robustness by isolating
probes having a obviously erratic behavior.
 One can notice that the hybridization signal must depend on the
nature of the probe (binding energy for example)
and of the nature of the target(s) (cross-hybridization,
self-hybridization, ...), and this is not accounted for in most
of the methods proposed to "summarize" probe intensities into an
"expression value".
Methods for "averaging" the probe intensities in a probe set are
only beginning to consider the physico-chemical properties of the
respective probes in a probe set (in bioconductor, check the packages
'gcrma' and 'affypdnn').
What is needed to make the "summary expression value" paradigm
more reliable is a transformation of individual probe
intensities that discards probe-specific signal, to keep only
an experiment-specific signal. One can see an analogy with the
"between-chips" normalization step, but this time this is at the
probe set level.  
Waiting for this "between-probes normalization", some have taken
the approach of considering the individual probes in a probe set
as separate measures for differential expression. One can see it
as each probe being a referee in a jury, and each referee voting
(or not) for differential expression. The package 'affy' offers
facilities to explore this approach, check the function 'ppsetApply'.

An another issue is the association of probes in probe sets.
The probes were designed to match subsequences in target RNA.
For any given RNA chosen to be address by a chip type, 20 (or 16
or whatever number) probes were chosen to match this RNA. The
problem is that in many cases the gene sequences in databases are
not 100% certain (and get eventually corrected). This what be called
the Dorian Gray syndrome. Currently the probes on any given
Affymetrix chip, the association of probes in probe sets, and the
association of the probe set with a RNA (and its functional
annotation) remain frozen in an apparent eternal youth, while 
their ability to monitor biological phenomena... degrades.
I have built alternative mapping using very recent sets of RNA
reference sequences (NCBI's RefSeqs) and modern Affymetrix chips
(HG-U133A) and observed significant differences using both mappings.
Those differences were enough to cause quite some difference in
the outcome of an analysis.
The (new) package 'altcdfenvs' contains tools to help building
one's own mapping.

These two issues are not completely independent, since "bad mapping"
will certainly result in probe sets with probes showing an erratic
behavior.

>From my last experiences in analyzing Affymetrix data, I would
consider carrying on several analysis (same data, different mappings,
summary and non-summary approaches), especially before going for
long and expensive complementary experiments.


Hopin' it helps,



Laurent



> Hi Bioconductor,
> 
> I'm a new list member and am not quite sure if this question is
> appropriate for the list, but will shoot anyway. I'm analyzing a bunch
> of data from Affy MgU74Av2 chips and am a bit perplexed as to how to
> treat conflicting expression data from multiplicate probe sets (that
> is a gene that has  >1 probe set designed against it (for example,
> 97569_r_at and 97658_r_at are both probes for the Insulin gene). 
> 
> Specifically, if probe #1 for geneX indicates significant fold change
> for that gene, but probe #2 indicates something else (no fold change,
> or even fold change in the opposite direction! (rare, but possible)),
> how can the expression status of geneX be properly evaluated?  Can one
> probe's measurement be considered more reliable than another's (and
> thus toss the one you suspect is wrong (although this could introduce
> experimental bias))? Or is it most appropriate to average the signal
> values for multiplicate probes together? Or is there some other
> method?
> 
> On the MgU74Av2 chip at least, by my calculations there are at least
> 1079 genes that have >1 probe agianst them (2323 probes total that are
> 'multiplicates'), so the numbers are great enough to potentially
> impact my analysis.  Any ideas/suggestions/criticisms will be much
> appreciated.
> 
> with thanks,
> 
> Jonathan Johnnidis
>