[BioC] PCR Validation threshold in dChip normalized data

[Mark Cowley]

Hi Benjamin,
I agree that a single expression threshold is unlikely to work for all  
probesets, however if you would still like to find a single cutoff,  
then i'd suggest that you plot histograms or density plots of the data  
from each array to identify an approximate threshold at which 2  
populations of probes separate into a clearly low, background set, and  
a higher set of expressed signals.
This method works very well for gcRMA (the 2 populations are very  
distinct), moderately well for RMA (the 2 populations are somewhat  
distinct) and I have no idea for dChip.
Caveat - your data probably needs to be on log scale, so you'll need  
to deal with the negative values in some appropriate way... add an  
offset?

cheers,
Mark

-----------------------------------------------------
Mark Cowley, BSc (Bioinformatics)(Hons)

Peter Wills Bioinformatics Centre
Garvan Institute of Medical Research, Sydney, Australia
-----------------------------------------------------

On 04/09/2008, at 10:45 PM, Benjamin Otto wrote:

> Hi Sean,
>
> First, thanks for the quick reply!
>
> Second, I wouldn't expect a perfect cutoff between possible  
> validation and junk. That would be more a question about judging the  
> data signal range with some tolerance where I can expect the gene to  
> be expressed at all and where the bets are, that I don't see much  
> more than a lot of noise.
>
> Third, probably I was just a little bit too slow to attach my PS  
> comment. What irritated and animated me to post the question was my  
> observation of really negative values in a GEO dataset (GSE3446)  
> which should only be normalized with dCHip without additional  
> transformation. I never, and I must say I rarely used dChip for  
> normalization ... still I never observed negative values in pure  
> dChip normalization. And this dataset has a range from -11000 up to  
> 17000. So that is where I just lost, or still don't have, the  
> slightest feeling for what the dataset range tells me.
> It is not z-score normalized. The sd's are unequal 1. Probably  
> something similar. But then, would you say:
>
> "No judgement possible anymore about "present/absent"-states because  
> the normal states are already curated by some mean value?"
>
> Best regards,
>
> Benjamin
>
>
>
> -----Ursprüngliche Nachricht-----
> Von: seandavi at gmail.com [mailto:seandavi at gmail.com] Im Auftrag von  
> Sean Davis
> Gesendet: Thursday, September 04, 2008 2:34 PM
> An: Benjamin Otto
> Cc: bioconductor at stat.math.ethz.ch
> Betreff: Re: [BioC] PCR Validation threshold in dChip normalized data
>
> On Thu, Sep 4, 2008 at 8:01 AM, Benjamin Otto <b.otto at uke.uni-hamburg.de 
> > wrote:
>> Hi,
>>
>> Given a dataset for Affymetrix arrays normalized with mas5 or rma  
>> we usually
>> made the experience that signals below 80 (6,3 in log2 format) are  
>> hard to
>> validate with PCR.
>>
>> Can somebody tell me, how I can judge on dChip normalized data in a  
>> analog
>> way? Where can I draw a threshold to tell, which signal has good  
>> chances to
>> withstand a verification with with PCR or even on protein level?  
>> And which
>> signals usually indicate a much to low expression level?
>
> I don't think this is an answerable question, exactly.  See Rafael
> Irizarry's work on gene expression barcoding.
>
> http://www.ncbi.nlm.nih.gov/pubmed/17906632
>
> In short, each probeset has a different threshold for expression,
> potentially.  Also, keep in mind that PCR, while held out as a "gold
> standard" is not without its own biases.  Finally, for proteins, all
> bets are off, as there are a number of highly relevant mechanisms for
> regulation of protein expression that occur after transcription.
>
> Not really an answer, but it is reality, I think.
>
> Sean
>
>
>
> -- 
> Pflichtangaben gemäß Gesetz über elektronische Handelsregister und  
> Genossenschaftsregister sowie das Unternehmensregister (EHUG):
>
> Universitätsklinikum Hamburg-Eppendorf
> Körperschaft des öffentlichen Rechts
> Gerichtsstand: Hamburg
>
> Vorstandsmitglieder:
> Prof. Dr. Jörg F. Debatin (Vorsitzender)
> Dr. Alexander Kirstein
> Ricarda Klein
> Prof. Dr. Dr. Uwe Koch-Gromus
> _______________________________________________
> 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