[BioC] PAM: Applying published classifiers

[Ryan C. Thompson]

Take my advice with a grain of salt. I've just started working with PAM 
and I'm not certain of all the particulars.

On Fri 17 May 2013 03:02:23 PM PDT, Ed Siefker wrote:
>
>
>
> On Fri, May 17, 2013 at 3:43 PM, Ryan C. Thompson
> <rct at thompsonclan.org <mailto:rct at thompsonclan.org>> wrote:
>
>     I can't see how the output of pamr.listgenes would be sufficient
>     to reproduce a trained classifier. I think your only choice would
>     be to re-run PAM starting from the CEL files.
>
>
> Thank you, this is why I got so stuck.  The table listed in their
> paper that's labeled "classifier" is not actually a classifier.   Do
> you have any idea what the list of centroids is used for if not to
> create a classifier?  Wouldn't it be more useful to publish something
> like "pamrtrained.RData.gz", so people can just download it, load the
> object and start classifying?
>
>
>     Also, consider whether their classifier would even be applicable
>     to your microarray samples, since your samples and theirs are
>     normalized separately.
>
>
> One of the papers I'm working off of (DeSousa 2013,
> doi:10.1038/nm.3174) has a flow chart in the supplementary figures
> that shows how they trained the classifier from one dataset of 90
> patients, and then applied that classifier to 5 different datasets
> from several different platforms.  There are no loops in the flow
> chart that indicate they retrained the classifier for each dataset.
> Are they doing it wrong, or is this a valid procedure?
>
> I hope DeSousa 2013 is on topic, as they even provided a bioconductor
> package to repeat their analysis.  I can use that package to recreate
> their classifier pretty easily, but others aren't so convenient.
> Thanks a bunch for the clarification.
> -Ed