[BioC] Common microarray clustering method

[Jenny Bryan]

Dear Ilya,

For better or for worse, the most commonly used method is
agglomerative hierarchical clustering.  For gene clustering, people
then tend to prune the resulting tree very high -- that is, one often
chooses K (= the number of clusters) to be much smaller than G (the
number of genes/probes/etc).  Given the output of a hierarchical
algorithm, however, it is wiser to do this with a tree produced by a
*divisive* algorithm.  This choice produces a more stable result
statistically, i.e. small perturbations in the input data tend to
create 'small' perturbations in the output.  An easy, well-established
way to do this is to use the 'diana' function from the R 'cluster'
package.  Agglomerative methods are also, of course, available via
'hclust' (in mva) and 'agnes' (in cluster).

Jenny

Ilya Venger writes:
 > Hi,
 > I'm looking for a most commonly used microarray clustering method. The 
 > main problem is in deciding upon the appropriate amount of clusters to 
 > use in the clustering, both in the agglomerative and teh partitioning 
 > methods.
 > I know there are certain procedures such as MSS or v-fold cross 
 > validation, which I might run. This would allow me to compare 
 > clusterings resulting from incrementing the K (amount of clusters to use).
 > The main point is not only to find the best clustering algorithm, but a 
 > not less importantly, commonly used.
 > 
 > Thanks,
 > Ilya.
 > 
 > 
 >