1
$\begingroup$

Suppose I have 50 scale parameters, these are all genes measured for one sample from a subject at the clinic, after data reduction by PCA, two meaningful components were extracted. This was followed by cluster analysis and turned out to be 4 meaningful clusters of subjects based on the two components of these 50 genes.

Since investigating 50 genes for one subject would be costy, one would like to reduce that number so that the same clustering pattern can still be obtained but with minimal costs possible ( there should be some measures here to say acceptable clustering or not, I wonder what kind of measures would fit this case though).

Of course, the more genes investigated, the more information gained, but there should be some measure to tell when to stop wasting more money when the same result is satisfactorily achievable will less number of genes.

Is there any R package that already implemented this approach? what would be the statistical approach in this case to select the most important genes that would preserve the clustering pattern? what criteria to be used in order to reach the minimum clustering pattern?

Improve this question
$\endgroup$
10
  • 1
    $\begingroup$ Some of your questions (all interesting and worthy) reveal nonetheless a curious pattern. You first go into a doctoral discourse, as if posing a problem. Next the description of your own approach should be coming... But instead, abruptly coming is request of an R package already impementing some "right" approach. $\endgroup$
    – ttnphns
    Commented Nov 25, 2013 at 22:04
  • $\begingroup$ Assuming the clusters are theoretically meaningful, you could use PLS followed by LDA using the package plsgenomics (in R).PLS finds orthogonal axes in the data which are maximally related to Y(by decomposing the crossprod(Y, X) matrix),where Y would be a priori groups (your meaningful clusters). One can then use the latent variables which maximize the PRESS statistic in LDA to determine decision rules for discriminating. From there you could assess the loadings of the PLS model along with variable importance scores to determine which genes are needed to classify individuals into the groups. $\endgroup$
    – Patrick
    Commented Nov 26, 2013 at 1:52
  • $\begingroup$ @Patrick, I don't have Y (response) variable in the original data. But I have like two groups of X variables (like 10 and 40) they contribute to the construction of the first two components, so how to do PLS in this case? All 4 clusters were from 50 genes or the two components. $\endgroup$
    – doctorate
    Commented Nov 26, 2013 at 7:49
  • $\begingroup$ @Patrick, do you know any reading material that implemented such approach by the mentioned R package? that would be helpful. $\endgroup$
    – doctorate
    Commented Nov 26, 2013 at 7:54
  • $\begingroup$ Considering, LDA the variables don't show normality of distribution (at least not all) so AFAIK LDA is not tenable when such assumption is not met, right? so what would be the alternative in this case? $\endgroup$
    – doctorate
    Commented Nov 26, 2013 at 8:42

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.