Hi All,
I am interested in developing a prognostic classifier in the following way:
- Use a large RNA-Seq cohort with SVM and LOOCV with RFE to develop a prognostic classifier that is able to distinguish between two phenotypes.
- I want take this classifier and test it on a different Microarray cohort to see how well these genes can discriminate between the same two phenotypes as described within the RNAseq data.
Since the distributions of these two data-types vary, we simply variance stabilize transform the counts which makes it more normal instead of neg. binomial.
The first point above is straightforward. My question is within the second point above. How can one take a classifier that was already independently created and validated to distinguish between two phenotypes using one cohort (RNAseq) but then test the accuracy of that multi-gene classifier to distinguish between the same two phenotypes using another cohort (microarray)?
Yours,
M.B.
This is probably not a solution to your problem, but have a look at this paper, where they also have to address the multiplatform issue. Maybe that'll at least provide a springboard for a solution.
Classification with binary gene-expression was an idea. I see it was put out there in 2009 but I was completely unaware of this paper. Cheers, this may provide useful for us.
If you switch classification methods to work with a score for relative expression between two groups, this might also help.
For example, I was able to apply a classifier from a microarray dataset to an RNA-Seq dataset using BD-Func:
https://peerj.com/articles/159/
Cheers guys. I'll meditate a bit on these approaches and test them out. I still think it is quite surprising though that still you can not take a classifier from one study and put it on to another, using expression, to see how it performs.