Classification (supervised learning) of expression data on pathway level

Question

I was wondering if there is any way to apply classification algorithms (e.g random forest) on microarray data but not using the genes as predictors/features but the pathway they belong to.

The thing is that the expressions of the genes that belong to the same pathway should somehow, grouped together and transformed into a new value that will describe that pathway for each specific sample.

Let me give an example.

Assuming that we have g1, g2, g3, g4, g5, g6, g7, g8, g9, g10 as genes and p1, p2, p3 as pathways.

And g1, g2, g3 belong to p1 while g3, g4, g5, g6belong to p2 and g2, g6, g7, g8, g9, g10 belong to p3

One thought is to calculate the mean expression of the above gene groups per sample and then run the classification algorithm on top of these values. But I would like to hear/read your opinion on such an approach and also your suggestions on using a different method to do this.

The final goal would be the extraction of the most significant classification predictors (pathways).

Answer 1

Have a look at the GSVA package. It allows to convert a matrix with genes x Samples to a pathways x Samples using several methods ssgsea, gage, gsva...

Afterwards you can use that matrix as input for differential expression of pathways or classification algorithms or whatever.

---

However it depends on the input of the "pathways" you give it. Make sure that the pathways are correct. Also consider that the pathway topology might change with conditions.