Currently I'm building a big matrix (using microarray, mass spectrometry, RNAseq data) that consist by pathways (rows) and treatment/drugs (columns).
The values of that matrix are scores that describe the effect/relation of that specific treatment/drug on each pathway.
Now, some of the drugs is known to target specific disease (target disease) and thus is expected to have a high score with that disease's pathways. (A)
The rest of the drugs is known to target other diseases.(B)
My thought now is that, if I label the known drugs as "1" (or whatever), could I use a prediction model to find out new relationships (links) between drugs from the B class and the significant disease pathway from A ?
The whole thing, we can represent it as a bipartite network shown below.
For example, here the denoted as green, drugs are the known drugs belong to the class A that mentioned before and they are connected highly with some pathways of the target disease.
From the other hand blue and red ones are drugs taken from other cases (class B) that are used to target other diseases, but as my plot shows, blue seems to has an "affinity" to some of the pathways of the target disease and thus should be an alternative candidate for that disease. The red from the other side, doesn't have any "affinity" to any significant pathway and thus we can not consider it as possible candidate.
To find a solution to that kind of problem, I looked for algorithms and methods developed for social networks (e.g recommendation etc.) and mostly I read about random walks with return.
But before to proceed, I would like to know if someone else tried such an approach here and having more experience.