search

TobiasRubel / Pathway-Reconstruction-Tools

2 stars 0 forks source link

PCSF added (it's an UNDIRECTED output) #11

Open annaritz opened 4 years ago

annaritz commented 4 years ago

Make sure that the PR code can handle the undirected graph. In the PCSF outfile, the edge (u,v) may be actually represented as (v,u) in the interactome.

When I convert the interactome to an undirected graph, I take the MIN COST of bidirected edges (corresponding to the highest weight of the edge). I'll write this up in the doc. (Also feel free to change if you have a better idea).

TobiasRubel commented 4 years ago

I'm somewhat puzzled about how best to implement this. When computing PR for PCSF (and whatever methods we are comparing it against), should we treat the output of all other methods as undirected as well, or just PCSF?

If we have the code handle PCSF specially then I'm worried that the result will be a bit misleading, since PCSF is being graded on how well it does a (presumably) easier task than the other algorithms.

Seems like taking the minimum cost is a good way of doing it. The only other thing I can think of making a case for would be combining the weights of the edges, such that the cost is actually lower than the min. The intuitive idea being that an undirected edge a-b tells us that either a->b or b->a, so if P is our confidence that a->b and Q is our confidence that b->a then P+Q - Prob(b->a and a->b) should be our confidence that a-b, which is just P+Q-(P*Q) assuming independence.

annaritz commented 4 years ago

(1) I agree that we shouldn't plot PCSF on the same plot as others if they're using different criteria. Can you make (yet another) option to evaluate all methods in an undirected way? That is, if a method predicts (u,v) but (v,u) is a positive, that edge should be a TP?

(2) Good point, I like your idea to combine the weights. We can change that once you're able to run the code.