Exploring weighting schemes for the discovery of informative generalized between pathway models to uncover pathways in genetic interaction networks.

In S. cerevisiae, a large and rich collection of epistasis data has been collected. When this data comes from double knockouts, it has a natural representation as a signed and weighted graph, where the weight on an edge is computed based on deviation from the expected sickness or health of the double-deletion mutant as compared to its constituent single deletion mutants. Different probabilistic null models (minimum, multiplicative, and logarithmic) to set edge weights appropriately were studied empirically by Mani et al. where the goal was to determine the best weighting scheme for detecting the presence or absence of epistasic effect in an individual double knockout in isolation. On the other hand, approaches such as the LocalCut algorithm of Leiserson et al. look at the entire network, and search for graph-theoretic structure indicative of compensatory pathways. The effect of different edge weighting schemes on the biological pathways returned by algorithms such as LocalCut has not been previously studied. We compare the generalized Between Pathway Models produced by LocalCut under multiple different ways of calculating edge weights, and analyze the resulting collections of putative redundant pathways that are produced. We recover some known pathways, find some interesting new pathways as well as give broad recommendations for how to set the parameters of LocalCut to produce the most biologically relevant gene sets.