Multitask learning for host–pathogen protein interactions

Kshirsagar, Meghana; Carbonell, Jaime; Klein-Seetharaman, Judith

The interactions listed in the table form the positive class. Because there is no experimental evidence about proteins that do not interact, we construct the 'non-interacting' (i.e. negative) class using a technique commonly used in PPI prediction literature. We use random pairs of proteins sampled from the set of all possible bacteria-human protein pairs. The number of random pairs chosen as the negative class is decided by what we expect the i221 interaction ratio to be. We chose a ratio of 1:100 meaning that we expect 1 in every 100 random bacteria-human protein pairs to interact with each other. In general, there is no basis for choosing a more meaningful ratio, as there are few known interactions. We rely on previous work on better-studied organisms, where a ratio of 1:100 was used, based on the number of known interactions. Further, prior studies (Dyer et al., 2007; Tastan et al., 2009 ) also use a similar ratio. This random selection strategy is likely to introduce $1% false negatives into the training set.