Figure 5. Method comparison.

Comparison of interaction transfer sets from various methods for H. sapiens with known H. sapiens interactions from iRefIndex. We compare interaction sets from STRING [62], InteroPORC [21], InterologFinder [24], BIPS:BIANA [25] and our Random-Forest-Filter (RFF). From the STRING database only interactions with interaction transfer information from other species and a combined score over 0.7 are included (STRING(1)). The combined score uses information from all information sources including knowledge on experimental interactions for the respective species (direct evidence). Therefore, an additional interaction set is created where the combined STRING score is recomputed excluding the scores from the direct evidence of databases, experiments and text-mining (STRING(2)). In general, the intersections between the different sets and the known interactions are small. a.) With the RFF and with STRING(1) 10% of the predicted interactions can be found in the experimental data. The modified STRING(2) interaction set is 43% smaller and only 4% of the predicted interactions are consistent with the experimental data showing a clear performance advantage of the RFF for species with no experimentally determined interactions. b.) We compare the interaction sets of RFF, STRING(1), a combined set of unique interactions from InteroPORC, InterologFinder and BIPS:BIANA and a set of known H. sapiens interactions. With the RFF 42% of predicted interactions can also be found in one of the other sets.