Are topological properties of drug targets based on protein-protein interaction network ready to predict potential drug targets?

Identification of potential druggable targets utilizing protein-protein interactions network (PPIN) has been emerging as a hotspot in drug discovery and development research. However, it remains unclear whether the currently used PPIN topological properties are enough to discriminate the drug targets from non-drug targets. In this study, three-step classification models using different network topological properties were designed and implemented using support vector machine (SVM) to compare the enrichment of known drug targets from non-targets. Surprisingly, none of the models was able to identify more than 75% of the true targets in the test set. It appears that the currently used simple PPIN topological properties are not likely robust enough for prediction of potential drug targets with high confidence, which also echoes similarly unsatisfying prediction data reported previously. However, we proposed that quality and quantity improvement of the protein-protein interactions (PPI) data for model training will help increasing the prediction accuracy.