Centrality as a Predictor of Lethal Proteins: Performance and Robustness

The Centrality-Lethality Hypothesis states that proteins with a higher degree centrality are more likely to be lethal, i.e. proteins involved inmore interactions aremore likely to cause deathwhen knocked off. This proposition gave rise to several new investigations in which stronger associations were obtained for other centrality measures. Most of this previouswork focused on thewell known protein-interaction network of Saccharomyces cerevisiae. In a recent study, however, it was found that degree and betweenness of lethal proteins is significantly above average across 20 different protein-interaction networks. Closeness centrality, on the other hand, did not perform as well. We replicate this study and show that the reported results are due largely to a misapplication of closeness to disconnected networks. A more suitable variant actually turns out to be a better predictor than betweenness and degree in most of the networks. Worse, we find that despite the different theoretical explanations they offer, the performance ranking of centrality indices varies across networks and depends on the somewhat arbitrary derivation of binary network data from unreliable measurements. Our results suggest that the celebrated hypothesis is not supported by data.