One of the biggest challenges in drug development is increasing costs of bringing new drugs to the market. Many candidate drugs fail during phase II and III trials due to unexpected side effects and experimental methods remain cost ineffective for large scale discovery of adverse effects. Alternatively, computational methods are used to characterize drug side effects, but they often rely on training predictors based on drug and side effect similarity. Moreover, these methods are typically tailored to the underlying data set and provide little mechanistic insights on the predicted associations. In this study, we investigate the role of network topology in explaining observed side effects of drugs. We find that drug targets are closer in the interactome to the proteins inducing the known side effects of the drug compared to the proteins associated with the rest of the side effects. We show that the interactome based proximity can be used to identify side effects and we highlight a use case in which interactome-based side effect prediction can give insights on drug side effects observed in the clinic.