An integrative label propagation framework to predict drug-drug interaction (DDI) by integrating label side effects, off-label side effects, and chemical structures. A systematic comparison of the experimental results shows (1) side effect profiles are more predictive features than chemical structures in DDI prediction. It greatly benefits from the fact that clinical side effects are human phenotypic data obviating translation issues. (2) label propagation algorithm boosted the DDI prediction by considering high-order relationships between drugs. (3) our proposed integrative label propagation algorithm effectively integrated multiple drug properties and outperformed competitors. Furthermore, we applied the proposed algorithm to all known drugs which have one or more side effect profiles and obtained 145,068 predicted DDIs. These predicted DDIs can be leveraged for clinical surveillance and real-world drug discovery.