Abstract:
Drug-drug interactions (DDIs) may give rise to the risk of unanticipated adverse effects, but current DDIs detection is expensive and time-consuming. Recently, graph neural network has achieved significant improvement in DDIs prediction, but the non-negative graph modeled by most methods adapts to assortative relations. Some semantic relationships between drugs, such as degressive effects or adverse side reactions, are actually disassortative relations, which can be described as negative edges. In this study, a method based on signed network was proposed for DDIs prediction. The drug nodes were embedded through a signed graph convolutional network which took the spectral decomposition of the signed Laplacian as the initial input. The problem-specific loss function was used to end-to-end training network model. Through comparative experiments on three test datasets of two prediction problems, it is verified that our method shows good performance in term of evaluation metrics.