Predicting Drug Drug Interactions by Signed Graph Filtering-Based Convolutional Networks

Drug drug interactions (DDIs) are crucial for drug research and pharmacologia. Recently, graph neural networks (GNNs) have handled these interactions successfully and shown great predictive performance, but most computational approaches are built on an unsigned graph that commonly represents assortative relations between similar nodes. Semantic correlation between drugs, such as degressive effects or even adverse side reactions (ADRs), should be disassortative. This kind of DDIs networks can be represented as a signed graph taking drug profiles as node attributes, but negative edges have brought challenges to node embedding methods. We first propose a signed graph filtering-based convolutional network (SGFCN) for drug representations, which integrates both signed graph structures and drug profiles. Node features as graph signals are transited and aggregated with dedicated spectral filters that capture both assortativity and disassortativity of drug pairs. Furthermore, we put forward an end-to-end learning framework for DDIs, via training SGFCN together with a joint discriminator under a problem-specific loss function. Comparing with signed spectral embedding and graph convolutional networks, results on two prediction problems show SGFCN is encouraging in terms of metric indicators, and still achieves considerable level with a small-size model.