Modeling polypharmacy effects with heterogeneous signed graph convolutional networks

Pharmaceutical drug combinations can effectively treat various medical conditions. However, some combinations can cause serious adverse drug reactions (ADR). Therefore, predicting ADRs is an essential and challenging task. Some existing studies rely on single-modal information, such as drug-drug interaction or drug-drug similarity, to predict ADRs. However, those approaches ignore relationships among multi-source information. Other studies predict ADRs using integrated multi-modal drug information; however, such studies generally describe these relations as heterogeneous unsigned networks rather than signed ones. In fact, multi-modal relations of drugs can be classified as positive or negative. If these two types of relations are depicted simultaneously, semantic correlation of drugs in the real world can be predicted effectively. Therefore, in this study, we propose an innovative heterogeneous signed network model called SC-DDIS, to learn drug representations. SC-DDIS integrates multi-modal features, such as drug-drug interactions, drug-protein interactions, drug-chemical interactions, and other heterogeneous information, into drug embedding. Drug embedding means using feature vectors to express drugs. Then, the SC-DDIS model is also used for ADR prediction tasks. First, we fuse heterogeneous drug relations, positive/negative, to obtain a drug-drug interaction signed network (DDISN). Then, inspired by social network, we extend structural balance theory and apply it to DDISN. Using extended structural balance theory, we constrain sign propagation in DDISN. We learn final embedding of drugs by training a graph spectral convolutional neural network. Finally, we train a decoding matrix to decode the drug embedding to predict ADRs. Experimental results demonstrate effectiveness of the proposed model compared to several conventional multi-relational prediction approaches and the state-of-the-art deep learning-based Decagon model.