Multi-filter Based Signed Graph Convolutional Networks for Predicting Interactions on Drug Networks

The prediction of drug-drug interactions (DDIs) and drug-target interactions (DTIs) is currently a prominent area of interest within the domain of drug data analysis. Both types of interactions can be modeled as signed links in a graph, in which the attributes of nodes usually have multiple sources. Graph convolutional network (GCN) models, which have theoretical inspirations from graph signal processing (GSP), are effective deep learning methods for drug research. However, it is still open to explore GSP based GCN models to sufficiently utilize spectral information from both signed graph structures and multi-source node attributes. In this study, we propose a multi-filter based signed graph convolutional network (MFSGCN) to handle multiple features of nodes on signed networks. We first extend a rational filter, which is parameterized and has theoretical power and meanings, to signed graphs. Subsequently, we leverage multiple attributes as multi-channel graph signals and implement MFSGCN via learning different parameters of filters. For the sign prediction problems on homogeneous DDIs networks and heterogeneous DTIs networks, we put forward MFSGCN-DDI and MFSGCN-DTI, respectively. The experimental results verify the validity and generalization of MFSGCN and demonstrate the impact of different features and effectiveness of multiple filters.