Metapath-based Feature Aggregated Heterogeneous Graph Neural Network for Adverse Drug Reactions Predicting

Backgrount: Predicting adverse drug reactions (ADRs) is crucial for drug development, discovery, and design. Although traditional pharmacological experiments can accurately detect the ADRs and have high reliability, the process is costly in terms of time and monetary. With the development and rise of computational methods, graph neural network (GNN)-based ADR prediction methods have gained attention recently since they effectively save research and development costs. Nevertheless, the performance of existing methods is often limited by the size and variety of data they can effectively handle. Furthermore, existing GNN-based methods cannot fully and accurately capture the complex structures and rich semantics of drugs and ADRs in bio-heterogeneous graphs. Result: In this paper, a novel more effective GNN framework named metapath-based feature aggregated heterogeneous graph neural network model (MFAHGN) is proposed for ADRs prediction to achieve more effective and accurate results. In MFAHGN, a drugs-ADRs heterogeneous graph is constructed by integrating multiple well-known biological databases. Then a high-quality feature learning method of drugs and ADRs is designed by using a delicately customized short metapath strategy to capture the complete semantic information of all metapaths, which avoids information decay along the long metapath and fully considers the mutual influence between various metapaths. Conclusions: Extensive experiments are conducted on two biological heterogeneous datasets across six metrics to demonstrate the effectiveness of the proposed model. Medical literature validation confirms 80% of top predictions, including eight novel drug-ADR associations. The code is available at https://github.com/FJH886/MFAHGN.