Heterogeneous graph neural networks for link prediction in biomedical networks

Heterogeneous graph neural networks (HGNNs) are gaining popularity as powerful tools for analysing complex networks with diverse node types often referred to as heterogeneous graphs. While existing HGNNs have been successfully used within the context of social and information networks, their application in biomedicine remains limited. In this study, we posit the utility of readily available generic HGNNs in addressing the link prediction tasks in biomedical settings. Thus, we conduct a benchmarking study of 42 techniques including nine generic HGNNs across eight biomedical datasets using several evaluation metrics. Our results show that the recently developed and readily available generic HGNNs achieve comparable and sometimes better results when compared with the specialized biomedical methods across all evaluation metrics. For instance, the generic HGNN Simple-HGN achieves the best results in four of the eight datasets and shows equivalent performance to the biomedical methods on the remaining datasets. Furthermore, this work analyses and presents useful guidelines to practitioners on how to optimally set complex hyperparameters which underpin the HGNNs.

Finally, this work makes publicly available, via https://github.com/Zaiwen/Link_Prediction_in_Biomedical_Network, the benchmarking framework and source codes which underpin this study.