Predicting disease-metabolite associations based on the metapath aggregation of tripartite heterogeneous networks

Metabolomics reveals the intricate relationship between diseases and metabolites, and identifies disease-specific characteristics and physiological abnormalities by analyzing metabolic changes in organisms, which is of great significance in disease prevention, diagnosis, treatment and pharmaceutical research. However, traditional biological experiments to verify disease-metabolite associations are time-consuming and costly, and the current computational approaches mostly rely on constructed disease-metabolite interaction networks, ignoring the influence of other biological entities. Therefore, this study proposes a novel deep learning model based on Metapath Aggregation of disease-microbe-metabolite Heterogeneous Networks (MAHN) to predict disease-related metabolites.Firstly, the association data among microbes, metabolites and diseases were extracted by integrating multiple databases, and a disease-microbe-metabolite heterogeneous network was constructed based on known associations, metabolite-metabolite, disease-disease and microbe-microbe similarities. Then, the graph convolutional neural network is employed to learn the feature representations of diseases and metabolites separately on different semantic networks with metapath length of 3, followed by the node-level and semantic-level attention to aggregate the embeddings of diseases and metabolites on different semantic networks with metapath length of 2 that fuse microbial information. Finally, the associations between diseases and metabolites are reconstructed using a bilinear decoder. The results demonstrated that the performance of the MAHN model outperforms the four latest algorithms in the five-fold cross-validation. Case studies further confirmed the consistency of the prediction results with relevant experimental validations.