The metabolic level within an organism typically reflects its health status. Studying the relationship between human diseases and metabolites helps enhance medical professionals' ability for early disease diagnosis and risk prediction. However, traditional biological experimental methods often require substantial resources and manpower, and there is still room for improvement in the performance of existing predictive models. To tackle these, we propose a novel method based on the Neighborhood Aggregation Graph Transformer (NAGphormer) to predict potential associations between diseases and metabolites (DMNAG), aiming to provide guidance for biological experiments and improve experimental efficiency. First, we calculated the Gaussian kernel similarity of diseases and the physicochemical similarity of metabolites, and combined them with known associations to construct a bipartite heterogeneous network. We then calculated the semantic similarity of diseases and the Mol2vec similarity of metabolites, using them respectively as the similarity feature vectors for the disease nodes and metabolite nodes. Meanwhile, we calculate the positional information features of nodes and combine them with similarity features as the initial features of the nodes. Next, we input the bipartite heterogeneous network and node initial features into the Hop2Token module to capture multihop neighborhood information between nodes. Finally, we input the multi-hop features of nodes into the Transformer model for training and obtain the edge prediction probabilities through the decoder. Through experiments, our model achieved an AUC value of 0.9801 and an AUPR value of 0.9818 in five-fold cross-validation. In case studies, most DMNAG-predicted associations have been validated, showcasing the model's reliability and superiority.
Keywords: Disease-metabolite; Graph transformer; Mol2vec; Neighborhood aggregation; Physicochemical similarity.
Copyright © 2024 Elsevier Ltd. All rights reserved.