Predicting Disease-Metabolite Associations Based on the Metapath Aggregation of Tripartite Heterogeneous Networks

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Disease-metabolite interactions are crucial for diagnostics and therapeutics.
• Traditional experimental methods are inefficient; current computational methods lack holistic biological context.

Purpose of the Study:

• To develop a novel deep learning model, Metapath Aggregation of Heterogeneous Networks (MAHN), for identifying disease-related metabolites.
• To overcome limitations of existing computational methods by incorporating microbial influences.

Main Methods:

• Constructed a tripartite heterogeneous network including microbes.
• Employed graph convolutional network and enhanced GraphSAGE for feature learning (metapath length 3).
• Utilized attention mechanisms for feature aggregation (metapath length 2) and a bilinear decoder for association probability.

Main Results:

• MAHN achieved high performance in cross-validation (Acc 91.85%, AUC 97.39%).
• Outperformed four state-of-the-art algorithms.
• Case studies on irritable bowel syndrome and obesity validated MAHN's predictive accuracy.

Conclusions:

• MAHN is a reliable tool for discovering potential disease-related metabolites.
• Deep learning models integrating multi-omics data are the future of predicting biological entities related to diseases.