Identifying Metabolite-Disease Associations via Messaging in Hypergraphs.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Traditional machine learning struggles with integrating diverse biological data for metabolite-disease prediction.
• Pairwise graph models are insufficient for complex multi-way interactions between metabolites, diseases, proteins, and Gene Ontology (GO) annotations.

Purpose of the Study:

• To develop an advanced computational framework for accurately predicting metabolite-disease associations.
• To overcome limitations of existing methods in modeling intricate biological networks.

Main Methods:

• Developed a novel hypergraph-based framework (DHG-LGB) conceptualizing diseases as hyperedges.
• Utilized hypergraph neural networks (HGNNs) for encoding metabolite-disease relationships into low-dimensional vectors.
• Employed LightGBM (LGB) for building the predictive model.

Main Results:

• DHG-LGB achieved high performance metrics including 98.87% accuracy, 91.77% sensitivity, and 0.9983 AUC.
• The framework demonstrated excellent robustness and generalization across varying data ratios (AUC > 0.9954).
• Comparative analysis confirmed the superiority of DHG-LGB over existing methodologies.

Conclusions:

• The DHG-LGB framework provides a more comprehensive model of biological interactions than traditional methods.
• It significantly improves the predictive accuracy of metabolite-disease relationships.
• DHG-LGB is a promising computational tool for biomarker identification and advancing precision medicine.