GMAMDA: Predicting Metabolite-Disease Associations Based on Adaptive Hardness Negative Sampling and Adaptive Graph...

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Medical Informatics

Background:

• Metabolite concentrations are linked to disease development and progression.
• Accurate prediction of metabolite-disease associations is vital for early diagnosis and treatment.
• Existing prediction models often fail to integrate node features and consider graph structure effectively.

Purpose of the Study:

• To develop a novel model, GMAMDA, for enhanced metabolite-disease association prediction.
• To address limitations in existing methods regarding feature integration and negative sampling.
• To improve the stability and accuracy of predicting potential metabolite-disease links.

Main Methods:

• Constructed multiple heterogeneous graphs using multisource similarity information for metabolites and diseases.
• Employed adaptive graph multiconvolution to generate rich node representations from varied hop neighborhoods.
• Utilized adaptive hardness negative sampling with PCA for selecting high-information negative samples during training.

Main Results:

• GMAMDA demonstrated superior performance over state-of-the-art methods.
• Achieved high evaluation metrics: AUC (0.9962), AUPR (0.9967), and accuracy (0.9733).
• Case studies on Alzheimer's and kidney diseases confirmed GMAMDA's clinical utility in identifying metabolite markers.

Conclusions:

• GMAMDA offers a robust and accurate approach for metabolite-disease association prediction.
• The model's novel strategies enhance prediction stability and clinical relevance.
• GMAMDA holds significant potential for advancing precision medicine through biomarker discovery.