GMC-DMA: GNN-Mamba Co-Contrastive Optimization for Disease-Metabolite Association Prediction.

Area of Science:

• Biomedical Research
• Computational Biology
• Network Medicine

Background:

• Metabolite level changes are linked to disease development, making metabolite-disease association prediction vital.
• Traditional methods struggle with long-range dependencies and interpretability in complex biological networks.

Purpose of the Study:

• To develop an advanced computational framework for accurate metabolite-disease association prediction.
• To overcome limitations of existing methods in modeling complex biological relationships.

Main Methods:

• A dual-path dynamic contrastive learning framework (GMC-DMA) integrating Graph Neural Networks (GNN) and Mamba architectures.
• Utilized a multi-source heterogeneous network, Residual Graph Convolutional Network (ResGCN), Mamba for Selective State Space Model (SSM), and Fast Kolmogorov-Arnold Networks (FastKAN).
• Employed InfoNCE loss with dynamic negative sampling for robust contrastive learning and a bilinear decoder for probability output.

Main Results:

• GMC-DMA significantly outperformed baseline methods in comprehensive performance for metabolite-disease association prediction.
• Demonstrated effectiveness in identifying disease-related metabolites and high reliability in discovering potential novel associations.
• The Mamba architecture addressed global dependencies efficiently, mitigating the over-smoothing issue in GNNs.

Conclusions:

• GMC-DMA offers a powerful and interpretable approach for metabolite-disease association prediction.
• The framework enhances biomedical research by accurately identifying disease biomarkers and potential therapeutic targets.
• This study highlights the potential of integrating GNNs, Mamba, and FastKAN for complex biological network analysis.