Deep-DRM

Deep-DRM identifies disease-related metabolites by integrating metabolite chemical structures and disease similarities with graph deep learning to predict metabolite-disease associations.

Key Features:

• Graph-Based Methodology: Uses graph convolutional networks (GCNs) to encode features of integrated metabolite and disease networks.
• Chemical Structure Integration: Calculates metabolite similarities based on chemical structures to inform association predictions.
• Disease Similarity Analysis: Derives disease similarities from functional gene networks and semantic associations to construct disease networks.
• Dimensionality Reduction: Applies principal components analysis (PCA) to reduce feature dimensions while retaining 99% of the information.
• Deep Neural Network Application: Trains a deep neural network to identify true metabolite-disease pairs (MDPs).
• Validation and Performance: Validates predictions using 10-fold cross-validation across three testing setups, reporting AUC 0.952, AUPR 0.939, and independent corroboration of 10 of the top 15 predicted associations.

Scientific Applications:

• Biomarker Discovery: Identifies metabolites associated with diseases to support the discovery of potential biomarkers.
• Targeted Drug Development: Pinpoints specific metabolite-disease associations that can inform targeted therapeutic strategies.
• Mechanistic Insights: Enables investigation of how metabolic changes correlate with genetic and protein-level alterations to elucidate disease mechanisms.

Methodology:

Calculate metabolite similarities from chemical structures; derive disease similarities from functional gene networks and semantic associations; construct integrated metabolite-disease networks; encode network features with GCNs; apply PCA to retain 99% variance; train a deep neural network to predict MDPs and validate with 10-fold cross-validation across three testing setups.