MAGED: Multimodal attentive graph learning with gene expression dynamics on knowledge graphs for TCM target...

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Traditional medicine offers valuable insights for drug discovery, but linking documented effects to biological mechanisms is challenging.
• Integrating traditional knowledge, clinical symptoms, and molecular regulation is crucial for advancing herbal medicine research.

Purpose of the Study:

• To develop MAGED, a multimodal graph attention learning framework for enhanced Herb-Target Interaction (HTI) prediction.
• To improve accuracy and interpretability by integrating knowledge graphs and gene expression dynamics.

Main Methods:

• MAGED utilizes multi-source data: TCM properties, clinical symptoms, and gene expression dynamics.
• A multimodal fusion encoder embeds herb attributes, combined with functional representations.
• A hierarchical graph attention network links symptoms to gene associations and regulatory pathways.

Main Results:

• MAGED significantly outperforms baseline methods, with a 59.8% improvement in HR@10.
• The model demonstrates superior performance in cold-start scenarios.
• A case study on Scutellaria baicalensis showed 80% of top predicted targets supported by literature or validated.

Conclusions:

• MAGED offers an accurate, interpretable framework for predicting herb-target interactions.
• It effectively bridges traditional knowledge with modern molecular evidence.
• The approach facilitates herbal mechanism discovery and novel therapeutic target identification.