Drug Repositioning Based on Expert Knowledge Augmented Graph Neural Network.

Area of Science:

• Computational biology and cheminformatics.
• Drug discovery and development.
• Artificial intelligence in healthcare.

Background:

• Drug repositioning accelerates the identification of new therapeutic indications for existing drugs.
• Graph neural networks (GNNs) are effective for modeling drug-disease associations but often use randomly initialized node embeddings.
• Existing GNN methods lack interpretability and fail to incorporate valuable expert knowledge from biological databases.

Purpose of the Study:

• To develop a novel framework, DReKGNN, for drug repositioning that integrates expert knowledge into GNNs.
• To enhance the interpretability and accuracy of node embeddings used in drug-disease association prediction.
• To improve the efficiency and effectiveness of drug discovery pipelines.

Main Methods:

• DReKGNN utilizes large language models (LLMs) as a semantic bridge to incorporate expert knowledge from DrugBank and OMIM databases.
• Expert knowledge descriptions, focusing on biological mechanisms, are extracted directly from databases, avoiding prompt templates.
• LLM-generated node embeddings are integrated with GNNs using a mean aggregation strategy to mitigate noise and improve predictions.

Main Results:

• Experimental results demonstrate the superior performance of DReKGNN compared to existing methods in predicting drug-disease associations.
• Case studies further validate the effectiveness of the DReKGNN framework.
• The generated node embeddings are interpretable and aligned with expert biological knowledge.

Conclusions:

• DReKGNN effectively enhances drug repositioning by integrating expert knowledge through LLMs and GNNs.
• The framework provides interpretable node embeddings, advancing the field of AI-driven drug discovery.
• DReKGNN offers a promising approach for accelerating the identification of novel drug indications.