Empowering Graph Neural Network-Based Computational Drug Repositioning with Large Language Model-Inferred Knowledge...

Area of Science:

• Computational biology
• Artificial intelligence in medicine
• Pharmacology

Background:

• Drug repositioning aims to identify new therapeutic uses for existing drugs.
• Current graph neural network (GNN) methods for drug-disease association (DDA) prediction rely heavily on network topology, which can be incomplete or noisy.
• These GNN methods often overlook rich biomedical knowledge crucial for accurate predictions.

Purpose of the Study:

• To investigate the effectiveness of large language models (LLMs) in generating knowledge representations for drug repositioning and DDA prediction.
• To develop and evaluate novel computational models that integrate LLM-generated knowledge into DDA prediction frameworks.
• To determine the optimal method for fusing LLM-based embeddings with existing network data for improved DDA prediction.

Main Methods:

• Utilized a zero-shot prompting template with LLMs to extract high-quality knowledge descriptions for drug and disease entities.
• Generated continuous numerical representations (embeddings) from LLM-derived text descriptions.
• Proposed LLM-DDA, a model incorporating LLM embeddings, with three architectures (LLM-DDA_{Node Feat}, LLM-DDA_{Dual GNN}, LLM-DDA_GNN-AE) to explore fusion strategies.
• Conducted extensive experiments on four DDA benchmarks and compared performance against 11 baseline methods.

Main Results:

• The LLM-DDA_GNN-AE architecture achieved optimal performance, outperforming 11 baseline methods.
• Demonstrated significant relative improvements: 23.22% in Area Under the Precision-Recall Curve (AUPR), 17.20% in F1-Score, and 25.35% in precision.
• Case studies, including Prednisone for Allergic Rhinitis, validated the model's ability to identify reliable DDAs and knowledge descriptions supported by literature.

Conclusions:

• LLMs significantly enhance drug repositioning by providing rich, integrated biomedical knowledge representations.
• The proposed LLM-DDA_GNN-AE model represents a state-of-the-art approach for DDA prediction.
• LLMs offer broad utility and applicability for various biomedical relation prediction tasks beyond drug repositioning.