Large Language Model-Enhanced Drug Repositioning Knowledge Extraction via Long Chain-of-Thought: Development and...

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Drug Discovery
• Natural Language Processing

Background:

• Drug repositioning accelerates therapeutic discovery but faces challenges with complex, dispersed biomedical data.
• Traditional information extraction methods struggle with accuracy and generalizability.
• Large language models (LLMs) show potential but have issues with hallucinations and interpretability.

Purpose of the Study:

• To introduce Long Chain-of-Thought for Drug Repositioning Knowledge Extraction (LCoDR-KE), a lightweight framework to boost LLM accuracy and adaptability.
• To enhance the extraction of structured biomedical knowledge for drug repositioning applications.

Main Methods:

• Developed a domain-specific schema with 11 entities and 18 relationships.
• Utilized chain-of-thought prompt engineering for automatic annotation on 10,000 PubMed abstracts.
• Curated a specialized drug repositioning corpus from 1,000 expert-validated abstracts.
• Combined supervised fine-tuning of Qwen2.5-7B-Instruct with reinforcement learning and dual-reward mechanisms.

Main Results:

• LCoDR-KE achieved an entity F1 of 81.46% and triplet F1 of 69.04%, outperforming traditional models.
• The framework rivaled larger LLMs, demonstrating the effectiveness of the lightweight approach.
• Ablation studies confirmed significant contributions from supervised fine-tuning and reinforcement learning.

Conclusions:

• LCoDR-KE enhances LLMs' domain-specific adaptability for drug repositioning.
• The framework provides an open-source corpus and a scalable, interpretable solution for biomedical knowledge extraction.
• This approach supports drug discovery and knowledge reasoning, with potential for broader applications.