A large language model framework for literature-based disease-gene association prediction

Area of Science:

• Biomedical Informatics
• Computational Biology
• Precision Medicine

Background:

• The exponential growth of biomedical literature necessitates advanced tools for knowledge extraction.
• Current Large Language Model (LLM) approaches struggle with reliability, verifiability, and scalability in analyzing complex biological relationships.

Purpose of the Study:

• To develop a novel, unsupervised methodology for automated biomedical literature understanding using LLMs.
• To overcome limitations in current LLM applications for extracting complex biological relationships and advancing precision medicine.

Main Methods:

• Proposed LORE, a two-stage reading methodology utilizing LLMs.
• Modeled biomedical literature as a knowledge graph of verifiable factual statements and semantic embeddings.
• Applied LORE to PubMed abstracts for gene-disease relationship extraction.

Main Results:

• LORE effectively captured essential gene pathogenicity information from PubMed abstracts.
• Achieved 90% mean average precision in identifying relevant genes across 2097 diseases by modeling latent pathogenic flow with ClinVar supervision.
• Demonstrated scalability and reproducibility in biomedical literature analysis.

Conclusions:

• LORE offers a scalable and reproducible approach for leveraging LLMs in biomedical literature analysis.
• This methodology enhances the identification of therapeutic targets by efficiently extracting disease-gene relationships.
• Paves the way for improved precision medicine through automated knowledge discovery.