Basic Science and Pathogenesis.

Area of Science:

• Neuroscience
• Computational Biology
• Biochemistry

Background:

• Alzheimer's Disease (AD) involves complex molecular pathways beyond classical pathologies, including histone deacetylase 6 (HDAC6).
• The increasing volume of scientific literature necessitates advanced methods for knowledge extraction and pathway analysis in AD.
• Large language models (LLMs) offer a powerful solution for summarizing and discovering knowledge within extensive biomedical texts.

Purpose of the Study:

• To develop and evaluate an LLM-based approach for constructing knowledge graphs (KGs) of molecular interactions in Alzheimer's Disease (AD).
• To systematically map the intricate molecular pathways involving histone deacetylase 6 (HDAC6) in AD.
• To compare the performance of different LLMs in extracting structured information from scientific abstracts.

Main Methods:

• A corpus of 265 PubMed abstracts on AD and HDAC was processed using GPT-4o and Gemini.
• LLMs extracted subject-predicate-object triplets, standardizing molecular entities (via UniProt) and verb categories.
• Extracted triplets were aggregated into a knowledge graph (KG), with performance assessed by human evaluation of accuracy.

Main Results:

• A knowledge graph was constructed, identifying key molecular nodes such as Tau, compound 15, and HDAC6.
• GPT-4o achieved a higher accuracy of 78.4% in extracting molecular interactions compared to Gemini's 58.9%.
• The KG provides a structured overview of known molecular interactions and potential research gaps in AD.

Conclusions:

• LLMs, particularly GPT-4o, can efficiently extract structured knowledge from AD and HDAC research, facilitating systematic understanding.
• The developed KG approach serves as a foundation for identifying novel therapeutic targets in AD.
• This methodology shows significant potential for automating scientific discovery across various research fields.