eGoT: enhanced graph-of-thoughts for multi-hop knowledge retrieval and hypothesis generation in biomedicine

Area of Science:

• Biomedical Informatics
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Biomedical research is hindered by fragmented information in unstructured text.
• Large language models (LLMs) offer potential for question answering but struggle with accuracy and evidence provenance.
• Existing retrieval augmented generation methods lack richness and multi-hop reasoning capabilities.

Purpose of the Study:

• To introduce eGoT, a system combining knowledge graph construction and a graph-of-thoughts approach.
• To enable comprehensive, multi-hop reasoning question answering from biomedical literature.
• To improve accuracy and evidence traceability in LLM-based biomedical research.

Main Methods:

• Automated knowledge graph construction from biomedical literature using an LLM-based pipeline.
• Creation of graph and vector databases for storing entities and relationships.
• Multi-round LLM querying of databases for response generation.
• Implementation and benchmarking on MultiHopRAG, HotpotQA, and Ultradomain datasets.

Main Results:

• eGoT demonstrates superior performance compared to state-of-the-art retrieval methods (HopRAG, SireRAG, HiRAG, HippoRAG).
• Successfully generated responses to expert-curated questions on small cell lung cancer from 1046 publications.
• Identified plausible connections between Lupus and UV exposure, impacting disease trajectory.

Conclusions:

• eGoT effectively addresses limitations in current LLM-based biomedical research by providing accurate, traceable, and comprehensive answers.
• The graph-of-thoughts approach enables complex reasoning across diverse biomedical data.
• eGoT shows promise for advancing hypothesis generation and knowledge discovery in medicine.