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

Research on biological mechanisms and disease processes is limited by fragmented findings across unstructured text in publications. Question answering and hypothesis generation that can reason across multiple sources can overcome this limitation. However, Large language models (LLMs) are prone to inaccuracies and lack clear provenance to primary evidence. Retrieval augmented generation approaches that have provenance to the original source of evidence address these shortcomings. However, the response richness is dependent on the retrieval process design. Current approaches often fail to produce responses requiring multi-hop reasoning across multiple domains.

To address this, we propose eGoT, which combines automated knowledge graph construction from biomedical literature with a novel graph-of-thoughts approach to query the knowledge base and construct comprehensive responses to natural language questions. Given a corpus of documents, eGoT first uses an LLM-based pipeline to identify and normalize entities and relationships and constructs graph and vector databases. Given an input question, eGoT performs multi-round LLM-based querying of the databases to construct a response. Benchmarking on datasets like MultiHopRAG, HotpotQA, and Ultradomain demonstrates eGoT's superiority over state-of-the-art retrieval methods, including HopRAG, SireRAG, HiRAG, and HippoRAG. We demonstrate eGoT on two biomedical use cases: (i) generate responses to domain expert-curated questions on small cell lung cancer using 1046 PubMed Central publications, and (ii) demonstrate eGoT's ability to find plausible connections between Lupus and climate factors (UV exposure) that affect disease trajectory.

https://github.com/NNeuralDynamics/eGOT.git.