Let Me Cypher That for You: A Multi-Graph Agent for Deciphering Biomedical Knowledge

Background: Crosstalk between infectious and non-infectious diseases may contribute to the initiation and progression of neurodegeneration. The COMMUTE project investigates post-COVID-19 multimorbidities and their potential links to an increased risk of neurodegenerative diseases. Hypothesis-driven biomedical knowledge graphs offer a powerful framework to represent and analyze complex relationships and mechanistic links between COVID-19 and neurodegeneration. However, interacting with these graphs typically requires proficiency in Cypher, a graph query language, which poses a technical barrier for many researchers and clinicians. Results: We developed CGEx (Cypher Generating Expert), an interactive system that enables users to query multiple heterogeneous COVID-19 and neurodegenerative disease knowledge graphs using natural-language questions. CGEx translates user questions into Cypher queries, retrieves relevant subgraphs, and generates explanatory natural-language responses. The system incorporates a user-feedback mechanism and a dynamically updated few-shot example store to improve query generation. CGEx2.0 extends the original system by using GPT-5 and providing interactive solution-graph visualization with access to supporting evidence associated with individual relationships. We evaluated the original CGEx and CGEx2.0 using 52 manually curated biomedical questions covering biological pathways, molecular interactions, comorbidity relationships, and multi-hop reasoning. Accuracy was defined as generation of a syntactically and semantically correct Cypher query within three or fewer feedback iterations. CGEx2.0 achieved accuracies of 94.2% and 82.7% on the two graphs, compared with 82.7% and 55.8%, respectively, for the original system. Conclusions: Our results show that CGEx streamlines the query formulation process by removing the need to learn or manually write Cypher, while uncovering meaningful comorbidity insights. Rather than harmonizing heterogeneous graphs into a single schema, CGEx illustrates how multi-graph agents can serve as a complementary paradigm to harmonization, enabling researchers to query diverse evidence sources directly.