Demo: Preserving Evidence, Context, and Disagreement in Mechanistic Microbe-Host Graph QA

A microbe-to-disease association alone does not show how an organism may affect the host or which evidence supports the intermediate steps. We present an interactive mechanistic evidence graph compiled by a hybrid pipeline of deterministic construction and bounded large language model (LLM) calls, and exposed through constrained natural language graph question answering (QA). The deployed graph stores 131,335 total nodes (16,523 biological entity nodes across 8 rendered entity-class planes and 114,812 paper-provenance nodes) and 162,155 relationships. At serving time, the language model selects from a closed catalogue of read-only graph executors and later narrates the returned rows. Entity resolution, graph execution, evidence pairing, and citation substitution are handled outside the narrator. Multi-label proteins remain one identity with a set of valid roles; 71.3% of the 188 adjudicated protein identity conflicts retained a second valid role. Opposing signed claims remain separate rather than being majority-aggregated, and explicit condition spans are retained when they occur in the supporting sentence. This demonstration shows that a working biomedical graph-QA interface can preserve evidence pairing, contextual information, biological ambiguity, opposing claims, and explicit no-evidence states through natural-language access.