Agentic Curation of Disease Mechanisms (NCATS IFX Retreat)
Description
Invited talk at the Annual Retreat of the Informatics Core (IFX), Division of Preclinical Innovation (DPI), National Center for Advancing Translational Sciences (NCATS), 5 November 2026. Knowledge graphs and phenotype annotations tell us which genes go with which phenotypes, but not how a disease works in between. The talk introduces DisMech (dismech.monarchinitiative.org), a knowledge base of disease mechanisms in which each disease has a pathograph: a causal graph from genetic cause through molecular, cellular and tissue-level events to phenotypes. Every node is grounded in ontologies (HGNC, GO, CHEBI, CL, UBERON, HPO, MONDO), and every claim carries a quoted snippet from its source publication. Fabry disease serves as a worked example, showing how diagnostics, treatments, disease models and new approach methodologies (NAMs), competing hypotheses and open knowledge gaps attach to specific nodes of the graph.
Because treatments are linked to the mechanism node they act on, and nodes are shared across diseases through mechanism modules, the graph can be used to generate drug repurposing hypotheses. The talk presents examples already recorded in DisMech (DFMO in two polyamine disorders, epalrestat in PMM2-CDG, losartan on a TGF-beta node shared across aortopathies, glucosylceramide synthase inhibitors across lysosomal storage disorders, and a negative result for chlorpromazine in ALS), along with counts of shared mechanism nodes, and stresses that these are candidates for testing rather than predictions. It also describes how DisMech connects to the NCATS Biomedical Data Translator: a single LinkML-validated YAML source exports Biolink-compliant KGX edges to the Monarch knowledge graph, with mechanism-level edges as the next step.
The final part covers how DisMech is built: general-purpose AI coding agents (Claude Code, Codex) working inside a domain harness of skills, schema documentation, tools and validators, with every term ID checked against its label and every quote checked against its source before a pull request is merged. Humans govern and manage the loop, and the people driving the agents have included curators, a drug developer and parents of patients with rare diseases. Source: github.com/monarch-initiative/dismech.