Unlocking Disease Mechanisms: Agentic AI for Clinical Knowledge (Keynote)

This keynote presents Dismech, a mechanistic disease knowledge base designed to support explainable, ontology-grounded AI for clinical and translational reasoning. The talk argues that current AI systems are often effective at prediction, but much weaker at mechanistic explanation, leaving gaps in understanding why specific mutations cause disease, why therapies work, and how patients can be meaningfully stratified. To address this, Dismech represents disease biology as causal graphs that connect genetic etiologies, molecular and cellular processes, environmental modifiers, phenotypic endpoints, treatments, assays, datasets, and models within a LinkML-backed framework enriched with biomedical ontologies such as Mondo, GO, HPO, CL, Uberon, MAXO, and NCIT. The presentation contrasts simple triple-based knowledge graphs with richer causal world models containing latent mechanistic nodes, and outlines applications including interpretation of GWAS signals, Bayesian causal reasoning, and drug repurposing. A major focus is the use of agentic AI to accelerate curation while preserving rigor through deterministic validation, ontology constraints, and anti-hallucination checks for identifiers and references. The talk positions Dismech as an experimental, collaborative platform for computable disease hypotheses and exploratory analysis, while emphasizing that it is not yet ready for clinical deployment.