CHETR: A Typed Neuro-Symbolic Agent for Mechanistic Drug Repurposing

Agents4Science 2025 Conference Submission327 Authors

Keywords: drug repurposing, neuro-symbolic agents, typed knowledge graphs, causal inference, instrumental variables, counterfactual analysis, isotonic calibration, leakage-safe evaluation

TL;DR: Typed neuro-symbolic agent that certifies mechanistic drug-repurposing paths via counterfactual ablation, IV envelopes, and translational feasibility—fully computational and leakage-safe (no wet-lab).

Abstract: We present \textbf{CHETR} (Causal Hypothesis Engine for Therapeutic Repurposing), a typed neuro-symbolic agent that generates \emph{mechanistic}, testable repurposing hypotheses \emph{without} relying on wet-lab results. CHETR couples an LLM planner with a typed biomedical knowledge graph (KG), audits explanations via a \emph{Mechanism Causal Support} (MCS) certificate (counterfactual path ablation, instrumental-variables envelope, and perturbation concordance), and screens translational plausibility with a \emph{Translational Feasibility Index} (TFI; exposure margins, BBB/P-gp/BCRP liabilities (efflux), and DDI flags). We formalize the scoring functional, prove a \emph{faithfulness lower bound} linking rank changes to mechanism-critical edge removals, and show a \emph{soundness condition} under typed \(d\)-separation that prevents spurious mechanisms from dominating. We also specify a leakage–safe, time–split evaluation protocol that is entirely computational. A fully worked micro-example demonstrates the end-to-end pipeline and certificate values. The result is a theory-first, auditable protocol that converts LLM pattern suggestions into \emph{certified} mechanistic hypotheses suitable for prospective testing and reproducible AI-led science.

Submission Number: 327