DeepRoot AI system recovers 47.6% of drug leads from ancient medical texts
Tool-using LLMs hallucinate 87% of claims, but DeepRoot gets it down to 7%
Researchers from Columbia University and industry introduced DeepRoot, a knowledge graph-coordinated multi-agent system designed to mine therapeutic insights from pre-ontological historical medical texts. Traditional drug discovery from archives like the Shen Nong Ben Cao Jing faces challenges due to idiosyncratic taxonomies and prose lacking standardization. Existing LLM agents—whether RAG-based, tool-calling, or deep-research—fail to scale such conversion into verifiable drug-discovery leads. DeepRoot closes this gap by separating grounding (building a verified knowledge graph) from reasoning (using LLMs to compose logical steps), and then composing both for therapeutic reasoning.
Applied to the ancient Chinese medical text, DeepRoot recovered 10 of 21 held-out compound-disease treatment pairs at recall@20 (47.6%), vastly outperforming raw corpus LLM (4.8%) and random baselines (~2.4%). Crucially, tool-using LLMs hallucinated evidence on 87% of claims, while DeepRoot sustained only 7-10% hallucination. Graph-only inference achieved 0% hallucination but scored lowest on reasoning coherence, whereas DeepRoot's combined KG+LLM approach won on both axes. The work, accepted at ICML 2026 GenBio and ACM CAIS 2026, points toward a practical route for systematically repurposing historical medical knowledge into modern drug discovery pipelines.
- DeepRoot recovers 10/21 compound-disease pairs at R@20 (47.6%) from the Shen Nong Ben Cao Jing, vs 4.8% for raw LLM
- Tool-using LLMs hallucinate 87% of claims; DeepRoot cuts hallucination to 7-10%
- Graph-only inference has 0% hallucination but low reasoning coherence; DeepRoot KG+LLM excels on both axes
Why It Matters
Enables systematic, verifiable drug discovery from historical medical texts with drastically reduced hallucination.