Technical Summary: DeepRoot – A KG-Coordinated Multi-Agent System for Therapeutic Reasoning over Historical Medical Texts

Problem Statement

Historical medical archives and traditional medicines represent a vast, underutilized resource for drug discovery, having provided the basis for major drugs like morphine and artemisinin. However, converting these pre-ontological prose texts into verifiable leads for modern biomedical pipelines is hindered by idiosyncratic taxonomies and a lack of standardized mechanisms. Existing Large Language Model (LLM) agent systems—whether tool-calling, retrieval-augmented, or agentic deep-research—fail to scale the conversion of such text into verifiable drug-discovery leads. They often lack a reasoning trace grounded in verified biological evidence, leading to high rates of hallucination when attempting to bridge historical claims with modern mechanisms.

Methodology: The DeepRoot System

DeepRoot is a multi-agent LLM pipeline designed to jointly build and utilize a verified Knowledge Graph (KG) to perform therapeutic reasoning. The system is divided into two primary phases: Assembly and Discovery, operating over a Neo4j graph.

1. Knowledge Graph Construction (DeepRoot Assembly)

The Assembly phase employs seven specialized agents to populate the KG from raw text (specifically the Shen Nong Ben Cao Jing), ensuring strict verification against curated biomedical databases.

• Extraction & Auditing: An Extraction Agent emits nodes for Sources, Traditional Maladies, and Preparation methods. An Auditor Agent canonicalizes sources (e.g., resolving Latin binomials) and verifies evidence spans against the source text to prevent hallucinated citations.
• Grounding & Linking: Three linker agents ground audited entities to modern biomedical concepts:
  • Source-to-Compound: Links sources to chemical compounds using COCONUT 2.0 and PubChem.
  • Compound-to-Target: Links compounds to molecular targets using ChEMBL.
  • Target-to-Disease: Links targets to modern diseases using Open Targets, NCBI Taxonomy, and OLS4.
• Mapping: A Malady-to-Disease Mapper uses a "generate-then-verify" protocol. The LLM proposes canonical names, but ontology codes (ICD-10, MeSH, SNOMED) are recovered only via tolerant exact matches against external services, eliminating hallucinated identifiers.
• Schema: The resulting graph contains six node types (Source, Traditional Malady, Modern Disease, Chemical Compound, Biological Target, Preparation Method) and seven edge types. A therapeutic claim is considered "mechanistically closed" when a loop exists: Source → Compound → Target → Modern Disease, where the Source also TREATS TRADITIONALLY a Traditional Malady that MAPS TO the Modern Disease.

2. Therapeutic Reasoning (DeepRoot Discovery)

The Discovery phase leverages the constructed KG to evaluate claims and nominate candidates.

• Critic Agent: Evaluates historical Source–Malady claims by traversing the graph to find mechanistic loops. It scores plausibility based on path confidence tiers (T1–T4) and edge evidence types.
• Discovery Agent: Nominates novel compound candidates for a target Modern Disease by walking the graph backward (Disease ← Malady ← Source ← Compound) and forward (Compound → Target → Disease) to score mechanistic plausibility.
• Scoring Logic: Paths are ranked lexicographically by their "weakest link" tier (bucket) and then by the multiplicative product of edge confidences. This ensures that a chain is only as strong as its least verified component.

Key Contributions

1. Separation of Grounding and Reasoning: The paper demonstrates that grounding (retrieving verified facts) and reasoning (synthesizing them) are separable axes. DeepRoot composes these by using a verified KG for grounding and an LLM for reasoning, outperforming systems that conflate the two.
2. Agentic KG Construction: Unlike prior Graph-RAG approaches that rely on expert oversight or pure LLM generation, DeepRoot uses an agentic assembly process with strict, deterministic verification against external databases (COCONUT, PubChem, ChEMBL, Open Targets) to build a high-fidelity graph.
3. Hallucination Suppression: The system shows that building a verified KG a priori suppresses hallucination more effectively than on-demand tool-calling during inference.

Results

The system was evaluated on the Shen Nong Ben Cao Jing corpus (71 chunks) across several benchmarks:

• Held-Out Recovery: In a blind rediscovery task, DeepRoot recovered 10 of 21 held-out compound–disease treatment pairs at Recall@20 (47.6%). This significantly outperformed a raw-corpus LLM baseline (4.8%) and random chance (~2.4%).
• Mechanistic Recovery: DeepRoot Discovery achieved 1.95× higher source recall and 6.11× higher compound recall compared to an LLM baseline when ranking candidates from noisy historical text.
• Reasoning Quality Audit: An independent LLM judge (Claude Sonnet 4.6) evaluated reasoning quality across six dimensions. DeepRoot (KG + LLM) achieved the highest overall scores (3.70–3.83) and the lowest hallucination rates (7–10%).
  • Tool-Calling Baseline: LLMs with direct API access to the same databases hallucinated evidence on 87% of claims.
  • Graph-Only Baseline: Produced 0% hallucinations but scored lowest on reasoning coherence (2.69).
  • DeepRoot: Occupied the optimal middle ground, maintaining low hallucination rates while preserving reasoning synthesis capabilities.
• Ablation Studies: Shuffling KG edges demonstrated that the Critic agent's confidence in therapeutic plausibility degrades linearly with edge perturbation, confirming genuine reliance on graph structure rather than LLM priors.

Significance and Claims

The paper claims that DeepRoot provides a route for the systematic mining and repurposing of historical medical knowledge. By converting pre-ontological prose into an auditable biomedical knowledge graph, the system enables:

• Mechanistic Therapeutic Reasoning: Moving beyond text classification to traceable biological mechanisms.
• Scalable Drug Repurposing: Identifying candidates for de novo nomination and prioritization of natural-product hypotheses.
• Cost-Efficiency: The KG construction is a one-time cost (~$0.25 per corpus), avoiding the recurring expense and hallucination risks of on-demand database retrieval for every query.

The authors emphasize that DeepRoot is a research tool for hypothesis generation, not medical advice. The system's value lies in its ability to prioritize candidates for experimental follow-up by providing traceable, mechanism-grounded chains from historical claims to modern biology, while explicitly flagging cases where the graph cannot close a loop or where evidence is weak.