Paper deep dive
DeepRoot: A KG-Coordinated Multi-Agent System for Therapeutic Reasoning over Historical Medical Texts
Zijian Carl Ma, Sean J. Wang, Sijbren Kramer, Li Erran Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 6/20/2026, 7:33:44 AM
Summary
DeepRoot is a multi-agent LLM system designed for therapeutic reasoning over historical medical texts, specifically the Shen Nong Ben Cao Jing. The system uses a two-stage process: 'Assembly' (using seven specialized agents to build a verified Neo4j knowledge graph) and 'Discovery' (using critic and discovery agents to traverse the graph). The research demonstrates that combining a verified Knowledge Graph (KG) with LLMs significantly reduces hallucinations and improves the recovery of compound-disease treatment pairs compared to raw LLM or tool-calling-only approaches. The system successfully recovered 10 of 21 held-out treatment pairs (R@20 of 47.6%) and showed superior reasoning quality in benchmarks against baseline LLMs and graph-only inference.
Entities (10)
Relation Signals (5)
Source â contains â Chemical Compound
confidence 100% · the source contains a Chemical Compound
Traditional Malady â mapsto â Modern Disease
confidence 100% · a Traditional Malady that maps to a Modern Disease
Biological Target â relatesto â Modern Disease
confidence 100% · Biological Target which itself relates to that same Modern Disease.
Chemical Compound â targets â Biological Target
confidence 100% · Chemical Compound that targets a Biological Target
Source â treatstraditionally â Traditional Malady
confidence 100% · a Source treats a Traditional Malady that maps to a Modern Disease
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Historical medical archives and traditional medicines hold immense potential for drug discovery and remain a primary source for current drug development. However, pre-ontological prose and idiosyncratic taxonomies prevent the standardization and medical modernization of the data for use in current biomedical pipelines. Furthermore, no existing LLM agent system, whether tool-calling, retrieval-augmented, or agentic deep-research, can convert such text into verifiable drug-discovery leads at scale. We close this gap with DeepRoot, a multi-agent LLM system that jointly builds and utilizes a verified knowledge graph, showing that grounding and reasoning -- often conflated -- are separable axes the system can compose for therapeutic reasoning. Applied to the Shen Nong Ben Cao Jing, DeepRoot recovers $10$ of $21$ held-out compound-disease treatment pairs at R@$20$ ($47.6\%$ vs $4.8\%$ for a raw corpus LLM and $\sim\!2.4\%$ random) and dominates an LLM-as-judge audit for reasoning quality over baseline LLMs and LLMs with direct tool-call access to the same APIs DeepRoot itself queries. Tool-using LLMs hallucinate evidence on $87\%$ of claims, versus 7-10% for DeepRoot. Graph-only inference hallucinates $0\%$ but ranks lowest on reasoning coherence; DeepRoot KG+LLM is the only condition to win on both axes, pointing toward a route for systematic mining and repurposing of historical medical knowledge.
Tags
Links
- Source: https://arxiv.org/abs/2606.15931v1
- Canonical: https://arxiv.org/abs/2606.15931v1
Trouble viewing inline? Open PDF directly â
Full Text
76,201 characters extracted from source content.
Expand or collapse full text
DeepRoot: A KG-Coordinated Multi-Agent System for Therapeutic Reasoning over Historical Medical Texts Zijian (Carl) Ma * 1 2 Sean J. Wang * 1 2 Sijbren Kramer * 1 2 Li Erran Li 3 Abstract Historical medical archives and traditional medicines hold immense potential for drug discov- ery and remain a primary source for current drug development. However, pre-ontological prose and idiosyncratic taxonomies prevent the standard- ization and medical modernization of the data for use in current biomedical pipelines. Further- more, no existing LLM agent system, whether tool-calling, retrieval-augmented, or agentic deep- research, can convert such text into verifiable drug-discovery leads at scale. We close this gap with DeepRoot, a multi-agent LLM system that jointly builds and utilizes a verified knowledge graph, showing that grounding and reasoningâ often conflatedâare separable axes the system can compose for therapeutic reasoning. Applied to the Shen Nong Ben Cao Jing, DeepRoot re- covers10of21held-out compoundâdisease treat- ment pairs at R@20(47.6%vs.4.8%for a raw corpus LLM andâŒ2.4%random) and dominates an LLM-as-judge audit for reasoning quality over baseline LLMs and LLMs with direct tool-call access to the same APIs DeepRoot itself queries. Tool-using LLMs hallucinate evidence on87% of claims, versus7â10%for DeepRoot. Graph- only inference hallucinates0%but ranks lowest on reasoning coherence; DeepRoot KG + LLM is the only condition to win on both axes, pointing toward a route for systematic mining and repur- posing of historical medical knowledge. 1. Introduction Natural productsâchemical compounds synthesized by liv- ing organismsâremain the leading source of approved * Equal contribution 1 Department of Bioengineering 2 Stanford University 3 Amazon AWS AI. Correspondence to: Zijian (Carl) Ma <mazijian@stanford.edu>. Accepted at the 2026 Workshop on Generative and Agentic AI for Biology (ICML 2026), Seoul, South Korea. PMLR 306, 2026. Copyright 2026 by the author(s). drugs and provide scaffolds for developing more potent derivatives (Newman & Cragg, 2020; Koehn, 2012). Many natural products have been uncovered through mining tradi- tional medicines, including morphine from opium poppies and the antimalarial artemisinin, with the latter isolated by Tu Youyou after consulting a 4th-century Chinese medical text (Tu, 2011; Brook et al., 2017). ML, DL, and LLM approaches for mining historical medical texts at scale have been reported before but treat the text as pure input-classification problems without a reasoning trace grounded in verified biological evidence or mechanism ontologies (Li et al., 2024; Hui et al., 2020; Liu et al., 2025; Dai et al., 2024). In parallel, multi-agent LLM systems leverage a shared knowledge graph (KG) for coordinated reasoning (Ghafarollahi & Buehler, 2025; Rasmussen et al., 2025), but only qualitatively: they neither ablate the graph against agent decomposition, nor evaluate on the regimes we targetâhistorical clinical cases where traditional text lacks clean ontological anchors, and discovery problems with sparse ground truth. Building on these advances, we introduce DeepRoot (Fig- ure 1), a multi-agent LLM pipeline where agents collectively construct and reason over a shared KG (Neo4j). Closest to our work is OpenTCM (He et al., 2025), which uses a Graph-RAG architecture for LLM reasoning. However, its construction relied on expert oversight and pure LLM- generated outputs. DeepRoot Assembly agentically pop- ulates the knowledge graph via seven specialized agents that combine LLM canonicalization with strict verification against curated biomedical databases. DeepRoot Discov- ery then employs critic and discovery agents, leveraging Neo4j Cypher walks for subgraph traversal to evaluate ther- apeutic claims and identify potential therapeutics that are mechanistically grounded in the KG. 2. Methods and KG construction 2.1. Dataset and grounding sources Corpus. We evaluate on the Shen Nong Ben Cao Jing materia medica, segmented into 71 chunks. The corpus catalogues plants, animals, and minerals (sources), maladies, preparation methods, and claimed therapeutic uses. 1 arXiv:2606.15931v1 [cs.MA] 14 Jun 2026 Submission and Formatting Instructions for ICML 2026 Traditional Malady Source TREATS_ TRADITIONALLY Modern Disease Chemical Compound Biological Target Historical Corpus AGENT 3 Malady Disease Mapper AGENT 4 Source Compound Mapper AGENT 5 Compound Target Linker AGENT 6 Target Disease Linker AGENT 7 Reviewer Agent AGENT 1 Extraction Agent AGENT 2 Extraction Auditor AGENT 8 Critic Agent AGENT 9 Validator Agent DeepRoot Assembly Neo4j Knowledge Graph Preparation Method âFritillaria opens depression and resolves binding...â DeepRoot Discovery Betulin âPTNP1 â Diabetes Asiatic acid α-amyrin Figure 1. Schematic of DeepRoot. Graph nodes and edges are represented by rounded rectangles and black arrows. Gray arrows indicate creation of specific nodes and edges by particular agents. External grounding. Every entity is verified against curated biomedical databases. Sources are linked to compounds via COCONUT2.0 (Chandrasekhar et al., 2025) (natural prod- ucts) and PubChem (Kim et al., 2023) (chemicals); com- pounds are linked to molecular targets and clinical indica- tions via ChEMBL (Mendez et al., 2019); protein targets are linked to diseases via Open Targets (Ochoa et al., 2021), and pathogenic-organism targets via NCBI Taxonomy (Schoch et al., 2020) with OLS4 (McLaughlin et al., 2025). Modern disease nodes are anchored to ICD-10, MeSH, SNOMED, MONDO, and DOID identifiers via NLM and EBI lookup services. 2.2. Knowledge graph schema The graph has six node types and seven edge types (Fig- ure 1). A therapeutic claim is verifiable when its mechanistic loop closes: aSourcetreats aTraditionalMalady that maps to aModernDisease;the source contains aChemicalCompoundthat targets a BiologicalTargetwhich itself relates to that same ModernDisease.Identity for compounds is the RDKit-computed InChIKey and identity for targets is the curated ChEMBL ID, so equivalent entities arriving from different routes collapse onto the same node. Full schema is tabulated in Table S1. Assembly. Seven specialized agents populate the graph in dependency order: an extractor emits Source, Malady, and Preparation nodes from raw text; an auditor canoni- calizes sources and archives evidence spans that fail sub- string verification against their source chunk; three linkers ground audited entities to compounds, molecular targets, and target-to-disease associations using the databases above; a malady-to-disease mapper follows a generate-then-verify protocol in which LLMs propose canonical names and on- tology codes are recovered only by tolerant exact match, eliminating hallucinated identifiers; and a reviewer archives orphans and off-domain entities. Resulting graph. On Shen Nong Ben Cao Jing, As- sembly yields 21,111 active nodes (415 sources, 294 maladies, 129 modern diseases, 18,012 compounds, 2,211 targets, 50 preparations) and 52,467 active edges (32,909ISEXTRACTEDFROM, 16,696TARGETS, 1,841 RELATESTO, 431TREATSTRADITIONALLY, 257 MAPSTO, 301KNOWNTREATS, 32PREPAREDAS). A visual example of nodes originating from a single extracted source is presented in Figure S1. 3. Results 3.1. Knowledge-graph ablation: edge perturbation tests structural dependence First, to verify that DeepRoot Discovery genuinely relies on graph structure, we progressively shuffled the graph edges and tasked the critic agent with evaluating 30 extracted closed-loop sourceâmalady claims. As expected, the Criticâs self-confidence in the therapeutic plausibility of the source based on the text decreases as edge perturbation increases, demonstrating responsiveness to the KGâs integrity (Fig- ure 2A). Around50%perturbation, the criticâs confidence converges with the raw LLM baseline, suggesting that the KG signal has been degraded enough that the critic behaves similarly to an LLM without structured graph support. Fur- thermore, past50%, the score continues to decrease to 0.30, reflecting KG-dependent scoring. 3.2. KG-guided recovery of mechanistically supported candidates Next, we tested whether DeepRoot can use the KG to re- cover mechanistically grounded candidates from noisy his- torical text. For this, we synthesized evaluation cases by selecting sets of 3 closed-loop and 7 non-closed-loop dis- tractor sources. The associated paragraphs of those sources were then interweaved into a mini-corpus and fed to dif- ferent models to rank the sources and candidate chemical compounds (Figure 2B). We report source recall@3, com- pound recall@10 (the fraction of closed-loop compounds recovered within the top-10 candidates), and mean self- confidence (0â1) related to the therapeutic plausibility of candidate compounds. Because each passage may con- tain⌠500compounds, compound recall@10 directly tests whether KG-grounded scoring concentrates the likely leads. Over30mini-corpora, DeepRoot Discovery outperforms the LLM baseline, achieving1.95Ăhigher source recall and 6.11Ăhigher compound recall (Figure 2B). Surprisingly, despite DeepRoot Discovery being theoretically capable of fully traversing the KG, recovery was not perfect. This is likely due to two factors: framing evaluation as an inference 2 Submission and Formatting Instructions for ICML 2026 Figure 2. (A) Critic agent self-reported confidence for the therapeutic plausibility of each source-text pair vs. KG edge-shuffle fraction (n=30 source-text pairs). (B) Source-and-compound recovery of DeepRoot Discovery, LLM baseline (both using Gemini3.1 Flash Lite). *Batch evaluation by processing all30mini-corpora in a single invocation. Candidate discovery self-confidence refers to the mean self-reported confidence in each modelâs proposed compound candidates per mini-corpus (n=30 mini-corpora). task (Edwards & Camacho-Collados, 2024) and incomplete subgraph traversal, which could explain why batch evalu- ation across all mini-corpora improved all metrics since it could indirectly surface shared relevant paths for the Critic. Nevertheless, the recall and F1 gains validate that KG aug- mentation meaningfully enhances parsing and ranking. No- tably, the LLM baseline overstates therapeutic relevance, with a self-confidence of0.87, versus DeepRoot Discov- eryâs0.48, which closely aligns with the latterâs source recall@3of0.41(Figure 2A). This alignment suggests that self-confidence in a KG-augmented system is effectively bounded by retrieval accuracy. In contrast, the other modali- ties demonstrated high self-confidence hallucinations, which is a phenomenon previously reported for both LLMs and agents (Lin et al., 2022; 2025). Together with the KG abla- tion study, we establish that KG contributes meaningfully to the reasoning capabilities of an underlying LLM. 3.3. Blind rediscovery of held-out validated treatments Whereas Section 3.2 tested whether DeepRoot surfaces mechanistically grounded candidatesâcompounds for which the graph itself closes a compoundâtargetâdisease loopâthis experiment tests whether the system can blind- rediscover empirically validated compoundâdisease treat- ments after we hide them. Concretely, for each held-out pair (aKNOWNTREATSedge sourced from ChEMBL clinical indications) we delete the edge and all stereochemical sib- lings (planar-InChIKey prefix) from the validator, then ask DeepRoot Discovery to re-rank candidates for the disease. We evaluate on a21-pair historical set, and compare against a raw-corpus LLM given the full Shen Nong Ben Cao Jing and asked to rank the same top-K (Table 1). DeepRoot Discovery recovers 10 of 21 held-out pairs (R@20 = 47.6%), compared with the LLM baseline (1 of 21, R@20 = 4.8%). Per-disease candidate pools span87â 1,954compounds (median835), so randomR@20â 2.4%, suggesting that the result is far above random retrieval. Table 1. Held-outKNOWNTREATSrecovery on21historically reachable ChEMBL indication pairs. R@k in %, MRR unitless. MethodR@1 R@5 R@10 R@20 MRR DeepRoot Discovery9.528.633.347.60.161 Raw-corpus LLM0.04.84.84.80.012 3.4. Benchmarking DeepRootâs therapeutic reasoning against diverse baselines We audit critic-agent outputs with an independent LLM judge (Claude Sonnet 4.6, cross-family from the graded systems) on30stratified sourceâmalady claims across seven conditions (Table 2): the DeepRoot Discovery at three LLM tiers (Gemini 3.1 Pro / 2.5 Flash / 3.1 Flash Lite), a graph- only baseline (no LLM), an LLM-only baseline given corpus passages, and a tool-call LLM with direct access to the same APIs (ChEMBL, Open Targets, PubMed, MeSH) that DeepRoot Assembly itself queries. The judge scores six dimensions on[1, 5]and flags hallucinated evidence per claim. All three KG-augmented configurations outperform every baseline on overall score. Even DeepRootâLite (3.70) ex- ceeds both the graph-only condition (3.55) and the tool- calling LLM (2.47). This contrast highlights a tradeoff between grounding and synthesis. The tool-calling agent triggers the judgeâs hallucinated-evidence flag on87%of 3 Submission and Formatting Instructions for ICML 2026 Table 2. Reasoning-quality evaluation: seven conditions graded by Claude Sonnet 4.6 over a stratified sample of30sourceâmalady claims. Scores are means on[1, 5]; Hallu. is the rate of the judgeâshallucinatedevidenceflag in[0, 1]. Bold = best per column. SystemComponentsOverallâ EFâ VAâ RCâ CMâ UCâ ACâ Hallu.â DeepRoot â Gemini 3.1 Prograph + LLM3.834.53 4.47 3.97 4.07 3.73 3.670.10 DeepRoot â Gemini 2.5 Flashgraph + LLM3.774.67 4.37 3.83 3.63 3.57 3.630.07 DeepRoot â Gemini 3.1 Flash Lite graph + LLM3.704.60 4.27 3.73 3.70 3.60 3.670.07 Graph-onlygraph, no LLM3.554.55 4.55 2.69 3.21 3.31 2.930.00 Text + LLM (G3.1 FL)corpus + LLM3.173.10 2.80 3.47 3.67 3.27 3.170.13 Tool-call + LLM (G3.1 FL)ChEMBL/OT/PubMed/MeSH2.472.30 2.97 2.80 3.30 2.63 2.700.87 EF: evidence fidelity; VA: verdict alignment; RC: reasoning coherence; CM: clinical mapping; UC: uncertainty calibration; AC: actionability; Hallu.: hallucination rate. claims, despite having on-the-fly access to the same set of APIs. By contrast, the graph-only condition produces no hal- lucinated evidence by construction, but exhibits the weakest reasoning coherence (2.69). KG-augmented LLMs there- fore occupy a favorable middle ground: they maintain low hallucination rates (7â10%) while preserving the reasoning and synthesis capacity absent from graph-only scoring. 3.5. Human-expert evaluation of DeepRoot reasoning as a qualitative case study To evaluate the reasoning quality of DeepRoot Discovery as a traditional medicine knowledge assistant, we constructed a mini-corpus of 50 randomly sampled sourceâmalady pairs from Shen Nong Ben Cao Jing, each paired with its associ- ated textual evidence. This setting reflects a potential use case in which a scientist seeks to assess whether observed historical claims about a sourceâs therapeutic potential are grounded in modern biological evidence. Table 3. Classified modern disease agreement and verdict agree- ment with reference to DeepRoot. Verdict refers to the system- interpreted therapeutic plausibility of a sourceâmalady pair. SystemDA VA PR PB WU DeepRoot â (G3.1 FL)â33944 Biomni (w/ full agent env.)4230032 126 Text + LLM (G3.1 FL)3313111 19 19 DA: disease agreement (% of sourceâmalady pairs mapped to the same modern disease as DeepRoot); VA: verdict agreement (% of pairs assigned the same verdict as DeepRoot); PR: Previously Reported; PB: Plausible or Better (Previously Reported, Very Plausible, or Plausible); W: Weak; U: Unsupported. Comparing DeepRoot with Biomni (a biomedical reason- ing agent environment) (Huang et al., 2025), we find good alignment of disease classification and verdicts (Table 3). We provide full responses from DeepRoot and Biomni Ap- pendix A.10. Overall, DeepRoot Discovery clearly lever- ages the KG for reasoning. Biomni also shows strong reason- ing and tool-calling capabilities. Notably, both DeepRoot and Biomni cite the same biological targets for their reason- ing in example 1. But example 8 highlights a key limita- tion of KG over-reliance. Although DeepRootâs underlying LLM identified that the hydrolyzed version of certain com- pounds were bioactive, those downstream products were absent from the KG, preventing KG path completion and leading to low confidence despite biological plausibility. 4. Discussion and Conclusion DeepRoot shows that historical materia medica can be con- verted from pre-ontological prose into an auditable biomed- ical knowledge graph that supports mechanistic therapeu- tic reasoning. On the Shen Nong Ben Cao Jing, this con- struction pass enables held-out treatment recovery and sub- stantially lower hallucinated-evidence rates than LLM-only, tool-calling, and biomedical-agent baselines; with broader historical corpora, the same framework could support larger- scale drug repurposing, de novo therapeutic candidate nomi- nation, and prioritization of experimentally testable natural- product hypotheses. It is also important to highlight the gap between DeepRoot (LLM + KG) and LLMs given direct access to the same biomedical APIs. Our results suggest that building a veri- fied knowledge graph suppresses hallucination in a way that querying those resources at inference time does not. We show that building a verified knowledge graph suppresses hallucination in a way where querying those same APIs at inference time does not. For corpora that predate modern ontologies, retrieval-augmented and tool-using agents need a construction pass first, rather than on-demand calling. The same pattern may transfer to other historical materia med- ica, including Ayurvedic and broader ethnopharmacological archives, as well as structured ranking problems beyond traditional medicine. Furthermore, DeepRoot Assembly is a one-time invocation costing⌠$0.25/corpus and avoid- ing the recurring expense of on-demand database retrieval. In these settings, an agentically constructed KG offers the additional advantage that new curated claims, user submis- sions, and external evidence can be incorporated over time, expanding coverage while preserving traceability for future candidate ranking. 4 Submission and Formatting Instructions for ICML 2026 Limitations. (i) Corpus: a single71-chunk materia medica; transfer to other historical corpora is unverified. (i) Sample size: the held-out slice isN = 21pairs, single seed, no bootstrap CIs. (i) Priors: flat, face-validity, un- calibrated (Appendix A.5). (iv) Coverage: Open Targets is human-disease-only, leaving non-modern indications un- scored. (v) DeepRoot Discovery reasoning: rationale quality is bounded by the underlying LLM. (vi) Comparasion evalu- tions with other LLM modalities: While we reported the best result for LLM or biomni from prompt engineering, limited analysis was placed in this area, but previous studies have demonstrated that such interventions seldom provide sub- stantial improvements (Qian et al., 2024; Wu et al., 2024). Furthermore, we tried evaluating Biomni/Phylo within the LLM judge suite, but its access to recently updated papers and external validation evidence may exceed the judgesâ closed-book biomedical knowledge, particularly for claims grounded in post-cutoff literature. Impact Statement DeepRoot is a research tool for hypothesis generation, not medical advice. By converting historical materia medica into auditable sourceâcompoundâtargetâdisease chains, it may help researchers prioritize natural-product candidates for experimental follow-up and drug repurposing. The main risks are overinterpretation, unsafe self- medication, and misuse of traditional knowledge. Histori- cal claims may be ineffective, toxic, or culturally specific, and graph-supported plausibility does not establish safety or efficacy. Any downstream use requires expert review, provenance tracking, toxicity assessment, and experimental validation. Data Availability The Shen Nong BenCao Jing corpus and code for generating the KG and traversing it are provided at github.com/CarlisleMa/deeprootv1. Provided within are also the scripts and raw data generated for the evaluations. References Brook, K., Bennett, J., and Desai, S. P.The chemi- cal history of morphine: An 8000-year journey, from resin to de-novo synthesis. Journal of Anesthesia His- tory, 3(2):50â55, 2017. doi: 10.1016/j.janh.2017.02. 001. URLhttps://w.sciencedirect.com/ science/article/pii/S2352452916301293. Chandrasekhar, V., Rajan, K., Kanakam, S. R. S., Sharma, N., WeiĂenborn, V., Schaub, J., and Stein- beck, C.COCONUT 2.0: A comprehensive over- haul and curation of the collection of open nat- ural products database.Nucleic Acids Research, 53(D1):D634âD643,2025.doi:10.1093/nar/ gkae1063. URLhttps://academic.oup.com/ nar/article/53/D1/D634/7908792. Published online 2024. Dai, Y., Shao, X., Zhang, J., Chen, Y., Chen, Q., Liao, J., Chi, F., Zhang, J., and Fan, X.TCMChat: A generative large language model for traditional Chi- nese medicine.Pharmacological Research, 210: 107530, 2024.doi:10.1016/j.phrs.2024.107530. URLhttps://w.sciencedirect.com/ science/article/pii/S1043661824004754. Edwards, A. and Camacho-Collados, J. Language models for text classification: Is in-context learning enough? In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), p. 10058â10072, Torino, Italia, 2024. ELRA and ICCL. URLhttps:// aclanthology.org/2024.lrec-main.879/. Ghafarollahi, A. and Buehler, M. J.SciAgents: Au- tomating scientific discovery through bioinspired multi- agent intelligent graph reasoning.Advanced Mate- rials, 37(22):2413523, 2025.doi: 10.1002/adma. 202413523. URLhttps://doi.org/10.1002/ adma.202413523. Published online 2024. He, J., Guo, Y., Lam, L. K., Leung, W., He, L., Jiang, Y., Wang, C. C., Xing, G., and Chen, H. Opentcm: A graphrag-empowered llm-based system for traditional chi- nese medicine knowledge retrieval and diagnosis, 2025. URL https://arxiv.org/abs/2504.20118. Huang, K., Zhang, S., Wang, H., Qu, Y., Lu, Y., Roohani, Y., Li, R., Qiu, L., Li, G., Zhang, J., Yin, D., Marwaha, S., Carter, J. N., Zhou, X., Wheeler, M., Bernstein, J. A., Wang, M., He, P., Zhou, J., Snyder, M., Cong, L., Regev, A., and Leskovec, J. Biomni: A general-purpose biomedical AI agent.bioRxiv, June 2025. doi: 10.1101/2025.05.30.656746. URL https://w.biorxiv.org/content/10. 1101/2025.05.30.656746v1. Preprint. Hui, Y., Du, L., Lin, S., Qu, Y., and Cao, D. Extrac- tion and classification of TCM medical records based on BERT and Bi-LSTM with attention mechanism. In 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), p. 1626â1631, 2020. doi: 10.1109/BIBM49941.2020.9313359. URLhttps:// ieeexplore.ieee.org/document/9313359. Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., Thiessen, P. A., Yu, B., Zaslavsky, L., Zhang, J., and Bolton, 5 Submission and Formatting Instructions for ICML 2026 E. E.PubChem 2023 update.Nucleic Acids Re- search, 51(D1):D1373âD1380, 2023. doi: 10.1093/ nar/gkac956. URLhttps://academic.oup.com/ nar/article/51/D1/D1373/6777787. Koehn, F. E.Biosynthetic medicinal chemistry of natural product drugs.MedChemComm, 3(8): 854â865,2012.doi:10.1039/C2MD00316C. URLhttps://pubs.rsc.org/en/content/ articlelanding/2012/md/c2md00316c. Li, Y., Luan, Z., Liu, Y., Liu, H., Qi, J., and Han, D.Automated information extraction model en- hancing traditional Chinese medicine RCT evidence extraction (Evi-BERT): Algorithm development and validation.Frontiers in Artificial Intelligence, 7: 1454945, 2024. doi: 10.3389/frai.2024.1454945. URL https://w.frontiersin.org/journals/ artificial-intelligence/articles/10. 3389/frai.2024.1454945/full. Lin, S., Hilton, J., and Evans, O. TruthfulQA: Measur- ing how models mimic human falsehoods. In Proceed- ings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 3214â3252, Dublin, Ireland, 2022. Association for Com- putational Linguistics. doi: 10.18653/v1/2022.acl-long. 229. URLhttps://aclanthology.org/2022. acl-long.229/. Lin, X., Ning, Y., Zhang, J., Dong, Y., Liu, Y., Wu, Y., Qi, X., Sun, N., Shang, Y., Wang, K., Cao, P., Wang, Q., Zou, L., Chen, X., Zhou, C., Wu, J., Zhang, P., Wen, Q., Pan, S., Wang, B., Cao, Y., Chen, K., Hu, S., and Guo, L. LLM-based agents suffer from hal- lucinations: A survey of taxonomy, methods, and di- rections. arXiv preprint arXiv:2509.18970, 2025. doi: 10.48550/arXiv.2509.18970. URLhttps://arxiv. org/abs/2509.18970. Liu, Y., Yuan, Y., Yan, K., Li, Y., Sacca, V., Hodges, S., Cannistra, M., Jeong, P., Wu, J., and Kong, J. Eval- uating the role of large language models in traditional Chinese medicine diagnosis and treatment recommenda- tions. npj Digital Medicine, 8:466, 2025. doi: 10.1038/ s41746-025-01845-2. URLhttps://w.nature. com/articles/s41746-025-01845-2. McLaughlin, J., Lagrimas, J., Iqbal, H., Parkinson, H., and Harmse, H.OLS4: A new ontology lookup service for a growing interdisciplinary knowledge ecosystem.Bioinformatics, 41(5):btaf279, 2025. doi: 10.1093/bioinformatics/btaf279.URLhttps: //academic.oup.com/bioinformatics/ article/41/5/btaf279/8125017. Mendez, D., Gaulton, A., Bento, A. P., Chambers, J., De Veij, M., F Ì elix, E., Magari Ì nos, M. P., Mosquera, J. F., Mutowo, P., Nowotka, M., Gordillo-Mara Ì n Ì on, M., Hunter, F., Junco, L., Mugumbate, G., Rodriguez-Lopez, M., Atkinson, F., Bosc, N., Radoux, C. J., Segura- Cabrera, A., Hersey, A., and Leach, A. R. ChEMBL: Towards direct deposition of bioassay data. Nucleic Acids Research, 47(D1):D930âD940, 2019. doi: 10.1093/ nar/gky1075. URLhttps://academic.oup.com/ nar/article/47/D1/D930/5162468. Newman, D. J. and Cragg, G. M. Natural products as sources of new drugs over the nearly four decades from 01/1981 to 09/2019. Journal of Natural Prod- ucts, 83(3):770â803, 2020. doi: 10.1021/acs.jnatprod. 9b01285. URLhttps://pubs.acs.org/doi/10. 1021/acs.jnatprod.9b01285. Ochoa, D., Hercules, A., Carmona, M., Suveges, D., Gonzalez-Uriarte, A., Malangone, C., Miranda, A., Fu- mis, L., Carvalho-Silva, D., Spitzer, M., Baker, J., Ferrer, J., Raies, A., Razuvayevskaya, O., Faulcon- bridge, A., Petsalaki, E., Mutowo, P., Machlitt-Northen, S., Peat, G., McAuley, E., Ong, C. K., Mountjoy, E., Ghoussaini, M., Pierleoni, A., Papa, E., Pignatelli, M., Koscielny, G., Karim, M., Schwartzentruber, J., Hulcoop, D. G., Dunham, I., and McDonagh, E. M. Open Targets platform: Supporting systematic drug- target identification and prioritisation. Nucleic Acids Re- search, 49(D1):D1302âD1310, 2021. doi: 10.1093/nar/ gkaa1027. URLhttps://academic.oup.com/ nar/article/49/D1/D1302/6024045. Qian, C., Liu, W., Liu, H., Chen, N., Dang, Y., Li, J., Yang, C., Chen, W., Su, Y., Cong, X., Xu, J., Li, D., Liu, Z., and Sun, M. Chatdev: Communicative agents for software de- velopment, 2024. URLhttps://arxiv.org/abs/ 2307.07924. Rasmussen, P., Paliychuk, P., Beauvais, T., Ryan, J., and Chalef, D. Zep: A temporal knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956, 2025. doi: 10.48550/arXiv.2501.13956. URLhttps: //arxiv.org/abs/2501.13956. Schoch, C. L., Ciufo, S., Domrachev, M., Hotton, C. L., Kannan, S., Khovanskaya, R., Leipe, D., Mcveigh, R., OâNeill, K., Robbertse, B., Sharma, S., Soussov, V., Sullivan, J. P., Sun, L., Turner, S., and Karsch-Mizrachi, I.NCBI taxonomy: A comprehensive update on curation, resources and tools. Database, 2020:baaa062, 2020. doi: 10.1093/database/baaa062. URLhttps: //academic.oup.com/database/article/ doi/10.1093/database/baaa062/5881509. Tu, Y. The discovery of artemisinin (Qinghaosu) and gifts from Chinese medicine. Nature Medicine, 17(10):1217â 6 Submission and Formatting Instructions for ICML 2026 1220, 2011. doi: 10.1038/nm.2471. URLhttps:// w.nature.com/articles/nm.2471. Wu, Y., Yue, T., Zhang, S., Wang, C., and Wu, Q. State- flow: Enhancing llm task-solving through state-driven workflows, 2024. URLhttps://arxiv.org/abs/ 2403.11322. 7 Submission and Formatting Instructions for ICML 2026 A. Technical appendices and supplementary material A.1. Knowledge graph schema Edge types. Seven typed edges, each carrying a numericconfidencescore(flat prior, see Appendix A.5), an evidencetype tag (see Appendix A.6), and a sourcedb provenance field where applicable. âą TREATSTRADITIONALLY (Sourceâ TraditionalMalady): evidence span quoted from the source chunk text. âą MAPSTO(TraditionalMaladyâModernDisease):isprimary,mappingrole â primary, syn- dromecomponent,mappingsourceâgemini+icd10exact, gemini+meshexact, gemini+snomedexact, gem- iniunverified, mappingalternatives (JSON). âą ISEXTRACTEDFROM(ChemicalCompoundâSource): evidence type encodes COCONUT/PubChem provenance and resolution level (canonical vs. alias vs. formula). âą TARGETS(ChemicalCompoundâBiologicalTarget):pchemblscore,assayid,assaytype â B, F, assaydescription, mechanismaction. âą RELATES TO(BiologicalTargetâModernDisease):otoverallscore,matchtierâefoid, mondoid, doidid, meshid, normname. âą KNOWNTREATS(ChemicalCompoundâModernDisease):clinicalphaseâ1, 2, 3, 4, materialized from ChEMBL drug indications; held-out evaluation slice (§3.3). âą PREPAREDAS (Sourceâ PreparationMethod). Identity and write semantics.All writes are idempotentMERGE-on-identity. Compound identity is the RDKit-computed InChIKey, which is invariant to canonical-SMILES variants and to naming differences across COCONUT and PubChem. Target identity is the ChEMBL ID, which unifies SINGLE PROTEIN, PROTEIN COMPLEX, PROTEIN FAMILY, and ORGANISM target types under one key. Modern disease identity is the canonical name, with ontology codes coalesce- backfilled as later agents verify them against additional services. Figure S1. Example of a node cluster in Neo4j. 8 Submission and Formatting Instructions for ICML 2026 Table S1. Complete node schema for the DeepRoot knowledge graph. NodePropertyDescription Source nameCanonical Latin binomial or common name (primary key) aliasesAlternative names from source text evidencespanVerbatim passage from which the node was extracted sourcedocumentOrigin corpus file identifier canonicalnameAuditor-resolved canonical name canonical typeTaxonomic category (herb, mineral, animal, fungus, . . . ) canonicalpartPlant/animal part used (root, bark, seed, whole, . . . ) canonicalsourceDatabase used to resolve canonical form canonicalrawresponseRaw LLM response from canonicalization step linkerstatusCompound-linker outcome (ok, skipped, failed) linkerattemptedatISO timestamp of last linker run linkercompoundcountNumber of compounds linked from this source linkerevidencetypeEvidence type used (coconut, chembl) ChemicalCompound nameIUPAC or common compound name (primary key) smilesCanonical SMILES string inchikeyStandard InChIKey identifier molecular formulaMolecular formula (e.g. C21H23NO4) nplikenessNatural-product likeness score from COCONUT (â5to +5) annotation levelStructuralconfidencetier(1 = MS2confirmed, 5 = predicted) sourcedbSource database (COCONUT, ChEMBL, . . . ) coconut rowCOCONUT row index for traceability pubchemcidPubChem Compound ID targetlinkerstatusChEMBL target-linker outcome linkerattemptedatISO timestamp of target-linker run linker chemblidChEMBL molecule ID used for target lookup linkerlookupmethodMatch method (inchikey, smiles, name) linkertargetcountTargets written after filtering linkerdroppedcountTargets dropped below pChEMBL floor linkerpchemblfloorpChEMBL activity threshold applied linker maxtargetsCap on targets written per compound ktlinkerstatus KNOWNTREATS linker outcome ktlinkerattemptedatISO timestamp of KNOWNTREATS linker run ktlinkerindicationcountDrug indications written kt linkerdroppedcountIndications dropped below phase threshold ktlinkerminphaseMinimum clinical trial phase accepted BiologicalTarget nameTarget protein name (primary key) target prefnameChEMBL preferred target name genesymbolHGNC gene symbol uniprotidUniProt accession target chemblidChEMBL target identifier targettypeTarget class (SINGLE PROTEIN, PROTEIN COMPLEX, . . . ) ncbitaxidNCBI taxonomy ID of the target organism tdlinkerstatusTarget-disease linker outcome td linkerattemptedatISO timestamp of target-disease linker run tdlinkerassociationcount Disease associations written td linkerdroppedcountAssociations dropped below score threshold td linkerminscoreOpen Targets association score floor applied Continued on next page 9 Submission and Formatting Instructions for ICML 2026 Table S1 (continued) NodePropertyDescription ModernDisease nameDisease name (primary key) doididDisease Ontology identifier mondoidMONDO disease ontology identifier meshidMeSH descriptor efoidExperimental Factor Ontology (EFO) identifier icd10 codeICD-10 classification code snomedidSNOMED CT concept identifier verifiedbyAgent or curator that confirmed the mapping TraditionalMalady nameTCM ailment name (primary key) descriptionClassical definition from source text evidencespanVerbatim passage supporting the malady sourcedocumentOrigin corpus file identifier mapperstatusMalady-to-disease mapper outcome mapperclassificationOntology mapping confidence class mapperattemptedatISO timestamp of mapper run mapper rawresponseRaw LLM response from mapping step archivedReviewer flag: duplicate or low-quality node archivereasonFree-text reason for archiving reviewedbyReviewer agent identifier PreparationMethod namePreparation name (decoction, pill, powder, . . . ) routeAdministration route (oral, topical, inhaled, . . . ) evidence spanVerbatim passage describing the preparation A.2. Assembly agent protocols Table S2. The seven DeepRoot Assembly agents in dependency order. Each agent combines LLM proposal with deterministic verification against the listed grounding source. # AgentRoleGrounding source i Extractionemit Source / Malady / Preparation; TREATS, PREPAREDAS â i Auditor canonicalize; verify evidence spans; merge duplicates COCONUT, Pub- Chem i MaladyâDiseasegenerate-then-verify ontology map- ping (MAPSTO) MeSH, ICD-10, SNOMED iv SourceâCompoundnatural-product / chemical lookup (ISEXTRACTEDFROM) COCONUT, Pub- Chem v CompoundâTargetmolecular targets and bioactivities (TARGETS) ChEMBL vi TargetâDiseasemechanism-to-disease, dispatched on target type (RELATESTO) OpenTargets, NCBI Tax, OLS4 vii Reviewerrules + LLM archival pass (orphans, OCR artifacts, off-domain entities) â 10 Submission and Formatting Instructions for ICML 2026 Extraction.A Gemini call per text chunk emitsSourcenodes with aliases and species hints,TraditionalMalady nodes with descriptions,PreparationMethodnodes, and theTREATSTRADITIONALLYandPREPAREDASedges among them. Each edge records the literal evidence span from the source chunk. Confidence is the LLMâs self-assessed score, used only as a soft signal for downstream auditing (the auditor verifies span and identity independently). Auditor. Three deterministic post-extraction jobs. (1) Canonicalization. Gemini Flash-Lite (temperature 0, batched 20 sources/call, structured-JSON schema) labels each sourceorganism(Latin binomial + part),chemical(name or formula), oruncanonicalized. Per-type external lookup verifies: organisms hit a local COCONUT inverted index over 62,792species keys, chemicals hit PubChem REST. Canonical labels with external-DB hits are taggedgemini+coconut /gemini+pubchem; the rest are taggedgeminiunverified * ,uncanonicalized, orerror. (2) Source merge. Sources sharing the same(canonicalname,canonicalpart)collapse onto a keeper (highest-degree, alphabetical tiebreak). Merged-from nodes are soft-archived with reasonmergedinto:<keeper>; outgoing edges are re-targeted (parallel edges take maximum confidence) and aliases are unioned. (3) Evidence-span verification. Substring check (whitespace-normalized) of everyTREATSTRADITIONALLYevidence span against its source chunk. Hallucinated spans (typically LLM-introduced ellipsis) trigger soft-archival with reason hallucinatedevidence. MaladyâDisease (generate-then-verify). One Gemini call per malady (temperature 0, six-shot system prompt) emits a typed exit:disease,symptom,syndrome,ambiguous, ortcmnoequivalent. Crucially, the LLM emits canonical names onlyânever codesâtogether with an ontology hint. Each proposed name is then verified in parallel against ICD-10 (NLM Clinical Tables), MeSH (NLM RDF Lookup), and SNOMED (EBI OLS4). Verification accepts only tolerant exact match (case- and punctuation-insensitive); fuzzy hits, even from the API itself, are rejected. Syndromes can produce one primary plus up to twosyndromecomponentedges, but only components that pass exact-match verification are written (unverified components are dropped and logged in the primary edgeâsmappingalternatives). Default mode rejects unverified mappings entirely; an--allow-unverifiedflag stores them withrequiresreview=trueand a degraded prior. SourceâCompound. Routes bycanonicaltype.Organisms hit a local COCONUT inverted index (in- memory,âŒ700,000structures with species provenance) by exact normalized species name and the âfirst two to- kensâ (genus species) prefix. Chemicals hit a disk-cached PubChem REST client (4 RPS), with formula fallback via compound/fastformulafor formula-shaped canonicals. Compound identity is the RDKit-computed InChIKey from canonical SMILES, which collapses canonicalization variants and unifies COCONUT/PubChem name splits onto a sin- gle node. Edges are written withevidencetypeencoding the resolution path (coconutorganismcanonical, pubchemchemicalcanonical, alias, formula, unverified). Per-sourcelinkerstatusstamping makes the agent fully resumable across network interruptions. CompoundâTarget.ChEMBL queries by InChIKey (91.4%resolution rate), falling back to canonical SMILES then pre- ferred name. For each resolved compound, the agent retrieves mechanism-of-action records and bioactivity records. Activity filtering:pchemblvalueâ„ 5.0(10ÎŒM floor) for quantitative tier,assaytypeâB,F,standardrelation â=,⌠; data validity flags must be empty. Target type is not restricted to SINGLE PROTEIN: PROTEIN COMPLEX (subunit fan-out), PROTEIN FAMILY (broad-spectrum inhibitors), and ORGANISM (anti-pathogen evidence, e.g. Plasmod- ium falciparum for antimalarials) are all admitted. A--include-phenotypicflag additionally retrieves phenotypic activities (pchemblvaluenull,assaytypeF/B) at a lower confidence prior; for terpenes, sterols, and other natural products tested phenotypically rather than against named molecular targets, this opt-in is necessary to avoid silent loss of âŒ2,000compounds. Salt and tautomer parents are aggregated via the ChEMBL molecule hierarchy; without hierarchy expansion, 30â50% of activities are missed for multi-form compounds. TargetâDisease. Four-way dispatch ontargettype.SINGLE PROTEIN: Open Targets GraphQL via UniProtâEnsembl, returning disease associations with overall scores binned to confidence tiers. PROTEIN COMPLEX: subunit fan-out via ChEMBL, then per-subunit Open Targets, with per-disease max-score deduplication. PROTEIN FAMILY: intentionally skipped (tdlinkerstatus = skippedproteinfamily), as family-level evidence is too coarse for clinical association. ORGANISM: an EFO/DOID walk over OLS4 starting from the NCBItaxid, emitting both the specific disease class and its ancestors, falling back to a parallelized LLM safety net (Gemini proposes candidate disease names, NLM MeSH verifies via tolerant exact match) when the ontology walk returns empty. Three-tier disease matching: exact ontology-ID match first, then normalized-name exact match, then MeSH-synonym expansion. A plan-then-apply 11 Submission and Formatting Instructions for ICML 2026 phase performs all writes viaUNWIND-batched Cypher transactions (âŒ12transactions forâŒ5,500rows), withDELETE restricted to terminal-status rows so transient API failures cannot wipe live edges. Reviewer.Two-pass deterministic-then-LLM archival. Pass 1 catches OCR artifacts (single-character entities, mojibake), orphans (degree0), generic categories (âherbâ, âcompoundâ), and metaphysical concepts that escaped extraction. Pass 2 batches the residual ambiguous nodes (âŒ20) to Gemini for biomedical-relevance filtering. Cascade archival propagates to incident edges. Archival is soft (archived=true with reason); no records are deleted. A.3. Discovery agent protocols DeepRoot Discovery comprises two agent roles operating over the typed graph: a critic agent that scores existing SourceâMalady claims, and a discovery agent that nominates novel compound candidates for a target ModernDisease. Both consume the tier-bucketed path-scoring layer described in Appendix A.4. Critic agent. For each (Source, Malady) claim, the agent receives a structured payload assembled from the KG: the claim itself (Source name + aliases, Malady description, primary mapped ModernDisease, mapping rationale), deter- ministic Pass-1 signals (path bucket distribution, loop-closure counts, top-bucket score), the top-Kmechanistic chains (SourceâCompoundâTargetâDisease) with edge metadata, and four cross-cutting enrichmentsâcompound profiles (KNOWNTREATS for other diseases, target spectrum), target genericity (number of associated diseases per target), source- level target convergence (multi-compound hits on the same target), and sibling verdicts (Pass-1 verdicts of other claims on the same source). The model returns a structured JSONCriticVerdictwith: a verdict on the four-tier ladder (VALIDATED / PLAUSIBLE / WEAK / UNSUPPORTED); biologicalplausibility and evidencecoherence scores in[0, 1], defen- sively clamped to that range; akeyevidencelist of cited compoundâtargetâ disease triples; aconcernslist with typed enum values (generictarget,weakevidenceonly,indirectmechanism,wrongdiseasemapping, syndromeunderutilized,promiscuouscompound,unverifiedevidence); a free-form rationale; and a requireshumanreviewflag (auto-set when the LLM and deterministic Pass-1 verdicts disagree byâ„ 2rungs). The prompt instructs the model to quote specific input fields and never speculate beyond the provided evidence; numeric ranges are enforced via post-hoc clamping rather than relying on the model to obey them. Discovery (nominator) agent. Given a target ModernDisease queryd â , the agent walks the KG backward (disease âmaladyâsourceâcompound) to enumerate all corpus-supported candidate compounds, then walks forward (com- poundâtargetâdisease) to score each candidateâs mechanistic plausibility. A novelty filter drops compounds whose KNOWNTREATSedge already reachesd â (supporting an in-memory mask for held-out evaluation without mutating the graph). The remaining candidates are ranked lexicographically by (i)hasloopclosure(does at least one for- ward chain reachd â ), (i)forwardbucket(T1>T2>T3>T4; weakest-link tier of the strongest loop-closing path), (i)uniquesourcescount, (iv)uniquemaladiescount, and (v)forwardmaxscore(multiplica- tive product of edge confidences along the strongest path). The output is an ordered list ofCompoundCardentries containing top historical paths (source, malady, evidence span), top forward paths (target, assay description, OT score), and KNOWNTREATS for other diseases as polypharmacology context. The discovery agent is fully deterministicâno LLM is in the loopâmaking the ranking auditable and stable across re-runs. A.4. Tier-bucket path scoring A path is a sequence of typed edges connecting a source node to a disease node through compound and target in- termediaries. Each edge carries anevidencetypetag (e.g.,chemblmechanism,otassociationstrong, coconutorganismcanonical) which maps to one of four tiersT âT1, T2, T3, T4(Appendix A.5, Table S3) and a flat numeric prior câ [0, 1]. For a path p with edges e 1 ,...,e n , the path bucket B(p) and path score S(p) are B(p) = min i=1...n T(e i ), S(p) = Y i=1...n c(e i ). Paths are ordered lexicographically by(B(p),S(p))with the bucket as the primary key (highest-tier bucket wins) and the multiplicative score as tiebreak within a bucket. The bucket captures the qualitative claim âa chain is only as strong as its weakest edgeâ (weakest-link), while the score gives a continuous ordering inside each tier. 12 Submission and Formatting Instructions for ICML 2026 The same scoring layer is consumed by both Discovery agents (Appendix A.3) and by the deterministic Pass-1 signals fed to the critic. Because priors are flat (Table S3) rather than learned, raw external scores (otoverallscore, pchemblvalue,nplikeness) are preserved as edge attributes so downstream consumers can recalibrate without re-running Assembly. A.5. Confidence priors Tier ladder (used for path scoring).T1>T2>T3>T4. Path bucket is the minimum tier across edges (weakest-link); within a bucket, ranking uses the multiplicative product of edge confidences as tiebreak. Table S3. Per-edge-type confidence priors. Priors are flat (not learned), chosen on biomedical face validity, and never replaced by self-reported LLM confidence. Raw external scores (e.g., otoverallscore, pchemblvalue, nplikeness) are preserved as edge properties so downstream consumers can recalibrate without re-running Assembly. EdgeEvidence typeTier (prior) ISEXTRACTEDFROM coconutorganismcanonical / pubchemchemicalcanonical T1 (0.70â0.80) coconut organismaliasT2 (0.55) coconutorganismunverified / pubchemchemicalunverified T3 (0.50â0.55) pubchem chemicalformulaT4 (0.50) TARGETS chemblmechanismT1 (0.95) chembl activitystrong (pchemblâ„ 7)T2 (0.75) chemblactivitymoderate (pchemblâ„ 6)T3 (0.60) chemblactivityweak (pchemblâ„ 5) / chemblphenotypic T4 (0.40) RELATESTO ncbipathogenconsensusT1 (0.92) otassociationstrong (OTâ„ 0.7)T1 (0.85) otassociationmoderate (OTâ„ 0.4) / complex aggregate / pathogenllmverified T2 (0.65â0.75) ot associationweak (OTâ„ 0.2)T3 (0.45) MAPSTO icd10/mesh/snomed exact (primary)T1 (0.80â0.85) syndromecomponent / symptomT2 (0.65â0.75) KNOWNTREATS clinical phase = 4 (approved)T1 (0.95) clinical phase = 3T1 (0.85) clinical phase = 2T2 (0.65) clinical phase = 1T3 (0.45) A.6. Graph statistics Node and edge totals (active, post-Assembly).21,111nodes active,94archived.By type:415Source, 294TraditionalMalady,129ModernDisease,18,012ChemicalCompound,2,211BiologicalTarget,50Prepara- tionMethod. Edges:52,467active. By type:32,909 ISEXTRACTEDFROM,16,696 TARGETS,1,841 RELATESTO, 431 TREATSTRADITIONALLY,301 KNOWNTREATS,257 MAPSTO(208primary+49syndromecomponent),32 PREPARED AS. 13 Submission and Formatting Instructions for ICML 2026 Per-evidence-type breakdown.ISEXTRACTEDFROM:32,885organismcanonical,21chemicalcanonical,3formula. TARGETS:60mechanism,1,148strong,1,185moderate,2,264weak,12,039phenotypic.RELATESTO:936otweak, 666otmoderate,14otstrong,203complexaggregate,17llmverified,3pathogenconsensus,2efo.KNOWNTREATS:60 phase4,92phase3,84phase2,65phase1. The phenotypic-heavy distribution ofTARGETSreflects the natural-product corpus: terpenes, sterols, and flavonoids are predominantly characterized by phenotypic bioassays rather than named molecular targets. Convergence.4,605compounds appear inâ„ 2sources (classic phytomedicine pattern:ÎČ-sitosterol112Ă, quercetin87Ă, kaempferol 66Ă). Across the 129 ModernDisease nodes, 257 MAPSTO edges resolve to an average of 1.99 maladies per disease, indicating strong canonical convergence rather than fragmentation. Coverage.504of2,211targets (22.8%) link to at least one disease;88of129disease nodes (68%) are reached by at least one target.3,221of18,012compounds (17.9%) have at least oneTARGETSedge; the remainder either lack ChEMBL records or have no admissible target-class data, a documented limitation of the underlying databases rather than of the pipeline. A.7. Evaluation protocols Eval 1: edge-perturbation sensitivity.A fixed test set of closed-loop SourceâMalady claims is sampled from the KG. For each perturbation levelpâ0%, 20%, 40%, 60%, 80%, 100%, a fractionpof edges across four mechanistic edge types (TARGETS,RELATESTO,KNOWNTREATS,MAPSTO) is selected uniformly at random and their target endpoints are shuffled among themselves. A single perturbation is applied per level; all test claims are then evaluated by the Critic against the same perturbed graph. Self-confidence (mean biological plausibility over all claims) is reported per level. Eval2:sourceandcompoundrecoveryonmini-corpora. Tobuildeachmini-corpus, buildrecoveryevalcorpus.pyfirst queries the knowledge graph for two disjoint source pools: closed- loop sources (those with at least one completeSource â Compound â Target â Diseasechain where the source alsoTREATSTRADITIONALLYâ Maladyâ MAPSTOthe same disease) and distractor sources (all other non-archived KG sources). The full Shen Nong Ben Cao Jing text is split into paragraphs, each tagged by keyword regex against every source name in the KG. Gemini Flash then verifies which tagged sources a paragraph actually describes therapeutically (rather than merely cross-referencing in a compatibility list). Verified single-source paragraphs are banked into the two pools. Each synthetic eval case is assembled by a diversity-maximising greedy algorithm: it picks K=3least-used closed-loop source paragraphs andN=7least-used distractor source paragraphs, ensures no overlap between the two sets, then deterministically shuffles all 10 paragraphs into an interleaved mini-corpus. The label set for compound recovery includes both closed-loop compounds (retrieved via a Cypher walk confirmingCompound â TARGETS â Target â RELATESTO â Diseasefor a disease the source already treats; typically 1â5 per source) and distractor compounds (ISEXTRACTEDFROMcompounds of the distractor sources). Compound recall@kis reported separately against each label set so the closed-loop and broad-coverage signals can be distinguished. Thirty such mini-corpora are generated, each with a distinct closed-loop source signature enforced by deduplication. Eval 3: positive-control recovery of hidden known treatments. The 301KNOWNTREATSedges are filtered to those whose ModernDisease has a backward chaindâmaladyâsourceâcompound in the KG (the historical-reachability subset), yielding 21 (compound, disease) pairs across 10 diseases. For each test pair(c â ,d â ), we computec â âs planar InChIKey prefix (first 14 characters, dropping stereochemistry) and mask everyKNOWNTREATSedge from any compound sharing that prefix tod â (in-memory mask only; the graph is not mutated). The discovery agent is then run ond â with top-K = 20. A trial succeeds at rankrif any nominee within the toprsharesc â âs planar prefix. The candidate poolâall compounds reachable via the backward chain fromd â âis recorded per-disease (range 87â1,954, median 835), giving a uniform random recall@20 baseline ofâ 2.4%. Eval 4: LLM-as-judge reasoning quality. A stratified sample of 30 closed-loop SourceâMalady claims is drawn from the 431 candidate claims, with the strata chosen to exercise distinct verdict regimes (representative balance of unsupported, strongsupport, T1-bucket-without-loop-closure, mechanistic-only, and traditional-only). The same 30 claims are scored by six conditions: DeepRoot Discovery at three LLM tiers (Gemini 3.1 Pro, 2.5 Flash, 3.1 Flash- Lite), a graph-only deterministic baseline (Pass-1 verdict only, no LLM), an LLM baseline given just the corpus pas- 14 Submission and Formatting Instructions for ICML 2026 sages, and a tool-call LLM baseline given direct API access to ChEMBL, Open Targets, PubMed, and MeSH. Out- puts are graded by Claude Sonnet 4.6 (cross-family from the graded systems) on six dimensions in[1, 5]: Evidence Fidelity (does the critic cite evidence present in the payload?), Verdict Alignment (does the verdict follow from the visible evidence?), Reasoning Coherence (does the rationale explain this claimâs chain?), Clinical Mapping (does the critic responsibly handle the maladyâdisease mapping?), Uncertainty Calibration (are the scores, concerns, and review flag calibrated?), and Actionability (would a curator know what to inspect next?). The judge additionally returns six binary flags (hallucinatedevidence,unsupportedverdictjump,ignoredloopclosurestatus, overclaimsstrength,contradictoryscores,needshumanreview) and a recommended status in pass,weakpass,fail,humanreview. The judge sees only the criticâs visible artifacts (verdict, scores, keyevidence, concerns, rationale) plus, where applicable, the structured payload that the critic was given; it does not have access to ground truth and grades the quality of the criticâs argument rather than its absolute correctness. A.8. Prompt templates This section summarizes the four critical prompts in DeepRoot. Each template is described as a tuple of (role, input schema, output schema, key instructions); the verbatim text is in the released codebase. Extraction prompt.Role: extract typed entities from a single text chunk of the source corpus. Input: the chunk text plus asourcedocumentidentifier. Output schema (structured JSON):sources[],maladies[],preparations[], treatsedges[],preparedasedges[]; each entity carriesname,aliases,evidencespan(verbatim quote from the chunk), and a self-assessedconfidence. Key instructions: evidence spans must be substring-matchable to the chunk text; identifiers (binomials, ChEMBL IDs) are never to be invented; the LLM emits names, not codes. MaladyâDisease mapping prompt (generate-then-verify). Role: classify each TraditionalMalady into one of disease,symptom,syndrome,ambiguous,tcmnoequivalentand propose canonical English name(s). Input: the maladyâsname,description,evidencespan, and source classical context. Output schema: a top- levelclassificationfield plus amappings[]list, where each mapping carriesname(canonical English), an ontologyhintâicd10,mesh,snomed, aroleâprimary,syndromecomponent, and a one-sentence rationale. Key instructions: the model emits canonical names only, never codes; forsyndromemaladies, up to two syndromecomponententries may be returned in addition to the primary; six in-context examples cover the typical TCM patterns (wind heat, gu toxin, counterflow, etc.). Codes (ICD-10, MeSH, SNOMED) are recovered downstream by deterministic exact-match verification against the corresponding ontology servicesânever trusted from the LLM. Critic agent prompt. Role: biomedical reasoning expert evaluating whether a historical Source plausibly treats a ModernDisease via known mechanisms. Input: the structured payload described in Appendix A.3 (claim, Pass-1 signals, evidence paths, compound profiles, target profiles, source-level target convergence, sibling verdicts), serialized as a single JSON object. Output schema: theCriticVerdictobject described in Appendix A.3. Key instructions (highlighted in the system prompt): (i) assess TARGET QUALITY via the per-target disease count (pleiotropy ofâ„ 50associated diseases is flaggedgenerictarget); (i) ground COMPOUND PHARMACOLOGY in the compoundâsKNOWNTREATSrecord and target spectrum; (i) scrutinize the DISEASE MAPPING for clinical plausibility against the historical evidence span; (iv) upweight POLYPHARMACOLOGY / CONVERGENCE when multiple compounds from the same source hit the same target; (v) discount SPECIFICITY when the source has many unrelated claims (kitchen-sink remedy). The prompt instructs the model to quote specific input fields and never speculate beyond the provided payload. Baseline LLM prompt. Role: Candidate nominator. You are a pharmaceutical and natural products expert with broad knowledge of traditional Chinese medicine, pharmacognosy, and bioactive plant compounds. You will be given a passage from the Shen Nong Ben Cao Jing and will be asked to discover and rank plausibility of therapeutic compounds. Read this passage from a historical Chinese herbal text. List each medicinal source discussed; for each, propose up to 10 therapeutic chemical compounds it likely contains, with a plausibility score 0.0-1.0 (0.0 = incoherent or no known mechanism, 1.0 = biologically obvious) and a one-sentence reasoning grounded in known biochemistry or pharmacology. Biomni prompt. Role: Biomedical reasoning expert. Input: You are a pharmaceutical and natural-products expert with broad knowledge of traditional Chinese medicine, pharmacognosy, bioactive plant/mineral/animal compounds, and modern clinical pharmacology. You will be given an extracted corpus of historical medicinal entries. Each block is one claim: a 15 Submission and Formatting Instructions for ICML 2026 SOURCE (a medicinal substance) and a MALADY (the traditional ailment the historical text says it treats), followed by the original passage. You must judge to your fullest capability, using any databases and subagents available necessitating @ChemBL, @PubChem, @PubMed, @OpenTargets as places for you to find relevant compounds or links to mechanistically evaluate the plausibility that the source provided can treat the malady listed. Rules: Score each claim independently; do not let one claim bias another. Do not invent citations or database IDs. Reason from known pharmacology. If the source is an inert mineral or has no plausible bioactive route to the disease, say so and mark Unsupported. LLM-judge prompt.Role: independent biomedical evaluation judge grading the visible reasoning of an automated critic, not deciding whether the underlying therapeutic claim is true. Input:conditionunder judgment (one of the six in Eval 4), the claim, the Pass-1 deterministic signals, the visible payload the critic received (or empty for graph-free conditions), and the criticâs full structured output. Output schema: per-dimension scores in[1, 5]for the six rubric dimensions (Appendix A.7), six binary flags, a recommended status, and a free-text justification referencing specific input fields. Key instructions: treat Pass-1 as an input signal, not ground truth (a critic can be good even when it agrees with an imperfect Pass-1 verdict, if it explains the limitation correctly); penalize hallucinated citations (every compound, target, and disease named by the critic must trace to a payload field, except for widely accepted biomedical facts in text-only conditions); penalize rationales that confuse global top-bucket evidence with loop-closing disease support (a claim with 76 T1-bucket paths but zero of those paths reaching the mapped disease is not well-supported). A.9. Implementation notes Models. Extraction, Auditor canonicalization, Malady mapper, and Reviewer Pass 2 use Gemini 3.1-Flash-Lite at temperature 0 with structured-JSON schemas (rate-limited 1 request/sec, retried up to 4 times on transient failure). The Discovery LLM stages use Gemini Pro under the same rate-limit envelope. Storage. Neo4j AuraDB. All writes are idempotentMERGEstatements with status fields (e.g.,linkerstatus, tdlinkerstatus,mapperstatus) written last so transient failures retry safely. Default re-run filters select only nodes withstatus IS NULL;--retry-missesreprocesseserrorandno * branches;--force-relink deletes prior edges from terminal-status branches only, never from in-flight nodes. This makes the pipeline crash-safe end-to-end. Caching. PubChem responses are disk-cached by query; COCONUT is loaded once into an in-memory inverted index. Re-runs after network interruptions complete in seconds. End-to-end cost. Full Assembly over the 71-chunk corpus runs inâŒ30â40minutes wall-clock, dominated by per-edge Neo4j MERGE latency on AuraDB. Total LLM spend isâŒ$0.25(Flash-Lite); ChEMBL, COCONUT, PubChem, Open Targets, NLM, and EBI are all free. 16 Submission and Formatting Instructions for ICML 2026 A.10. Qualitative scoring of representative critic agent responses We assembled fifty sourceâmalady pairs, constrained so that every pair has both a traditionalTREATSTRADITIONALLY edge and a primary mapping to a modern disease. For each pair we extracted the corresponding historical passage from the Shen Nong Ben Cao Jing to form a mini-corpus. The mini-corpus was given to DeepRoot Discovery, Biomni, or baseline LLM and tasked to score each pair based on the plausibility that the particular source can treat the traditional malady it is claimed to treat. Shown below are the first 10 examples for DeepRoot and Biomni. Table S4. DeepRoot versus Biomni baseline): maladyâmodern-disease mappings and plausibility verdicts across ten (Source, Malady) pairs. DeepRootBiomni # Sourceâ MaladyMapped diseaseVerdictMapped diseaseVerdict 1 Epimediumâ impotence Erectile Dysfunction Very Plausible Erectile Dysfunction Very Plausible 2 Trichosanthes cucumeroidesâ menstrual block AmenorrheaUnsupported AmenorrheaPlausible 3 Rhizoma Arisaematis â heart pain Angina PectorisPreviously Reported Angina PectorisWeak 4 Calcareous Sparâ generalized fever FeverUnsupported Pyrexia (Fever)Unsupported 5 Aster tataricusâ cough CoughPlausibleBronchitis; Asthma Very Plausible 6 Fructus Ailanthi Altissimiâ impotence Erectile Dysfunction PlausibleErectile Dysfunction Weak 7 Ge Genâ toxicityPoisoningPlausibleAlcohol PoisoningPlausible 8 Sanguisorba officinalis â vaginal discharge LeukorrheaUnsupported Leukorrhea; Candidiasis Plausible 9 Hedgehog Peltâ hemorrhoids HemorrhoidsPlausibleHemorrhoidal Disease Unsupported 10 Calcareous Sparâ cough CoughUnsupported Cough; BronchitisUnsupported Each box below shows one verdict for a (Source, Malady) pair.VERY PLAUSIBLEPREVIOUSLY REPORTEDPLAUSIBLEWEAK UNSUPPORTEDbadges reflect the reported plausibility tier (green = strongest; PREVIOUSLY REPORTED denotes a source compound with an already-known therapeutic edge to the mapped disease). Example 1: Epimedium (Source) and Impotence (Malady) DeepRoot Verdict:VERY PLAUSIBLE Weakest Link: The KG contains several non-closing paths for icariin that distract from the primary, well-established PDE5 inhibitory mechanism. Reasoning: Epimedium is well-documented for its primary active constituent, icariin, a known phosphodiesterase-5 (PDE5) inhibitor. The KG confirms the icariinâPDE5Aâerectile dysfunction path, which is the canonical mechanism for treating impotence by enhancing nitric oxide-mediated vasodilation. Although other paths in the graph do not close the loop, the primary mechanism is highly robust and clinically validated. Biomni Verdict:VERY PLAUSIBLE Weakest Link: Icariinâs PDE5A IC 50 (âŒ5,900 nM) isâŒ1,000-fold weaker than sildenafil (âŒ3.5 nM); oral bioavail- ability of the native glycoside in humans remains the key uncertainty. Reasoning: Icariin (CID 5318997; CHEMBL553204) is a confirmed PDE5A inhibitor (IC 50 = 5,900 nM, ChEMBL; active in PubChem AID 362785 â human recombinant PDE5A1 assay), sharing the same mechanism as sildenafil: cGMP preservationâcorpus cavernosum smooth muscle relaxation. In vivo rat models confirm efficacy via GRK2 17 Submission and Formatting Instructions for ICML 2026 downregulation in penile tissue (PMID 39905744) and protection of cavernous tissue under hyperglycemia (PMID 40161546). The metabolite icariside I is significantly more potent. Icariin has reached ChEMBL Max Phase 3, making this the best-validated claim in the dataset. Example 2: Trichosanthes cucumeroides (Source) and Menstrual block (Malady) DeepRoot Verdict:UNSUPPORTED Weakest Link: The lack of any mechanistic path connecting cucurbitacins to reproductive hormonal regulation or uterine function. Reasoning: Trichosanthes cucumeroides contains cucurbitacins, which are potent cytotoxic and anti-inflammatory triterpenoids. While these compounds modulate pathways like IGF2BP1 and ITGAL, there is no clear mechanistic link to the regulation of the menstrual cycle or the treatment of amenorrhea. The KG paths converge on unrelated conditions like diabetes and malaria, failing to provide a coherent basis for this traditional claim. Biomni Verdict:PLAUSIBLE Weakest Link: The primary active agent,ÎČ-trichosanthin, is a large protein (âŒ28 kDa) with an abortifacient rather than emmenagogue mechanism â it does not address functional amenorrhea unrelated to pregnancy. Reasoning: Trichosanthes root containsÎČ-trichosanthin, a Type I ribosome-inactivating protein (RIP) that depurinates 28S rRNA, arresting protein synthesis in trophoblast cells and inducing abortion in vivo (PMID 3596896; 198 PubMed results for âtrichosanthin ribosome inactivatingâ). This coherently explains the TCM use for pregnancy-related amenorrhea. Cucurbitacin B (CID 5281316; CHEMBL2180736) contributes cytotoxic activity (IDH1 IC 50 = 640 nM; 79/122 active PubChem assays). The mechanism is abortifacient rather than broadly emmenagogue, limiting applicability to non-pregnancy amenorrhea. Example 3: Rhizoma Arisaematis (Source) and heart pain (Malady) DeepRoot Verdict:PREVIOUSLY REPORTED Weakest Link: The evidence tier for the adenosine-receptor interactions is limited to bronze, suggesting a need for higher-confidence binding data. Reasoning: Rhizoma Arisaematis contains adenosine, which acts on adenosine receptors (ADORA1, ADORA2A, ADORA2B, ADORA3) to modulate cardiac rhythm and coronary blood flow. These paths close the loop to angina pectoris, providing a plausible mechanism for alleviating heart pain through vasodilation and metabolic protection. The evidence is consistent across multiple receptor subtypes, though the overall tier remains bronze. Biomni Verdict:WEAK Weakest Link: Arisaema species (the actual botanical source) have no documented cardiac pharmacology in any modern database; the claimâs plausibility depends entirely on resolving âHu Zhangâ as Polygonum cuspidatum (resveratrol), which is a different plant. Reasoning: Rhizoma Arisaematis (Tian Nan Xing) derives from Arisaema species; PubMed and ChEMBL searches returned no cardiac or anti-anginal activity for any Arisaema compound. The classical textâs reference to âHu Zhangâ maps to Polygonum cuspidatum in modern pharmacopoeia â a rich source of resveratrol (CID 445154; CHEMBL165, Max Phase 3) with COX-1 (IC 50 = 380â2,840 nM), COX-2 (IC 50 = 3,490 nM), and SIRT1 inhibition relevant to anti-platelet and cardioprotective activity (PMID 41978708). If the source is confirmed as knotweed, the rating rises to Plausible; as Arisaema, it remains Weak. Example 4: Calcareous Spar (Source) and Generalized fever (Malady) 18 Submission and Formatting Instructions for ICML 2026 DeepRoot Verdict:UNSUPPORTED Weakest Link: The substance is an inert mineral with no pharmacological activity relevant to systemic fever. Reasoning: Calcareous Spar is primarily composed of calcium carbonate (CaCO3), an inert mineral in the context of systemic fever reduction. It lacks bioactive organic compounds capable of modulating thermoregulatory centers or inflammatory cytokines. The KG correctly shows no paths, as there is no plausible pharmacological mechanism for this substance to treat fever. Biomni Verdict:UNSUPPORTED Weakest Link: CaCO3 has no antipyretic target in any database; the antipyretic TCM mineral is Gypsum (CaSO4), a distinct compound with documented TLR4/NF-ÎșB inhibition. Reasoning: Calcareous Spar is calcium carbonate (CaCO3; CID 10112; CHEMBL1200539, Approved). Its 31 ChEMBL and 22 OpenTargets indications cover hypocalcemia, renal osteodystrophy, GERD, and osteoporosis â none include fever. No antipyretic bioactivity was found in ChEMBL or PubMed. The TCM antipyretic mineral is Gypsum (CaSO4, Shi Gao), which inhibits TLR4/NF-ÎșB signaling, reduces IL-6/PGE2/TNF-α, and upregulates AVP in LPS-fever models (PMID 37769495). The TCM âcold nature clears heatâ rationale has no pharmacological correlate for CaCO3. Example 5: Aster tataricus (Source) and Cough (Malady) DeepRoot Verdict:PLAUSIBLE Weakest Link: The low-confidence wood-tier evidence for the strobopinin-PTGS interaction limits the certainty of this specific mechanism. Reasoning: Aster tataricus is traditionally used as an antitussive, and its constituents like strobopinin target PTGS1/2 (cyclooxygenases). Inhibition of these enzymes can reduce airway inflammation and prostaglandin-mediated cough reflexes. While the KG confirms the loop to cough, the evidence is wood-tier, and other paths for quercetin and kaempferol do not close the loop to respiratory conditions. Biomni Verdict:VERY PLAUSIBLE Weakest Link: Shionone (CID 12315507) has zero PubChem bioassay data and is absent from ChEMBL; specific antitussive IC 50 values for any Aster compound against a defined respiratory target have not been reported. Reasoning: Aster tataricus (Zi Wan) root contains shionone, epifriedelanol (CID 119242; 2 active PubChem assays), and triterpenoid saponins. Multiple in vivo mouse studies confirm mechanistically specific anti-inflammatory activity in respiratory disease: PI3K/AKT/NF-ÎșB inhibition in asthma (PMID 38081397), HIF-1α/VEGFA pathway suppression in chronic bronchitis, and iNOS/LPS-mediated mediator suppression in acute bronchitis. Saponins additionally provide expectorant activity via mucociliary stimulation. Comprehensive reviews (PMIDs 40271061, 36080154) confirm the traditional antitussive use is well-corroborated by modern pharmacology. Example 6: Fructus Ailanthi Altissimi (Source) and impotence (Malady) DeepRoot Verdict:PLAUSIBLE Weakest Link: The reliance on wood-tier evidence for the identified targets makes the specific pharmacological contribution of these compounds to erectile function speculative. Reasoning: Fructus Ailanthi Altissimi contains flavonoids like quercetin and rutin, which exhibit antioxidant and vasorelaxant properties. The KG identifies paths such as quercetinâAKR1B1 and rutinâADRA2A/C, which are linked to erectile dysfunction. While these targets are relevant to vascular health and smooth muscle tone, the evidence 19 Submission and Formatting Instructions for ICML 2026 is primarily from wood-tier sources, limiting the strength of the mechanistic link. Biomni Verdict:WEAK Weakest Link: Ailanthone and related quassinoids have no demonstrated activity at PDE5, androgen receptors, eNOS, or any pro-erectile target; the pharmacological profile is anticancer and antimalarial only. Reasoning: Ailanthus altissima fruit contains quassinoids (ailanthone, CID 72965; CHEMBL161015) with documented anticancer (HepG2 IC 50 = 240 nM), antiparasitic (P. falciparum IC 50 = 0.003ÎŒg/mL), and NF-ÎșB/PKM2 inhibitory activity. PubMed returned zero results for âAilanthus altissima erectile dysfunction.â No PDE5A, androgen receptor, or nitric oxide pathway activity was found in ChEMBL (8 activities) or PubChem (11 assays, 1 active). The TCM bark use is for spermatorrhea â a distinct condition â and even that lacks modern mechanistic validation. Example 7: Ge Gen (Source) and Toxicity (Malady) DeepRoot Verdict:PLAUSIBLE Weakest Link: The term âtoxicityâ is too non-specific to definitively map to the ACHE-inhibition mechanism identified in the KG. Reasoning: Ge Gen (Pueraria lobata) contains isoflavones like daidzin and flavonoids like rutin and apigenin. The KG identifies paths where these compounds target ACHE (acetylcholinesterase), which is a critical enzyme in managing certain types of neurotoxicity and organophosphate poisoning. While the loop closes, the clinical application of Ge Gen for general âtoxicityâ is broad and requires more specific evidence regarding the type of toxin. Biomni Verdict:PLAUSIBLE Weakest Link: The detoxification evidence is specific to alcohol-related hepatotoxicity and does not extend to the broader traditional claim of resolving âvarious toxins.â Reasoning: Ge Gen (Radix Puerariae) contains puerarin (CID 5281807; CHEMBL486386, Max Phase 2) and daidzein (CID 5281708; CHEMBL8145). Puerarin modulates macrophage polarization (M1âM2) and activates PI3K/AKT signaling to suppress alcohol-induced hepatocellular cytotoxicity (PMID 42010989; 14 PubMed papers). Daidzein inhibits ALDH2 (IC 50 = 9,000 nM) and acts as a phytoestrogen via ERÎČ(IC 50 = 303 nM). OpenTargets confirms Phase 2 clinical registration for alcohol abuse/dependence. The mechanism is coherent for alcohol detoxification specifically, but the broader âvarious toxinsâ claim lacks support. Example 8: Sanguisorba officinalis (Source) and Vaginal discharge (Malady) DeepRoot Verdict:UNSUPPORTED Weakest Link: The absence of any KG paths linking the known astringent tannins of Sanguisorba to the treatment of vaginal discharge. Reasoning: Sanguisorba officinalis is rich in hydrolyzable tannins and ellagic acid, which provide astringent and antimicrobial properties useful for treating inflammatory exudates. However, the KG paths for its constituents like astragalin and genistein do not close the loop to vaginal discharge or related gynecological infections. The current graph evidence focuses on unrelated targets like ESR2 and CA12, which do not explain the traditional use. Biomni Verdict:PLAUSIBLE Weakest Link: The most direct anti-Candida evidence uses the charred/processed form (not raw root), and no human clinical trial for vaginal discharge specifically has been identified. Reasoning: Sanguisorba officinalis contains ellagic acid (CID 5281855; CHEMBL6246; Src IC 50 = 300 nM, DNA 20 Submission and Formatting Instructions for ICML 2026 gyrase IC 50 = 3.3â47ÎŒg/mL), sanguiin H-6 (CID 16130897), and ziyuglycoside I/I (CID 71609288). Direct anti-Candida albicans activity is confirmed for the whole extract (PMID 12169406). Ziyuglycoside I synergizes with tetracycline against MDR S. aureus (MIC = 32ÎŒg/mL; FICI = 0.25â0.375; PMID 41385864) via membrane disruption and ROS generation. Astringent ellagitannins reduce mucosal secretion and inflammation, providing a dual antimicrobial-astringent mechanism supporting the traditional use. Example 9: Hedgehog Pelt (Source) and Hemorrhoids (Malady) DeepRoot Verdict:PLAUSIBLE Weakest Link: The reliance on wood-tier evidence for the interaction between the peltâs constituents and the identified receptors. Reasoning: Hedgehog Pelt contains various organic acids and signaling molecules that the KG links to ADRA1A and NR3C1 (glucocorticoid receptor). These targets are involved in vasoconstriction and anti-inflammatory responses, which are relevant to reducing the swelling and pain associated with hemorrhoids. While the mechanism is plausible, the source material is complex and the evidence tier is low. Biomni Verdict:UNSUPPORTED Weakest Link: No bioactive compounds from hedgehog pelt (Erinaceus europaeus skin) have been characterized in any database; the entire mechanistic chain is unestablished. Reasoning: Hedgehog pelt (Wei Pi, Pellis Erinacei) from Erinaceus europaeus returned zero PubMed results for âhedgehog pelt hemorrhoidsâ or âCi Wei Pi TCM hemorrhoids.â No ChEMBL entries exist for E. europaeus skin compounds. Compounds found in searches (erinacines, hericenones) derive from the unrelated hedgehog mushroom (Hericium erinaceus). The traditional use likely reflects sympathetic magic or empirical use of charred animal material as a non-specific hemostatic, neither of which constitutes a credible pharmacological mechanism. Example 10: Calcareous Spar (Source) and Cough (Malady) DeepRoot Verdict:UNSUPPORTED Weakest Link: The substance is an inert mineral with no pharmacological activity relevant to cough. Reasoning: As a mineral composed of calcium carbonate, Calcareous Spar possesses no antitussive properties or mechanisms to modulate the cough reflex. It does not interact with respiratory receptors or inflammatory pathways involved in cough. The lack of KG paths reflects the biological reality that this substance is inert for this indication. Biomni Verdict:UNSUPPORTED Weakest Link: CaCO3 has no antitussive, bronchodilatory, or anti-inflammatory mechanism in any database; PubMed returned zero results for âcalcium carbonate antitussive cough.â Reasoning: Calcareous Spar is the same CaCO3 (CID 10112) as Example 4, with no antitussive indication in ChEMBL or OpenTargets. While calcium signaling is mechanistically relevant to cough â Ca2+-activated Cl- channels (TMEM16A) regulate airway smooth muscle tone and cough reflex sensitivity (PMID 39608849) â these are endogenous calcium-signaling mechanisms unrelated to exogenous CaCO3 supplementation. The TCM rationale of minerals âdescending rebellious qiâ to suppress cough has no pharmacological correlate for this compound. 21