DeepTCM1.0: A Multi-Expert AI Agent for Deciphering Mechanisms of Chinese Herbal Formulae Based on General Large Language Models

Methodological gap

• Traditional data-mining methods remain largely confined to descriptive associations and phenomenological observations, failing to reveal complex synergistic relationships among herbal-formula components or molecular regulatory networks[8].Network pharmacology datasets are often fragmented and updated infrequently, restricting reliability.Source: 1.Introduction · p. 3
• Molecular docking remains poorly aligned with pattern differentiation, holistic regulation, and formula compatibility, while its functional, cost, and efficiency limitations restrict TCM applicability.Its core design and technical logic primarily target small chemical molecules rather than TCM’s systemic principles.Source: 1.Introduction · p. 4

AI rationale

• In ophthalmology, hepatitis serological testing, and rheumatoid arthritis, distilled LLMs fine-tuned on high-quality domain-specific datasets enhanced domain-knowledge representation and logical-reasoning performance in specialized medical fields [19,12,54].Source: 1.Introduction · p. 4
• Multi-agent systems are highly compatible with TCM’s holistic perspective, pattern differentiation–based treatment, and formula compatibility, and employ consensus-seeking and evidence-based querying mechanisms [36].Developing such systems is urgently required to extend general-purpose large language models such as GPT and DeepSeek for innovative research [36].Source: 1.Introduction · p. 5

Proposed framework

• DeepTCM1.0 proposes a multi-expert intelligent agent framework for deep exploration of TCM mechanisms via general large language models.The framework uses DeepSeek V3.2 as its central model and applies specialized expert roles, progressive task decomposition, and iterative quality control.Source: 1.Introduction · p. 5

Medicine Mechanism Research

• DeepTCM1.0 is designed around TCM principles including pattern differentiation-based treatment, holistic regulation, and compound prescription.The framework aims to provide a reproducible technical basis for classic-formula research and syndrome-mechanism elucidation.Source: Medicine Mechanism Research · pp. 5–6
• The framework integrates classical TCM, formula compatibility, pharmacology, network pharmacology, molecular biology, and molecular docking to analyze herbal-formula mechanisms from multiple perspectives.It establishes dynamic mappings between abstract TCM concepts and measurable modern-biological indicators, enabling complex interdisciplinary tasks to be decomposed and executed efficiently.Source: Medicine Mechanism Research · p. 6
• Guizhi Decoction serves as the representative case for demonstrating DeepTCM1.0’s mechanistic-analysis effectiveness and reproducible research framework.The stated application supports the transition from experience-based TCM knowledge transmission toward intelligence-driven scientific innovation.Source: Medicine Mechanism Research · p. 6

2.1 Overall Design Strategy

• The framework uses configurable LLM expert roles to simulate multidisciplinary collaboration while prioritizing TCM theory, lightweight implementation, and interpretable evidence-based reasoning.DeepSeek V3.2 supplies embedded multidisciplinary knowledge without additional knowledge-graph construction or model fine-tuning, while iterative discussion supports traceable, self-consistent conclusions.Source: §2.1 · p. 6
• DeepTCM1.0 uses a three-tier collaborative architecture comprising task decomposition and team formation, multi-expert reasoning and emergent insight, and integrative validation and output.The architecture is presented as a modular closed loop for structured research analysis.Source: §2.1 · pp. 6–7, §2.2 · p. 7

2.2 Three-Tiered Collaboration Framework

• The three-tier workflow parses research questions, assembles interdisciplinary teams, conducts multiple rounds of cross-disciplinary reasoning, and produces traceable hypotheses and standardized research protocols.Critic agents and iterative multi-agent review are used to reduce reasoning bias and AI hallucinations.Source: §2.2 · p. 7, Figure 2 · p. 7

2.2.1 First Tier: Task Decomposition and Team Formation

• The first tier acts as the decision-making hub, with a Principal Investigator decomposing complex TCM prescription-mechanism questions into standardized multidimensional sub-tasks.Its design follows a TCM-led interdisciplinary research philosophy.Source: §2.2.1 · pp. 7–8
• The Recruiter combines predefined fixed experts with task-specific dynamic experts to provide comprehensive disciplinary coverage without perspective gaps.This configuration supports systematic integration of traditional TCM with modern scientific technologies.Source: §2.2.1 · p. 8

Integration

• A hybrid team of fixed TCM specialists and dynamically recruited biomedical experts operationalizes “TCM-guided, multidisciplinary collaboration, and specialized division of labor.”Domains include classical TCM theory, formula systems, modern medicine, pharmacology, immunology, neuroendocrinology, metabolomics, and microbiomics.Source: Integration · p. 9

2.2.2 Second Tier: Multi-Expert Collaborative Reasoning Layer

• Critic–expert bidirectional reasoning examines logical integrity, evidence sufficiency, and TCM–modern compatibility while reducing hallucinations and enabling novel mechanistic insights.The critic identifies conflicts, blind spots, and omissions, providing quality-control support through sustained questioning and targeted recommendations.Source: §2.2.2 · p. 10
• The second tier coordinates independent expert analyses, reverse-evidence critique, theory-first integration, and targeted optimization across three iterative discussion rounds.The sequence is independent interpretation, integration and deepening following “TCM theory first, modern pharmacology second, and cross-validation third,” then controversy resolution with supplementary evidence and hypothesis generation.Source: §2.2.2 · pp. 10–11

2.2.3 Third Tier: Integrative Deepening and Solution Output Layer

• The third tier integrates viewpoints, consensus, controversies, and unverified hypotheses into a traceable report structured as classical theory, macroscopic logic, microscopic mechanisms, and clinical application.A report integration specialist works with the Principal Investigator and critic to deliver a logically complete, evidence-supported compound-mechanism analysis report.Source: §2.2.3 · p. 11
• The third tier synthesizes expert viewpoints into scientifically coherent hypotheses, complete research plans, implementable research schemes, and consensus conclusions.The Principal Investigator directs integration, the critic performs corrective quality review, and the report specialist standardizes final documentation.Source: §2.2.3 · p. 12

2.3.1 Development Environment

• The system was developed in Python 3.10 using an OpenAI-compatible DeepSeek API setup, JSON processing, and a Windows 11 environment with 16 GB memory and an Intel Core i7 CPU.Source: §2.3.1 · p. 12

2.3.2 Implementation of AI Agents

• AI experts use a shared Agent class with role, prompt, and expertise attributes, plus analyze() and discuss() methods, while prompts define roles, knowledge, objectives, and boundaries.Source: §2.3.2 · p. 12

2.3.3 Multi-Round Discussion Process Control

• The Discussion Manager preserves dialogue histories, invokes critic quality control after each round, supports resumption and caching, and locally persists prompts, outputs, and reports for reproducibility.Source: §2.3.3 · p. 12

2.3.4 Distinctive Advantages

• The framework requires no additional domain-data or fine-tuning costs, adapts through prompt engineering and role simulation, and supports reproducible replication for other classical TCM formulae.Source: §2.3.4 · pp. 12–13

2.4 Case Analysis Design

• Guizhi Decoction was selected as a representative case to assess knowledge integration, mechanism discovery, and evidence traceability across classical TCM and modern scientific perspectives.The planned output covers theory interpretation, compatibility, material constituents, network mechanisms, integrative conclusions, and experimental validation protocols, using fixed and dynamically augmented experts across relevant disciplines.Source: §2.4 · p. 13

3.1 Agent System Execution Process

• DeepTCM 1.0 completed a full-process analysis of Guizhi Decoction’s efficacy mechanism using three-layer collaboration and three-round iteration.Source: §3.1 · p. 14

3.1.1 Task Deconstruction and Team Formation

• The Principal Investigator decomposed the task into six sub-tasks spanning classical theory, formula compatibility, pharmacokinetics, network regulation, omics, and clinical evidence.Source: §3.1.1 · p. 14
• The system autonomously assembled 11 experts, combining seven fixed core experts with four task-specific specialists across complementary biomedical and TCM domains.Specialized roles included immunology and neuroendocrinology, metabolomics and microbiomics, systems biology and network pharmacology, and pharmacokinetics and pharmaceutical analysis.Source: §3.1.1 · p. 14

3.1.2 Multi-Expert Collaborative Reasoning

• In the first collaborative round, 11 experts independently developed perspectives, underwent critical inquiry and quality-control audits, and produced an initial multidimensional evidence foundation.The core conclusions of each expert were presented in Table 1.Source: §3.1.2 · p. 14, Table 1 · p. 15

Classical TCM Interpretation

• Guizhi Decoction’s classical compatibility is interpreted through balanced dispersion and restraint, warm porridge support, pulse differentiation, and two functional modules: the ying–wei and middle-energizer units.The Guizhi–Shaoyao 1:1 ratio supports the balance between dispersion and astringency, while warm diluted porridge replenishes ying yin and may enhance absorption.Source: Interpretation · pp. 16–17

Chinese Medicine Mechanism

• The formula’s macro–micro interpretation maps cinnamaldehyde to TRPV1–CGRP effects, paeoniflorin to TLR4/NF-κB inhibition, glycyrrhetinic acid to 11β-HSD2 inhibition, and herb compatibilities to immune–inflammatory regulation.Guizhi–Shaoyao synergy is assigned to TRPV1–NF-κB signaling, while Guizhi–licorice and Shaoyao–licorice correspond to β2-AR activation and inhibition of IL-6 and TNF-α; formula decomposition remains necessary for validation.Source: Interpretation · pp. 17–18

Syndrome Differentiation

• The syndrome model links fever, sweating, aversion to wind, and a floating–moderate pulse with PGE2, AQP5, TRPM8, vascular tone, and cholinergic nerve–sweat gland dysfunction.Differentiation from the Taiyang cold-damage excess pattern relies on sweating versus anhidrosis and moderate versus tight pulse, with microscopic indicators required for support.Source: Interpretation · pp. 18–19

Integrative Chinese–Western Medicine

• A “three-tier, dual-axis, one-foundation” model links the NEI regulatory layer, vascular–microcirculatory effector layer, gut–metabolic support layer, thermoregulatory axis, and barrier immunity axis.Ying–wei disharmony is interpreted as acute NEI maladaptation, while Guizhi Decoction acts through a neural–immune–endocrine–metabolic network to restore homeostatic resilience.Source: Interpretation · pp. 19–20

Systems Biology and Neuroendocrinology

• Systems analyses identify AKT1, TNF, IL6, and CHRM3 as predicted targets and associate the formula with NF-κB, cholinergic-synapse, HPA-axis, immune, neural, and metabolic modules.The interpretation highlights syndrome-specific targets as a bottleneck requiring clinical multi-omics, while NEI imbalance is represented by TNF-α/IL-10, cortisol rhythmicity, HRV, and sIgA biomarkers.Source: Interpretation · pp. 21–22

Metabolomics, Microbiomics, and Pharmaceutical Analysis

• The interpretation extends to microbiota, metabolism, and pharmacokinetics: warm porridge and herb polysaccharides may increase SCFAs, while metabolites, CYP3A4 inhibition, and gut conversion shape in vivo activity.Cinnamaldehyde has less than 1% oral bioavailability and acts predominantly through cinnamic acid, whereas paeoniflorin is hydrolyzed into more active metabolites; UPLC–Q-TOF/MS and PK–PD modeling are proposed for validation.Source: Interpretation · pp. 23–24

Critic-guided revision

• The critic required traceable classical citations, clarified “moderate pulse” as reduced vascular tension rather than slowed rate, and reformulated unsupported wei qi–SCFA links as testable hypotheses.The revisions emphasized accuracy, standardization, and consistency with classical sources.Source: Interpretation · p. 25

Classical–modern translation

• The revised interpretation defined “moderate pulse” as reduced vascular tone and diminished elasticity, linking it to relative ying yin insufficiency and proposing PWV and vascular tone indices for measurement.Classical commentaries were incorporated to support the standardized interpretation.Source: Interpretation · p. 26

Hypothesis standardization

• No direct evidence currently supports a wei qi–SCFA linkage; the proposal was framed as a gut microbiota–immune hypothesis requiring experimental validation.A proposed test compares SCFA levels and immune-related biomarkers between patients with ying–wei disharmony and healthy controls.Source: Interpretation · p. 26

Round 3 optimization

• Round 3 re-examined core TCM concepts through classical interpretation, modern translation, and mechanistic analysis, promoting deeper integration with research design.Five classical theoretical constraints were incorporated as fixed conditions in a dynamic network integration model.Source: Interpretation · p. 27

Validation assurance

• A comprehensive assurance system combined theoretical constraints, falsifiability criteria, and data-driven validation, including TMCAD to test whether bioactive substances improve symptoms through network-level changes.The framework also established multilevel hypothesis falsification and bidirectional statistical analysis to strengthen macro–micro integration.Source: Interpretation · p. 27

Critic Feedback)

• Precise tracing of original classical sources was achieved, with key provisions from Shanghan Lun Articles 12, 2, 53, and 54 comprehensively addressed.Alignment between classical theory and the core consensus was established, and supporting evidence from classical commentators was incorporated.Source: Critic Feedback) · pp. 27–28

3.1.3 Integration, Verification and Output

• The integration process maintained participation by the Principal Investigator, critic, and report integration specialist, producing structured consensus conclusions, mechanistic hypotheses, and research plans.The cuozhou hypothesis concerning gut microbiota-derived metabolites was explicitly marked as requiring experimental validation.Source: Critic Feedback) · pp. 27–28, §3.1.3 · p. 28

3.2.1 Core Positioning at the Classical Theory Level

• The consensus identifies Guizhi Decoction’s core indication as ying–wei disharmony, emphasizing a 1:1 Guizhi–Shaoyao ratio and the supportive roles of middle-energizer ingredients and warm porridge.The formulation balances dispersion and astringency, while ginger, jujube, licorice, and warm administration support therapeutic efficacy.Source: §3.2.1 · p. 28

3.2.2 Mechanism Analysis at the Modern Science Level

• The modern mechanism is modeled as three-tier network harmonization across NEI regulation, vascular–microcirculatory effectors, and gut–metabolic support.Reported mechanisms include TRPV1, NF-κB, HPA-axis, cholinergic sweat-gland, vascular-tone, PGE2, microbiota, SCFA, barrier, and immune-cell regulation.Source: §3.2.2 · pp. 28–29

3.2.3 Material Basis and In Vivo Process

• Guizhi Decoction’s in-vivo bioactive basis is a dynamic ensemble of prototype compounds, metabolites, and microbiota-derived products, with formula-level CYP3A4 inhibition increasing active-constituent bioavailability.Cinnamaldehyde has less than 1% oral bioavailability and acts mainly through cinnamic acid, while paeoniflorin and glycyrrhizic acid undergo microbial or metabolic conversion.Source: §3.2.3 · p. 29

3.3.1 Definition of Pharmacodynamic Material Basis

• Iterative discussion resolved the pharmacodynamic-material controversy by adopting a dynamic ensemble of prototype compounds, metabolites, and microbiota-derived products.Different material forms may predominate in different contexts.Source: §3.3 · p. 29, §3.3.1 · p. 30

3.3.2 Construction of Macro-Micro Connection Logic

• A three-level framework connects classic syndromes to intermediate physiological phenotypes and microscopic molecules, mapping syndrome features to measurable indicators and molecular targets.Mapped elements include PGE2, AQP5, TRPM8 sensitivity, vascular tone, NF-κB, CHRM3, and TRPV1.Source: §3.3.2 · p. 30

3.3.3 Standardization of Research Scheme

• After three rounds of discussion among 11 experts, research validation was standardized around a wind–cold/LPS composite animal model and a deep-phenotyping clinical study.Both approaches standardized warm-porridge administration and maintenance of warmth.Source: §3.3.3 · p. 30
• The protocol administers 200 mL millet porridge at 60–70 °C and maintains a thermally stable environment at 32–34 °C for 1 h.Source: §3.3.3 · p. 30
• Animal evaluation spans behavior, neuro–endocrine–immune indices, vascular microcirculation, and gut microbiota–metabolite profiles, while clinical sampling covers T0–T5 with Z-score and mixed-effects modeling.Source: §3.3.3 · p. 30

3.4 Innovative Mechanism Hypotheses and Research Plans

• Based on the consensus, DeepTCM 1.0 proposes testable mechanistic hypotheses and corresponding validation research plans.Source: §3.4 · p. 30

• Guizhi Decoction is hypothesized to achieve ying–wei harmonization through coordination among core regulatory, core effector, and foundational support layers.These layers comprise the neuroendocrine–immune network, vascular–microcirculatory system, and gut–metabolic system, respectively.Source: Hypothesis 1: Three-Layer Network… · pp. 30–31
• Cinnamaldehyde and paeoniflorin may initiate peripheral signaling through cutaneous TRPV1 activation, followed by NF-κB-mediated immune–inflammatory modulation and HPA-axis neuroendocrine stabilization.Source: Hypothesis 1: Three-Layer Network… · p. 31
• Gut microbiota-derived SCFAs may optimize metabolic homeostasis, completing a “peripheral initiation-central integration-foundational support” closed-loop regulatory system illustrated in Figure 4.Source: Hypothesis 1: Three-Layer Network… · p. 31, Figure 4 · p. 31

Research Plan

• Guizhi and Shaoyao compatibility is proposed to produce immunomodulatory and anti-inflammatory effects through a combined “dispersion–astringency” mechanism, while middle-energizer herbs regulate gut microbiota.These mechanisms synergize with fermentable substrates from warm, diluted porridge to promote SCFA production, improve metabolic homeostasis, and support the spleen–stomach theory.Source: Research Plan · p. 31
• The proposed Compatibility–Metabolism–Microbiota Synergy Hypothesis integrates pharmacological effects, microbiota regulation, SCFA production, and metabolic homeostasis.It is presented as support for the classical concept that the spleen and stomach generate ying–wei transformation.Source: Research Plan · p. 31, Figure 5 · p. 32

Hypothesis 3: Dynamic Effector Substance Collection Hypothesis and Its Research Plan

• Guizhi Decoction is hypothesized to act through a dynamic ensemble of prototype components, metabolites, and microbiota-transformed products rather than a single component.Examples include cinnamic acid, paeoniflorin derivatives, and glycyrrhetinic acid.Source: Hypothesis 3: Dynamic Effector… · p. 32, Figure 6 · p. 33
• This dynamic ensemble is proposed to act synergistically on cholinergic synapses, NF-κB, and the HPAaxis through a multi-component–multi-target–multi-pathway regulatory paradigm.The hypothesis is illustrated as the Dynamic Ensemble Hypothesis of Bioactive Substance Assemblages.Source: Hypothesis 3: Dynamic Effector… · p. 32, Figure 6 · p. 33

3.5.1 Double-Blind Scoring Experimental Design

• Evaluation used a double-blind anonymous Likert five-point design to assess five AI-generated Guizhi Decoction mechanistic reports, including one DeepTCM1.0 report and four single-model reports.The single-model reports were generated by DeepSeek V4, Doubao, Kimi k2.6, and Qwen3.5.Source: §3.5.1 · p. 33
• Four independent evaluators—DeepSeek, Doubao, Kimi, and Qwen—scored every anonymous report using a predefined standardized rubric to reduce single-model evaluation bias.Reports were anonymized before scoring and unblinded afterward for comparative analysis.Source: §3.5.1 · p. 33

3.5.2 Evaluation Dimension Design

• The rubric assessed five dimensions: classical TCM completeness and accuracy, modern scientific comprehensiveness and rigor, interdisciplinary integration, systematic logical integrity, and academic-expression standardization.Each dimension used a Likert five-point scale, with scores from 1 representing very poor performance.Source: §3.5.2 · p. 34

1. Evaluation Subjects

• The evaluation subjects comprised the DeepTCM1.0 multi-agent collaborative report and reports independently generated by DeepSeek V4, Doubao, Kimi k2.6, and Qwen 3.5.These were organized as multi-agent and single-large-language-model groups.Source: §1 · p. 34

2. Double-Blind Scoring Procedure

• The scoring procedure anonymized five reports, randomized presentation order, used four independent evaluators for five repetitions, and produced 100 independent scoring assessments.Unified prompts and newly initialized dialogue contexts constrained scoring behavior and reduced order, historical-context, and single-instance stochastic effects.Source: §2 · pp. 34–35

1. Reliability Testing

• Reliability testing combined Krippendorff’s alpha for scoring consistency among evaluators with a two-way random-effects absolute-agreement ICC(2,1) model for repeated scores within evaluators.Raw five-dimensional scores supported inter-evaluator analysis, while total scores across five rounds supported intra-evaluator analysis.Source: §1 · p. 35

2. Validity and Difference Testing

• Validity testing used repeated-measures ANOVA with eta-squared, Mann–Whitney U testing for multi-agent versus single-model totals, and corresponding effect-size analyses.Statistical significance was interpreted as effective report discrimination, while larger η² values indicated greater explanatory power; Mann–Whitney testing was selected for the relatively small sample and non-normality assumptions.Source: §2 · p. 36

3.6.1 Reliability and Validity Analysis

• ICC values for Doubao, DeepSeek, Kimi, and Qwen were 0.867, 0.819, 0.789, and 0.798, respectively, all exceeding the acceptable threshold of ICC ≥0.7.These results indicated stable repeated scoring and controllable random error.Source: §3.6.1 · p. 37
• Inter-evaluator reliability was excellent, with dimension-level Krippendorff’s α values of 0.842–0.867 and an overall total-score coefficient of 0.868.The coefficients exceeded the 0.8 threshold for excellent consistency and indicated strong evaluator agreement.Source: §3.6.1 · p. 37
• One-way repeated-measures ANOVA identified highly significant intergroup differences (F=66.969，P＜0.001), with η²=0.738 indicating that 73.8% of total-score variance was explained by report-quality differences.The result confirmed excellent discriminative validity.Source: §3.6.1 · p. 37

2. Validity and Difference Testing

• Mann–Whitney U testing showed highly significant differences between the multi-agent and single-model groups, with a large effect size of r=0.5403.The analysis supported the conclusion that DeepTCM1.0 significantly outperformed single general-purpose large language models.Source: §3.6.1 · pp. 37–38

3.6.2 Radar Chart Visualization Results

• DeepTCM1.0 significantly outperformed single general-purpose large language models, achieving near-ceiling scores of 4.8–5.0 across all five evaluation dimensions, with highly convergent evaluator contours indicating strong agreement and stability.Source: §3.6.2 · p. 39

3.6.3 Summary of Scoring Statistics

• Among single-model baselines, Kimi k2.6 achieved the highest overall mean score, followed by DeepSeek V4, while Doubao and Qwen3.5 scored comparatively lower.Performance divergence was also evident across classical TCM interpretation, modern scientific reasoning, and interdisciplinary integration.Source: §3.6.3 · p. 40
• The results objectively reflected capability differences among models in mechanistic interpretation of traditional Chinese medicine formulae.Together with radar visualization results, they demonstrated the performance advantage of DeepTCM1.0 over single general-purpose large language models.Source: §3.6.3 · p. 40
• DeepTCM1.0’s multi-expert collaboration, three-round iterative reasoning, and critic-based quality control better addressed TCM formula research complexity.The relevant complexity involves multi-component interactions, multi-target regulatory mechanisms, interdisciplinary knowledge integration, and dependence on classical TCM theory.Source: §3.6.3 · p. 40
• DeepTCM1.0 achieved the highest scores across all five evaluation dimensions, and its mean total score significantly exceeded those of all single-model systems.These quantitative results are presented in Table 2 and summarized through repeated scoring statistics.Source: Table 2 · p. 40, §3.6.3 · p. 40

4 Discussion

• DeepTCM 1.0 achieved multidimensional analysis of Guizhi Decoction mechanisms, integrating cross-disciplinary knowledge and aligning logically coherent hypotheses with TCM principles.The framework demonstrated advantages in integrating knowledge, constructing mechanistic hypotheses, and adhering to TCM theoretical principles.Source: §4 · p. 41

4.1.1 Completeness of Cross-Disciplinary Knowledge Integration

• Eleven specialized agents dynamically linked classical theory, clinical practice, pharmacology, and systems biology into a classical theory–mechanisms–clinical application knowledge chain.This addressed the fragmented databases, static analyses, and inefficient cross-domain integration associated with conventional network pharmacology [26, 34, 35, 41].Source: §4.1.1 · p. 41

4.1.2 Logical Closed-Loop Nature of Mechanism Analysis

• The framework imposed a “Classic Traceability - Evidence Support - Logical Deduction” loop, with Critic quality control and PK–PD modeling supporting traceable, coherent mechanism conclusions.Its Three-Layer Network Harmonization Hypothesis was grounded in Ying-Wei theory and supported by network pharmacology, metabolomics, and microbiota analyses.Source: §4.1.2 · p. 42

4.1.3 Precision of TCM Theory Adaptability

• The proposed framework consistently prioritized classical TCM theory, emphasized cuozhou wenfu and zhongjiao, and strictly adhered to the classical 1:1 ratio.Source: §4.1.3 · p. 42

4.2 Implications for TCM Methodology

• DeepTCM 1.0 offers a lightweight, low-cost, reproducible route to TCM mechanism analysis without additional database construction or model fine-tuning.Role prompt engineering transforms general large models into TCM domain experts, while shareable prompts and process designs support migration to other classic formulas.Source: §4.2 · p. 42
• The approach may help shift TCM research from experience-based knowledge inheritance toward intelligence-driven scientific research, while its generality remains to be tested across formulas.The authors propose applying the framework to Mahuang Decoction and Xiao Chaihu Decoction to assess generalizability and transferability.Source: §4.2 · p. 42, §4.3 · p. 43

4.3 Limitations and Future Directions

• Current limitations include incomplete coverage of rare or region-specific TCM knowledge, dependence on prompt quality, and mechanism hypotheses derived primarily from knowledge-based inference.The proposed hypotheses therefore still require experimental verification.Source: §4.3 · p. 43
• Future research could incorporate RAG to dynamically expand domain knowledge coverage without model fine-tuning using real-time updated databases such as TCMSP [44] and SymMap, and develop automatic prompt optimization algorithms based on historical discussion outcomes.Source: §4.3 · p. 43

5 Conclusion

• DeepTCM 1.0 successfully enabled multidimensional mechanistic analysis of Guizhi Decoction as a representative TCM formula case.Source: §5 · p. 43
• The framework used a three-layer collaboration architecture and three-round iterative discussion mechanism.Source: §5 · p. 43
• It integrated TCM classics, formula compatibility, and modern pharmacology to generate innovative mechanism hypotheses with a logical closed loop.Source: §5 · p. 43
• Compared with existing technologies, DeepTCM 1.0 demonstrated advantages in cross-disciplinary knowledge integration, TCM theory adaptability, and interpretability.Source: §5 · p. 43
• The framework provides a lightweight and reproducible paradigm for analyzing complex TCM systems.Source: §5 · p. 43
• Future work will focus on experimental validation and framework extension to promote applications in TCM modernization research.Source: §5 · p. 43

Funding Declaration

• The authors received no specific funding for this work.Source: Funding Declaration · p. 44

Appendix

• Table A1 reports dimension-level and total scores from five rounds of repeated scoring.Source: Table A1 · p. 50