PaperlayerView exact revision on arXiv

Source-linked AI summary

LingShu: A Large-Scale Symptom-Centric Contextualized Knowledge Graph Bridging Traditional Chinese Medicine and Modern Biomedicine

Rui Hua, Zixin Shu, Kai Chang, Dengying Yan, Jianan Xia, Hui Zhu, Shujie Song, Shurui Yang, Tongxin Wang, Yue Yin, Yu Wei, Lijuan Pei, Yunhui Hu, Hao Xu, Mingzhong Xiao, Xiaodong Li, Haibin Yu, Runshun Zhang, Wenjia Wang, Baoyan Liu, Xuezhong Zhou

arXiv:2608.20402v1cs.CLcs.AI

TL;DR

Existing medical knowledge graphs separate symptom, TCM, and biomedical concepts and struggle to represent conditional relations. LingShu addresses this with a symptom-centric hybrid graph whose 17.33 million entity records include semantic triples and contextualized quadruples for linking these knowledge domains.

  • Problem

    Existing resources lack symptom-centric integration across TCM and biomedicine, and conventional triples inadequately represent conditional medical knowledge.

  • Method

    LingShu integrates heterogeneous clinical, TCM, and biomedical sources using a symptom-centric hyper-relational model with typed triples and contextualized quadruples.

  • Results

    17.33 million atom-level entity records, 17.19 million semantic triples, and 22.29 million contextualized quadruples were integrated into LingShu.

  • Takeaways & Limitations

    LingShu provides a scalable infrastructure for cross-paradigm knowledge alignment and evidence-grounded reasoning through symptom-centric organization and contextualized representation.

  • Takeaways & Limitations

    Population-level associations come from limited COVID-19 and cirrhosis cohorts, relation strengths are not quantitatively weighted, and text extraction remains vulnerable to OCR errors and incomplete context.

Abstract

from arXiv · show

Biomedical knowledge graphs (KGs) are pivotal for knowledge organization, yet traditional binary relations often struggle to represent the conditional nature of biomedical knowledge. Symptoms provide a shared phenotypic layer for linking Traditional Chinese Medicine (TCM), which relies on symptom patterns for syndrome differentiation and treatment selection, with modern biomedicine, which connects clinical manifestations to diseases and molecular mechanisms. We present LingShu, a large-scale symptom-centric contextualized knowledge graph designed to bridge TCM and modern biomedicine. The exported version of LingShu analyzed in this study comprises 17.33 million atom-level entity records and 39.47 million relation records, including 17.19 million semantic triples and 22.29 million contextualized quadruples. LingShu integrates multi-source data, including clinical electronic medical records, authoritative TCM texts, biomedical ontologies, and curated knowledge bases, through a pipeline combining natural language processing, terminology normalization, and human-in-the-loop verification. A key innovation of LingShu is its hybrid data model: it maintains 64 typed triple relation patterns to ensure broad connectivity, while incorporating 35 contextual quadruple relation patterns to capture conditional medical associations. This dual-structure approach explicitly encodes conditional knowledge, providing a granular representation of the contexts associated with medical relations. These contextualized relations cover syndrome-dependent herb efficacy, disease-contextualized drug effects, population-specific clinical associations, and mechanism-related therapeutic responses. Furthermore, we developed a web platform (http://www.tcmkg.com/) that integrates graph visualization, graph-based reasoning, and an evidence-grounded knowledge question-answering agent.

1 Introduction

LingShu is introduced as a symptom-centric knowledge graph that bridges TCM and modern biomedicine while explicitly representing conditional biomedical knowledge. It combines a hyper-relational schema, large-scale integrated entities, and a web platform for context-aware exploration, reasoning, and question answering.

  • Motivation: Existing medical KGs often subordinate symptoms to diseases, struggle to align heterogeneous TCM and biomedical vocabularies, and cannot preserve conditional relations with conventional triples.Biomedical knowledge is inherently conditional, with relations depending on specific clinical or molecular contexts.
  • Core knowledge graph: LingShu prioritizes symptom phenotypes as core entities to bridge TCM and modern biomedicine, integrating 17.33 million atom-level entity records.Its symptom-centric design reflects TCM syndrome differentiation and biomedicine’s use of symptom patterns to understand pathology.
  • Contextualized schema: 35 typed contextual quadruple patterns encode conditional variables and enable context-aware graph retrieval and reasoning beyond conventional triples.The hyper-relational model represents biomedical facts with contextual qualifiers such as disease contexts and molecular targets.
  • Applications: The web-based platform supports graph exploration, interpretable context-aware reasoning, multi-mode knowledge question-answering, and human-in-the-loop data management.The platform is designed for continuous knowledge curation alongside graph-based applications.

2 Related Work

Prior biomedical, phenotypic, pharmacological, and TCM resources support terminology normalization, evidence retrieval, and knowledge integration but do not adequately represent symptom-centric, cross-domain, conditional medical knowledge. LingShu addresses these gaps with a symptom-centric graph bridging TCM and modern biomedicine through unified atom–concept representations and contextualized relation schemas.

  • Biomedical resources: Biomedical terminologies, literature resources, and molecular databases support semantic interoperability, entity alignment, information extraction, and evidence retrieval.Examples include UMLS, MeSH, PubMed, Gene Ontology, and UniProt.
  • Phenotype resources: Clinical phenotype resources structure symptoms and disease-associated abnormalities, including hierarchical symptom concepts and standardized phenotypic representations.Relevant resources include Symptom Ontology, HPO, MalaCards, Orphanet, ICD-11, and SNOMED CT.
  • Biomedical knowledge graphs: Biomedical knowledge graphs connect diseases, drugs, genes, proteins, chemicals, pathways, phenotypes, and clinical observations for mechanistic and pharmacological reasoning.Hetionet exemplifies heterogeneous integration for disease biology and drug repurposing, while later graphs extend the precision-medicine paradigm.
  • TCM knowledge graphs: TCM knowledge organization has advanced through ontologies, databases, named entity recognition, relation extraction, graph construction, and emerging large-language-model applications.These efforts structure knowledge from classical texts, clinical records, medical cases, and domain databases, but existing resources remain fragmented and insufficiently symptom-centric.
  • Research gaps: Three gaps remain: limited symptom-centric integration, incompatible TCM and biomedical conceptual systems, and insufficient representation of conditional medical knowledge.These gaps impede alignment among symptom patterns, syndromes, interventions, molecular mechanisms, prescription modification, population-specific associations, and evidence provenance.
  • LingShu’s approach: LingShu addresses these gaps by integrating heterogeneous resources through unified atom–concept representations and typed triple and contextual quadruple schemas.Hyper-relational quadruples encode the contexts under which medical relations are valid, enabling a symptom-centric bridge between TCM and modern biomedicine.

3 Materials and Methods · 3.1 Ontology Design

LingShu is constructed through a reproducible symptom-centric knowledge acquisition and graph-construction pipeline that bridges TCM and modern biomedicine. Its ontology combines standardized entity types, 64 typed triple patterns, and 35 contextual quadruple patterns to encode both broad semantic connectivity and condition-dependent medical knowledge.

  • 3 Materials and Methods: LingShu construction integrates ontology design, heterogeneous data sources, preprocessing, entity alignment, verification, inference, and graph storage.The workflow includes atom–concept normalization and source-specific knowledge extraction.
  • 3.1 Ontology Design: The symptom-centered ontology standardizes entities, predicates, typed relation patterns, and contextualized relations for linking TCM concepts with biomedical mechanisms.Its comprehensive schema covers entity types, relation categories, and hyper-relational quadruple edges.
  • 3.1.1 Entity Types: 16 entity types span TCM and modern biomedicine, including symptoms, symptom clusters, herbs, prescriptions, diseases, drugs, genes or proteins, acupoints, and populations.Population entities encode disease-specific patient cohorts, while Gene entities connect clinical phenotypes, drugs, chemicals, herbs, and diseases.
  • 3.1.2 Triple Relations: 64 typed triple relation patterns are instantiated from 13 basic relations, with entity-type constraints separating predicate semantics from admissible head and tail types.The patterns cover therapeutic, causal, pharmacological, hierarchical, membership, and TCM-specific localization structures.
  • 3.1.3 Contextual Quadruple Relations: Contextual quadruples represent relations as (h, r, t, c), where c specifies the condition under which the relation holds.This model addresses context-dependent therapeutic efficacy, adverse effects, mechanistic associations, and clinical co-occurrence patterns.
  • 3.1.3 Contextual Quadruple Relations: 35 typed contextual quadruple patterns organize condition-aware treatment, condition-aware causation, and other contextual medical associations.Contexts include syndrome, disease, gene, therapeutic efficacy, chemical, population, and herb-related roles such as constituent, added, and removed herb.

3.2 Data Sources

LingShu uses the ISPO symptom ontology as its core and integrates heterogeneous TCM, biomedical, clinical, literature, and curated database sources. The repository includes 4,490 clinical electronic medical records, 1,541 classical TCM texts, 435 modern TCM books, 6 medical ontologies or terminologies, and 11 biomedical and TCM databases.

  • Repository composition: 4,490 clinical electronic medical records, 1,541 classical TCM texts, 435 modern TCM books, 6 medical ontologies or terminologies, and 11 biomedical and TCM databases comprise the integrated repository.These sources span clinical records, authoritative TCM materials, biomedical ontologies, and curated knowledge bases.
  • TCM textual sources: 1,541 classical TCM texts provide core TCM concepts and evidence for terminology variants, prescriptions, herb efficacies, syndromes, meridians, and acupoints.The corpus covers canonical theory, materia medica, formula composition, syndrome differentiation, acupuncture, disease etiology, and historical clinical cases.
  • TCM textual sources: 435 modern TCM books complement classical texts by standardizing terminology, defining symptom clusters, aligning traditional disease names with contemporary descriptions, and supplying clinical evidence.The sources include pharmacopoeias, national terminology standards, textbooks, clinical references, formula books, and case collections.
  • Biomedical terminology sources: 6 biomedical ontologies or terminologies supply controlled vocabularies, synonyms, identifiers, and cross-resource alignment anchors for symptom and disease normalization.ISPO serves as the core TCM-oriented symptom phenotype reference, while UMLS, MeSH, HPO, ICD-11, and SNOMED CT provide complementary biomedical normalization and classification support.
  • Structured databases: 11 biomedical and TCM databases provide curated herb–target, herb–symptom, drug–target, disease–gene, symptom–disease, and side-effect associations that are difficult to recover from free text alone.Records were normalized and retained when endpoints mapped to LingShu entity types and relation semantics fit predefined triple or contextual quadruple schemas.
  • Literature and clinical sources: PubMed supplements structured resources with sentence-level evidence for biomedical treatment, adverse-effect, and mechanism-oriented relations involving Disease, Symptom, Drug, and Gene entities.Clinical records contribute real-world symptom, syndrome, disease, herb, and population-context associations, including 1,788 COVID-19 and 2,702 cirrhosis records.

3.3 Knowledge Acquisition

LingShu acquired structured knowledge from TCM books, clinical electronic medical records, and biomedical literature using source-specific extraction strategies under a unified ontology. Concept normalization, human review, ontology-constrained extraction, and schema-based validation supported consistent integration of entities and contextual relations.

  • Cross-source concept normalization: LingShu normalized structured resources into type-specific Concept IDs, preserving source-specific atom records while merging semantically equivalent atoms into unified concepts.Extracted entities were mapped to existing concepts when possible; otherwise, new type-specific Concept IDs were assigned incrementally.
  • TCM book acquisition: TCM books were processed by an ontology-constrained LLM agent combining preprocessing, joint extraction, entity anchoring, type verification, and closed-loop validation.The agent used Qwen3.5-27B with predefined entity types, triple patterns, and contextual quadruple patterns.
  • Clinical record acquisition: Clinical electronic medical records were annotated with HCPSAS, whose model-assisted workflow identified fine-grained positive symptoms, negative symptoms, and symptom trajectories for human correction and review.Population-context knowledge aggregated patient-level herb–symptom co-occurrences within disease-specific cohorts characterized by demographic attributes.
  • Biomedical literature acquisition: Biomedical literature extraction targeted mechanism-oriented relations among Disease, Symptom, Drug, and Gene entities using PubMed evidence, distant supervision, dictionary matching, and LLM-assisted classification.The classification distinguished treatment, causation, and mechanism-mediated semantics under the typed contextual quadruple schema.

3.4 Terminology Normalization and Entity Alignment

LingShu normalized heterogeneous terms and aligned extracted mentions within a unified concept-level schema built on Atom–Concept identifiers. Type-specific rules, cross-source mappings, and layered matching methods supported provenance preservation, retrieval, reasoning, and cross-source integration.

  • Unified schema: Atom–Concept identifiers provided a unified concept-level schema for terminology normalization and entity alignment while preserving source-level provenance.The schema also supported concept-level retrieval, reasoning, and cross-source integration.
  • TCM normalization: TCM normalization used type-specific lexical and semantic rules, including compositional analysis of syndrome expressions and mapping of synonymous, historical, and alternative-translation forms.Syndrome structure typically followed etiology, pathological location, and pathological state; mappings used exact matching, rule-based synonym expansion, and concept validation.
  • Biomedical alignment: Biomedical alignment combined cross-source identifiers, curated mappings, and ontology-derived synonyms across disease resources including UMLS, MeSH, MalaCards, DisGeNET, ICD-11, Orphanet, and SymMap.Chinese–English synonym relationships expanded alignments without changing underlying concept identifiers, while analogous procedures covered chemicals, drugs, and genes.
  • Mention alignment: Extracted mentions were anchored to the normalized concept inventory using exact name matching, normalized name matching, and vector-based semantic retrieval.Normalized matching removed irrelevant modifiers, punctuation, and formatting differences, while vector retrieval handled unresolved mentions.

3.5 Post-verification and Quality Control

LingShu introduced ontology-guided post-verification after entity normalization and alignment to improve semantic validity, entity-type consistency, and schema compatibility. The procedure grounded entity and relation checks in source context and retained only ontology- and schema-consistent graph elements.

  • Procedure goals: Ontology-guided post-verification targeted semantic validity, entity-type consistency, and schema compatibility for extracted entities, triples, and contextual quadruples.It was applied after entity normalization and alignment rather than relying directly on raw extraction outputs.
  • Source-grounded preparation: Source-grounded preparation linked extracted entities and relations to original textual evidence and examined inconsistent concept assignments.This supported verification with contextual evidence rather than isolated entity names.
  • Entity-level semantic verification: Entity-level verification evaluated candidates against the LingShu ontology and normalized concept inventory, mapping matching candidates to Concept IDs.Ambiguous or previously unseen candidates were checked using source context for medical-entity validity and possible type correction.
  • Relation-level schema verification: Relation-level verification retained triples only when endpoints mapped to valid concepts and satisfied their typed relation patterns.For contextual quadruples, the head, tail, and contextual entities also had to satisfy the predefined contextual schema.
  • Quality-control outcome: Post-verification reduced extraction noise and based graph construction on semantically valid entities and ontology-consistent relations.This outcome followed the staged entity- and relation-level checks.

3.6 Logical Inference over Verified Relations

LingShu applied deterministic, ontology-defined logical inference after verification to improve contextual completeness and graph connectivity. The process bidirectionally derived schema-consistent contextual quadruples and triples, retaining only relations that satisfied typed ontology constraints.

  • Inference scope: Inference operated on concept-level triples and contextual quadruples after ontology-based validation, rather than during raw text extraction.The goal was to improve LingShu’s contextual completeness and graph connectivity.
  • Triple-to-quadruple inference: Verified triples were composed into contextual quadruples when compatible entities and predefined semantic patterns satisfied an admissible contextual schema.For example, Treats and ManifestationOf assertions could yield a syndrome-contextualized therapeutic quadruple.
  • Quadruple-to-triple inference: Selected contextual quadruples were projected into triples when their relations entailed valid binary relations under the ontology’s typed triple schema.Contextual treatment, gene-context drug, or disease-context drug assertions could support corresponding therapeutic, drug–symptom, or drug–disease triples when permitted.
  • Schema validation: All inferred relations were rechecked against typed triple and contextual quadruple schemas, retaining only those whose entity types met ontology constraints.This unified directly extracted, statistically derived, and structured database knowledge within a relation system containing triples and contextual quadruples.

3.7 Knowledge Graph Storage

LingShu stores verified knowledge in Neo4j using a concept-level Property Graph model that preserves provenance and supports retrieval, exploration, contextual querying, and reasoning. Typed triples use directed semantic edges, while contextual quadruples encode contextual entities as structured relationship attributes.

  • Concept-level graph construction: Verified knowledge was stored in Neo4j using the Property Graph model, with normalized concepts represented as nodes carrying identifiers, types, names, synonyms, and source mappings.Concept-level construction preserves provenance from heterogeneous sources while supporting concept-level retrieval and reasoning.
  • Relation representation: Typed triples connect head and tail concept nodes with directed edges labeled by semantic predicates constrained by ontology relation patterns.Contextual quadruples connect the same concept nodes while storing the contextual entity as structured relationship attributes, including its identifier, type, and role.
  • Evidence traceability: Stored relations retain data sources, source identifiers, and available text evidence, preserving multiple supporting records within a unified concept-level schema.The graph database supports Cypher retrieval, subgraph exploration, contextualized relation querying, and downstream reasoning over triples and contextualized relations.

4 Results

LingShu’s results show that multi-level verification substantially reduced extraction noise before integration into a large-scale hybrid graph spanning TCM and modern biomedicine. The resulting graph combines atom-level provenance with normalized concepts, typed triples, contextualized quadruples, and platform-based reasoning and exploration.

  • Book-scale verification: 914,156 unique entities were retained after verification, following removal of 253,007 of 1,167,163 candidate entities.The removal rate was 21.7%, with exclusions including non-medical expressions, vague references, incorrectly typed entities, and compound expressions requiring decomposition.
  • Book-scale verification: 10,839,759 triples and 622,340 contextual quadruples remained after relation-level verification of book-derived records.Verification removed records with unverified endpoints or schema-incompatible type combinations; contextual quadruples mainly captured prescription modification involving herb addition and removal.
  • Graph-wide coverage: 17.33 million atom-level entity records were normalized into 1,035,937 unified concepts across LingShu’s 16 entity types.The atom–concept design preserves source-level provenance while supporting concept-level retrieval, integration, and reasoning.
  • Graph-wide coverage: 17,185,769 verified triples and 22,286,936 contextualized quadruples were distributed across 64 typed triple patterns and 35 typed quadruple patterns.The quadruples encode contextual variables such as syndromes, therapeutic contexts, populations, and mechanisms, using extraction, clinical associations, literature extraction, and logical inference.
  • Graph-enabled use cases: The LingShu web platform supports graph visualization, querying, retrieval, reasoning, and evidence-grounded question answering for symptom-centred, context-dependent medical knowledge.Its reasoning scenarios include symptom-driven recommendations, disease-driven retrieval, and syndrome–prescription–herb mechanism exploration with candidate gene targets.

5 Discussion

LingShu bridges TCM and modern biomedicine through a unified, symptom-centric knowledge graph that combines large-scale heterogeneous data with contextualized relation representation. Its reasoning platform supports interpretable cross-paradigm exploration, while limited cohorts and unweighted relations remain important limitations.

  • Integrated knowledge graph: 17.33 million atom-level entity records, 17.19 million semantic triples, and 22.29 million contextualized quadruples integrate heterogeneous TCM and biomedical sources.Sources include clinical electronic medical records, classical and modern TCM books, and biomedical databases.
  • Integrated knowledge graph: Symptom-centric organization elevates symptoms to first-class entities, supporting TCM syndrome differentiation and modern phenotypic and mechanistic analysis.The architecture organizes the graph around symptom phenotypes rather than treating symptoms only as disease attributes.
  • Contextualized representation: Contextualized quadruples explicitly bind medical relations to conditions such as syndromes, molecular targets, and specific populations.The representation extends triples to (h, r, t, c), preventing syndrome constraints from being disconnected from treatment relations.
  • Reasoning and exploration: Graph-based reasoning and a RAG interface enable knowledge retrieval, candidate prioritization, and interpretable inference across symptoms, interventions, syndromes, and molecular targets.Global graph ranking and path-based semantic reasoning expose logical pathways connecting entities across TCM and biomedical domains.
  • Limitations: Current population-level associations rely on limited COVID-19 and cirrhosis cohorts, and relation strengths are not yet quantitatively weighted.The paper identifies expanded cohorts and probabilistic or quantitative relation weights as improvements for inference representativeness and precision.

6 Conclusion

LingShu is presented as a large-scale, symptom-centric contextualized knowledge graph integrating TCM and modern biomedicine through unified entity records and multimodal biomedical sources.

  • 6 Conclusion: 17.33 million atom-level entity records form LingShu’s unified symptom-centric representation linking TCM and modern biomedicine.The graph combines classical literature, clinical electronic medical records, and biomedical resources using a hyper-relational quadruple structure.

Data Quality

LingShu emphasizes data quality through atom-level provenance, traceable source identifiers, schema validation, and LLM-assisted verification of extracted knowledge against source text.

  • Provenance and verification: LingShu records atom-level provenance and preserves source database names and accession identifiers for traceability.For text-derived knowledge, candidate entities, triples, and contextualized quadruples undergo rule-based schema validation and LLM-assisted evidence verification against source snippets.
Loading 2608.20402v1…