Biolink Model: Translational KG Schema

• Biolink Model is an open source, platform-agnostic data model that unifies clinical, biomedical, and translational knowledge via hierarchical classes and predicates.
• It standardizes identifier normalization and mapping with preferred namespaces and SKOS-based external ontology alignments to enable seamless integration.
• Implemented using LinkML, the model supports multi-format deployments—from RDF to property graphs—enhancing FAIR compliance and facilitating machine learning applications.

Biolink Model is an open source, platform-agnostic data model for representing clinical, biomedical, and translational knowledge in knowledge graphs. It was introduced as a universal schema intended to standardize heterogeneous graph-based resources whose integration had been impeded by ad hoc formats, inconsistent identifiers, sparse controlled-vocabulary usage, and the absence of a common open model. Its design combines object-oriented classification with graph-oriented features: a hierarchy of classes for biomedical entities, a complementary hierarchy of predicates for relationships, preferred identifier namespaces and ontology alignments, and a first-class association schema for provenance, evidence, and qualification. In this role, Biolink functions as an upper-level schema for translational biomedical knowledge graphs rather than as a domain ontology or a clinical data standard (Unni et al., 2022).

1. Motivation and problem setting

Biolink Model emerged in response to persistent barriers to knowledge discovery across biomedical knowledge graphs. The motivating problems identified include heterogeneous data formats such as tabular, RDF, and property graphs; inconsistent or sparse use of controlled vocabularies and identifiers; and the lack of a universally accepted, open-access model for standardization across biomedical knowledge graphs. These conditions forced downstream consumers to reconcile incompatible sources and undermined adherence to FAIR principles—findability, accessibility, interoperability, and reusability (Unni et al., 2022).

Within this setting, graph-based data models were already valuable because they elucidate interconnectedness between biomedical concepts, support easy updating of data structures, and enable intuitive queries, visualizations, and inference algorithms. Biolink addresses the gap not by replacing existing ontologies, but by formalizing relationships between data structures in translational science and by supplying a shared semantic layer across heterogeneous graph ecosystems.

A common misconception is to treat Biolink as a competitor to all biomedical schemas. The source material positions it more narrowly and more precisely. It is contrasted with OBO Foundry ontologies, SIO, schema.org, Bioschemas, and Wikidata’s ontology, and it is described as a unifying upper-level schema for translational knowledge graphs. Models such as FHIR, OMOP CDM, and BioPAX are explicitly noted as not discussed in the 2022 paper; the emphasis is on bridging heterogeneous knowledge graphs rather than aligning to clinical data standards per se (Unni et al., 2022).

2. Core representational architecture

At the core of Biolink Model is a hierarchical set of classes, also called categories, for foundational biomedical entities. The paper identifies exemplars including Disease, PhenotypicFeature, Gene, SequenceVariant, ChemicalSubstance, AnatomicalStructure, Pathway, and Publication. These classes are interconnected and organized hierarchically, allowing inheritance and mixins for cross-cutting concerns while preserving graph-oriented usability (Unni et al., 2022).

Biolink also defines a hierarchy of predicates for edges. All predicates descend from biolink:related_to, which enables polymorphic queries: a query written against a more general predicate can match edges asserted with more specific descendant predicates. Examples named in the source material include biolink:has_phenotype, biolink:positively_regulates, biolink:affects, and biolink:associated_with. This predicate hierarchy is central to the model’s support for coarse-grained querying over fine-grained assertions.

In addition to classes and predicates, the model defines node and edge properties. Node properties are intrinsic to instances of biolink:NamedThing and include examples such as biolink:id, biolink:name, and biolink:symbol. Edge properties are intrinsic to statements or edges and include examples such as biolink:publications and biolink:has_evidence. The resulting design is neither a bare triple model nor a simple property graph vocabulary; it is a typed schema in which nodes, predicates, and edge metadata are jointly constrained.

This architecture suggests that Biolink’s main abstraction is not merely “typed nodes and edges,” but a computable semantic envelope around graph assertions. The class hierarchy supports abstraction and reuse, while the predicate hierarchy supports subsumption and query polymorphism.

3. Association model, constraints, and identifiers

The central graph-oriented feature of Biolink Model is the association model. A Biolink Association is a first-class class representing a computable statement comprising a subject, predicate, and object—the core triple—plus structured edge attributes for provenance, evidence, and other qualifiers. The 2022 paper illustrates this pattern with biolink:DiseaseToPhenotypicFeatureAssociation, in which the subject is biolink:Disease, the object is biolink:PhenotypicFeature, and the predicate is biolink:has_phenotype; an edge property such as biolink:publications anchors the provenance trail (Unni et al., 2022).

Association classes carry domain and range constraints that define which kinds of entities can participate in a statement, along with slots that ensure consistent provenance and evidence capture across knowledge sources. The paper explicitly notes that, in OWL, Associations are akin to axioms, and in RDF they correspond to rdf:Statement. This makes the association the locus of edge-level metadata rather than an incidental annotation on a binary edge.

Biolink also makes CURIE handling and identifier normalization first-class concerns. Nodes are identified using IRIs or compact URIs drawn from preferred community ontologies, and the model defines id_prefixes within class descriptions to enumerate preferred identifier namespaces in priority order. The examples given include MONDO for diseases, HGNC or NCBIGene for genes, ChEBI or ChEMBL for chemicals, and HPO for phenotypic features. The Biomedical Data Translator’s Node Normalization Service and the Knowledge Graph Exchange framework consume Biolink’s prefix guidance and mappings to resolve multiple equivalent identifiers to a single preferred identifier, thereby harmonizing nodes across disparate sources (Unni et al., 2022).

A canonical pattern described in the source material is disease–has_phenotype–phenotype, for example MONDO:… biolink:has_phenotype HP:…, with the association carrying publications and evidence. Another concrete Translator example starts with NCBIGene:23221 (RHOBTB2) and traverses a path through biolink:entity_regulates_entity and biolink:genetically_interacts_with to biolink:Protein/Gene, then through biolink:related_to to biolink:SmallMolecule, returning fostamatinib disodium (CHEMBL.COMPOUND:CHEMBL3989516) and ruxolitinib (CHEMBL.COMPOUND:CHEMBL1789941) with supporting evidence.

4. Implementation in LinkML and interoperability mechanisms

Biolink is specified and implemented using the LinkML framework. This allows the schema to be authored in YAML and compiled into multiple artifacts for different implementation contexts: JSON-Schema for data validation, SQL-DDL for relational implementations, Python and Java classes for application integration, RDF for semantic web deployments, and UML diagrams for visual documentation (Unni et al., 2022).

This multi-artifact distribution means that the same model can underpin RDF triple stores and property graph databases such as Neo4j. The paper also states that model elements can be validated using standard toolchains, including JSON Schema validators and database constraints. LinkML-generated artifacts therefore provide a mechanism for enforcing slot usage and other schema-level constraints across heterogeneous infrastructure.

Alignment to external ontologies and vocabularies is another major interoperability mechanism. Where possible, each Biolink element is mapped to external terms using SKOS relations—exact, broad, narrow, close, or related—and the paper specifically references broad_mapping (skos:broadMatch). Biolink is informed by and interoperates with OBO Foundry ontologies, draws inspiration from SIO’s treatment of scientific data and knowledge, and complements schema.org and Bioschemas by providing a domain-specific upper-level schema with explicit graph constraints and evidence/provenance slots (Unni et al., 2022).

The practical implication is that Biolink does not seek interoperability through a single canonical ontology. Instead, it establishes a harmonization layer in which mappings, preferred identifier namespaces, and association semantics provide the computational basis for cross-graph integration.

5. FAIR alignment and deployment in translational knowledge graphs

Biolink advances FAIR compliance on several fronts. Findability and accessibility derive from open-source distribution, public documentation, and machine-readable multi-format artifacts. Interoperability is achieved through standardized node categories and edge predicates, enforced id_prefixes and identifier normalization, and SKOS-based mappings to external ontologies. Reusability follows from platform-agnostic design, generated code and schemas, and explicit evidence and provenance capture that make assertions auditable and reusable across systems (Unni et al., 2022).

The principal use cases emphasized in the 2022 paper are the NIH Biomedical Data Translator Consortium and the Monarch Initiative. Translator adopted Biolink as its upper-level common dialect so that participating knowledge sources could annotate nodes and edges to Biolink categories and predicates, enabling federation across property graphs and semantic web stores. Because predicates inherit from biolink:related_to and classes are hierarchical, Translator queries can be written at a coarser granularity while still returning results asserted with more specific predicates. Translator’s architecture is described as returning chemically typed candidates with evidence trails spanning multiple independent sources.

Similar harmonization and query capabilities underpin Monarch’s integrated view of genomic, phenomic, and biochemical data. The paper also identifies additional projects that depend on Biolink for reconciling heterogeneous identifiers, ontologies, and graph structures, including IDG, KG-COVID-19, ECO-KG, KG-ENVPOLYREG, and KG-Microbe (Unni et al., 2022).

These deployments establish Biolink’s role as an operational schema rather than a purely descriptive one. Its value lies not only in representational uniformity but in cross-KG querying, inference, and reconciliation over independently produced biomedical data.

6. Operationalization in KG-Hub and graph machine learning

KG-Hub places Biolink Model at the center of a platform for standardized construction, exchange, and reuse of biological knowledge graphs. In this setting, Biolink is operationalized throughout a modular extract–transform–load pattern, storage and versioning practices, and downstream graph machine learning tooling. KG-Hub’s ETL pattern is explicitly designed to produce Biolink-compliant property graphs (Caufield et al., 2023).

During transformation, Koza applies declarative templates that parse raw inputs into KGX TSV node and edge files while validating instances against Biolink typing rules. The paper gives a concrete example: Koza raises an error if a Human Phenotype Ontology identifier such as HP:0002362 (“Shuffling gait”) is erroneously typed as biolink:Gene. KGX automatically assigns Biolink categories based on namespaces and curated mappings in Biolink, and KGX merges subgraphs according to a merge.yaml configuration, performing identifier normalization across sources while preserving Biolink typing.

KG-Hub further integrates OBO Foundry ontologies through KG-OBO, which transforms ontologies into JSON and KG-compatible KGX formats and maintains multiple versions across 201 ontologies. Common ontologies cited include Gene Ontology, Mondo, HPO, and ChEBI. ROBOT can be used as an interlocutor to import, filter, and reconcile ontology content before representing it in the graph. This workflow makes Biolink the organizing schema through which ontology-derived and instance-derived knowledge are merged.

The paper reports seven biomedical KG projects and gives representative sizes and use cases. KG-COVID-19 is reported at 574k nodes and 24,145k edges; KG-IDG at 560k nodes and 4,431k edges; KG-Microbe at 276k nodes; KG-Monarch at 794k nodes and 6,970k edges; and KG-Phenio at 275k nodes and 1,183k edges. KG-IDG uses Koza transforms and KGX merges, then applies graph machine learning through NEAT and GRAPE on Biolink-typed subgraphs. A supplemental NEAT YAML uses Biolink categories for node types and a default edge predicate of biolink:related_to to compute embeddings and train a logistic regression classifier for protein–protein link prediction (Caufield et al., 2023).

This suggests that Biolink’s contribution extends beyond interoperability into experimental reproducibility and machine learning task specification. Standardized categories and predicates permit coherent subgraph selection, stable feature construction, and cross-project reuse of transformed subgraphs via permanent URLs and versioned artifacts.

7. Community governance, evolution, and use in ontology-constrained extraction

Biolink is described as open-source, publicly accessible, and under active development with frequent releases. The project maintains a GitHub repository for the schema and generated artifacts, comprehensive documentation, and a public issue tracker, and it invites contributions from biologists, clinicians, data curators, developers, and ontologists. The 2022 paper presents this collaborative governance model as a reason Biolink has evolved in response to real-world integration needs (Unni et al., 2022).

The same paper also acknowledges a persistent challenge: multiple ontologies may exist for the same domain concept, making identifier management and mappings an ongoing issue. Biolink’s id_prefixes, SKOS mappings, and association schema are presented as practical mitigation mechanisms. Continued expansion of mappings, refinement of constraints, and deeper provenance and evidence modeling are implied directions.

A later development shows Biolink being used not only to normalize existing graphs but also to constrain knowledge extraction from unstructured text. RELATE uses the Biolink Model as a target schema for mapping free-text biomedical relations into standardized ontology predicates. Its three-stage pipeline includes ontology preprocessing with predicate embeddings, similarity-based retrieval enhanced with SapBERT, and LLM-based reranking with explicit negation handling. On ChemProt, the paper reports 52% exact match and 94% accuracy@10; in 2,400 HEAL Project abstracts, it rejected 10 relationships, approximately 0.4%, as unsuitable for Biolink predicate mapping and flagged 77 relationships as negated assertions (Olasunkanmi et al., 23 Sep 2025).

RELATE’s use of Biolink is notable for two reasons. First, it operationalizes Biolink’s predicate hierarchy and association semantics in a literature-mining context, including use of predicates such as biolink:related_to, biolink:coexists_with, biolink:affects, biolink:causes, biolink:located_in, biolink:part_of, biolink:temporally_related_to, biolink:precedes, biolink:has_part, and biolink:has_output. Second, it treats negation as a qualifier rather than as a pseudo-predicate, preserving the ontology while capturing polarity. A plausible implication is that Biolink’s association-centric design is sufficiently expressive to support not only graph integration and exchange, but also ontology-constrained graph augmentation from biomedical abstracts.

Taken together, these developments position Biolink Model as a computable blueprint for harmonizing biomedical knowledge across heterogeneous graphs, ETL systems, and text-derived relation extraction pipelines. Its class and predicate hierarchies, association-based edge semantics, LinkML implementation, identifier normalization, and external mappings make it a general semantic layer for translational biomedical knowledge graph infrastructure (Unni et al., 2022).