Ontologies

An ontology is a controlled vocabulary of defined terms, organised in a hierarchy so that data annotated by different groups can be compared and queried together. Ontologies describe what a dataset records about the world, its anatomy, cell types, molecular functions, phenotypes, behaviours, and diseases, as distinct from the metadata that describes the dataset as an object (covered in Data Discoverability). Most of the ontologies relevant to neuroscience are coordinated by the OBO Foundry, which sets shared design principles so the separate ontologies interoperate rather than overlap, each covering a distinct slice of the description problem and connecting to its neighbours. Several of the disease and phenotype ontologies are produced by the Monarch Initiative, a consortium that integrates gene, disease, and phenotype data across species and develops the ontologies that make that integration possible. The result is a division of labour across what a neuroscience dataset needs to describe: anatomy, cell type, molecular function, phenotype, behaviour, and disease.

Vocabularies, ontologies, and data models

Vocabulary, thesaurus, ontology, and data model are related terms that are often used loosely and interchangeably, though each adds a distinct layer of structure on top of the last. A vocabulary is an agreed list of approved terms, so a dataset consistently records jaguar rather than panther for the same animal. A thesaurus adds a little structure on top, recording that jaguar sits under big cat and is related to Panthera, without saying exactly how. An ontology says how: a jaguar is a kind of Panthera and preys on capybara. Because each link carries a stated meaning, software can follow it and reason over it, which is what the located in, performs, and acts on relationships described below actually are. A data model is a separate layer again, a form with fields that must be filled from an approved vocabulary, such as a species field that only accepts a listed term. Metadata is that form filled in for one dataset, species = jaguar. The data model supplies the blank fields, vocabularies and ontologies supply the words allowed in them, and metadata is the finished description.

In practice these categories blur. SNOMED CT has enough internal structure to be called an ontology by some, while MeSH is built as a thesaurus but is often used as a plain vocabulary.

graph TD
  vocab(["vocabulary<br/>approved term: <code>jaguar</code><br/>not <code>panther</code>"])
  thes(["thesaurus<br/>broader: <code>big cat</code><br/>related: <code>Panthera</code>"])
  onto(["ontology<br/><code>jaguar</code> <code>is-a</code> <code>Panthera</code><br/><code>preys-on</code> <code>capybara</code>"])
  vocab -->|adds structure| thes -->|adds structure| onto
  onto -->|supplies terms to| dm(["data model<br/>required: <code>species</code>"])
  dm -->|becomes| meta(["metadata<br/><code>species</code> = <code>jaguar</code>"])

Each layer adds structure: a vocabulary of approved terms, a thesaurus of broader and related links, an ontology of typed relations, a data model of required fields, and metadata as the filled-in result.

From biology to the clinic

Biology gives rise to observable traits, traits are grouped into diseases, and a disease is recorded in a patient’s record as a diagnosis code. Each step is a distinct relationship, a trait is a manifestation of biology, a disease is a grouping of traits, a diagnosis code is a record of a disease, and the standard naming one step maps to the standard naming the next. This is what lets a single disease be matched from the genetics that underlie it to the diagnosis code a clinician enters.

graph TD
  bio([biology<br/>cells, anatomy, molecules]) -->|manifests as| trait([traits<br/>phenotypes, behaviour])
  trait -->|grouped into| disease([disease<br/>one identifier])
  disease -->|recorded as| record([clinical record<br/>a coded diagnosis])

From biological substrate to a clinical diagnosis code, each arrow a distinct kind of relationship.

Anatomy, cell type, and function

For anatomy, UBERON provides cross-species brain region terms, so a structure can be named once and matched across human, mouse, and other model species, and it is used across EBRAINS, NWB, and the Allen Institute for Brain Science to enable search by anatomical location independent of species-specific naming. For cell type, Cell Ontology (CL) covers neuronal and glial classes down to specific subtypes and is required for single-cell data annotation in CELLxGENE and BICAN. For molecular function, GO (Gene Ontology) describes the biological processes, molecular functions, and cellular components of gene products, and draws on ChEBI for the chemical entities, drugs, and neurotransmitters those functions involve.

graph TD
  cl([Cell Ontology<br/>cell type])
  uberon([UBERON<br/>anatomy])
  go([GO<br/>molecular function])
  chebi([ChEBI<br/>chemical entities])

  cl -->|located in| uberon
  cl -->|performs| go
  go -->|acts on| chebi

Cell type, anatomy, molecular function, and chemistry: each ontology covers one layer and links to its neighbours.

Phenotype and behaviour

Phenotype is described by a paired set of ontologies split by species. HPO (Human Phenotype Ontology) is the standard for human clinical phenotypes and is widely used in rare disease genomics, while MP (Mammalian Phenotype Ontology) is its model-organism counterpart for mouse, rat, and other mammalian phenotypes. The two co-maintain a published cross-species mapping through the Mouse-Human Ontology Mapping Initiative (MHMI), so a phenotype observed in a mouse model can be matched to the corresponding human disease phenotype for model discovery and disease-gene prediction. uPheno (Unified Phenotype Ontology) builds on these mappings to provide a single species-neutral layer under which the human, mouse, and other species-specific phenotype terms are grouped, so cross-species queries run against one vocabulary rather than a set of pairwise mappings. NBO (Neurobehavior Ontology) complements both, covering behavioural and neurological phenotypes (motor, cognitive, sensory, affective, and circadian) in humans and model organisms alike.

graph TD
  hpo([HPO<br/>human phenotype])
  mp([MP<br/>mouse, rat<br/>phenotype])
  nbo([NBO<br/>behaviour])

  hpo <-->|co-maintained mapping| mp
  nbo -->|describes behaviour for| hpo
  nbo -->|describes behaviour for| mp

Human and model-organism phenotype ontologies, linked by a co-maintained cross-species mapping.

Disease

Disease is described separately from phenotype. MONDO (Monarch Disease Ontology) harmonises the major disease classifications, integrating ICD-10, ICD-11, OMIM, and ORDO into a single hierarchy with a unified identifier per disease that maps to the corresponding code in every source system, which makes a disease queryable across databases that each use a different classification. ORDO (Orphanet Rare Disease Ontology) provides the rare-disease backbone MONDO builds on, and OMIM supplies the curated gene-disease relationships.

graph BT
  omim([OMIM<br/>Mendelian genetics]) -->|maps to| mondo([MONDO<br/>one research disease ID])
  ordo([ORDO<br/>rare disease]) -->|maps to| mondo
  icd([ICD-10 / ICD-11<br/>clinical coding]) -->|maps to| mondo
  mesh([MeSH<br/>literature]) -->|maps to| mondo

Four disease classification systems mapped onto one unified research disease identifier.

The clinical terminologies that code diagnoses, procedures, and observations in health and trial data (SNOMED CT, ICD-10, LOINC, MeSH, and the drug vocabularies) are covered in the Health and Clinical Trials perspectives, since their primary role is operational coding rather than research annotation.

For the phenotyping standards and variant curation infrastructure that build on these phenotype and disease ontologies, see Rare Disease and Phenotyping. For the metadata, identifiers, and registries that describe and locate the dataset itself, see Data Discoverability.