Ontology

Definition and scope

The word ontology has two lives, and a reader who meets it in a philosophy seminar and then in a software specification can be forgiven for suspecting a coincidence of spelling. The uses are related but not the same, and this publication keeps them apart.

In philosophy, ontology is the study of what there is: which kinds of things exist, what it is for a thing to be the kind of thing it is, and how kinds and individuals depend on one another. The question is older than the word. Aristotle's Categories asks what can be said of a subject (a substance, a quantity, a quality, a relation, a place, a time) and his Metaphysics asks what it is to be at all. The Latin coinage ontologia appears in the early seventeenth century, and Christian Wolff gave it a settled place in the curriculum in 1730 with his Philosophia Prima sive Ontologia. In the twentieth century Quine reframed the question as one about commitment: to say what exists is to say what the variables of our best theory range over.

In informatics, an Ontology is a written artifact. The most quoted definition is Gruber's from 1993: an explicit specification of a conceptualization. Guarino and colleagues sharpened this in 2009: a formal, explicit specification of a shared conceptualization, where formal means machine readable, explicit means the types and constraints are stated rather than assumed, and shared means that a community has agreed to them for a purpose. The artifact lists the kinds of things a domain recognizes (its classes), the ways those things can be related (its relations), the constraints among them, and sometimes the particular things themselves (its instances). A computer can then check whether a set of statements is consistent with the ontology and draw conclusions that follow from it.

This monograph is about both, in that order. The scope ends where Epistemology begins: ontology asks what there is, epistemology asks how we know. It also stops short of Taxonomy, one restricted way of organizing kinds, and of the standards that encode formal ontologies, treated under OWL and in the foundations section.

Key distinctions

Philosophical versus formal ontology

The philosopher asks whether numbers exist, whether a disease is a thing or a process, whether a patient's tumor at two time points is one entity or two. The informatician usually does not settle those questions. She chooses a position for the purpose of a system and writes it down. Barry Smith, who has worked on both sides, describes the difference as one between ontology as a discipline and ontology as an artifact. The artifact inherits its vocabulary (entity, class, relation, instance) from the discipline, but it is a built thing, with a version number, an author, and a scope.

Entities and relations

An entity is anything that can be the subject of a statement: a molecule, a procedure, a person, a measurement, a time interval. A Relationship is a stated connection between entities, with a type and usually a direction. "Agent A binds Target P" names two entities and one directed relation. An ontology that recognizes molecules and targets but has no relation between them cannot say anything about binding.

Classes and instances

A class is a kind, a category under which many things can fall. An instance is a particular thing that falls under a class. "Radiopharmaceutical" is a class. The contents of a particular syringe prepared on a particular morning is an instance of it. The distinction sounds trivial and causes more modeling errors than any other, because the same word often does double duty. "Lutetium-177" can name a kind of atom (a class, every atom with 71 protons and 106 neutrons) or the material in a given vial (an instance, a particular quantity of that kind). An ontology must decide which it means, and a mature one will often need both.

A vocabulary is not an ontology

A list of approved names, even a well-governed one with identifiers and definitions, is a Controlled vocabulary. It becomes an ontology only when it states what kinds its terms denote and how those kinds relate, in a form that supports inference. Rector's 1999 account of why clinical terminology is hard turns on this point: a terminology that merely names things cannot say whether "fracture of femur" and "femoral fracture" are the same, whether one implies "injury," or whether "fracture of left femur" is a kind of "fracture of femur." These are ontological questions, and answering them requires more than a list.

Historical development

Aristotle's Categories (fourth century BCE) is the earliest surviving attempt to say systematically what kinds of predicate can be applied to a subject. His distinction between primary substance (this particular horse) and secondary substance (the species horse) is the ancestor of the instance and class distinction that every formal ontology still uses. The Metaphysics adds the question of what being is in general, and its treatment of genus and species shaped the European curriculum for nearly two thousand years.

The word itself is a product of early modern school philosophy. It appears in Latin in the first decade of the seventeenth century, and Wolff's 1730 Ontologia made it the standard name for the part of metaphysics that treats being in general, as distinct from the parts that treat God, the soul, and the world. Kant's criticism of that program did not kill the question; it changed its form.

Two twentieth-century developments matter here. The first is Heidegger's insistence in Being and Time (1927) that what a thing is cannot be separated from how it shows up within a practice that already understands it: a scanner, a tracer, a report exist for the physician as equipment within a task before they are catalogued as objects. We cite this not to borrow a vocabulary but to mark a real problem for formal ontology: the categories we write down are shaped by the practices we write them for, and a different practice will cut the world differently. The second is Quine's 1948 "On what there is," which replaced the search for a neutral inventory with a criterion of commitment: we are committed to the entities our accepted theories quantify over. For a formal ontology this is a liberation. It need not describe the furniture of the universe, only make explicit what a community is committed to when it accepts a body of statements.

Formal ontology as an engineering discipline is younger. Knowledge representation research in the 1970s and 1980s produced frame systems and semantic networks, and the need to share knowledge bases between projects created the demand for a portable specification of what a knowledge base assumes, which Gruber's 1993 paper answered. The 1990s and 2000s brought description logics as a formal foundation (see Description logics), the Web Ontology Language as a standard encoding, and large biomedical ontologies such as the Gene Ontology and the Foundational Model of Anatomy as evidence that the approach scales. Smith and Ceusters' 2010 program of ontological realism, discussed below, argued that biomedical ontologies should describe entities in reality rather than concepts in people's heads.

Philosophical or technical account

What a formal ontology contains

A formal ontology, at minimum, declares classes, relations, and constraints. Consider the smallest useful fragment.

Classes: Radiopharmaceutical, Radionuclide, Ligand, Molecular target, Administration.

Relations: has radionuclide (from Radiopharmaceutical to Radionuclide), has ligand (from Radiopharmaceutical to Ligand), binds (from Ligand to Molecular target), administers (from Administration to Radiopharmaceutical).

Constraints: every Radiopharmaceutical has exactly one radionuclide; every Administration administers exactly one Radiopharmaceutical; Radionuclide and Ligand are disjoint, so nothing is both.

With these declarations, a machine can already do things a list cannot. Told that an item is a Radiopharmaceutical with two radionuclides, it reports a contradiction. Told that an Administration administered item X, it infers that X is a Radiopharmaceutical even if no one said so. None of this is deep reasoning, but all of it depends on the ontology saying what kinds there are and how they relate.

Classes, instances, and the two readings of a name

Formal ontologies distinguish the class level (what is true of every member of a kind) from the instance level (what is true of a particular thing). Description logics make the distinction precise with a terminological component (the TBox, holding class axioms) and an assertional component (the ABox, holding facts about individuals).

A recurring difficulty is that domain experts name both levels with the same word. "The patient received Lu-177" names a kind. "This vial of Lu-177 was calibrated at 10:00" names an instance. A good ontology resolves the ambiguity by naming both: a class Lutetium-177 atom and, where needed, instances such as the quantity of Lutetium-177 in vial NX-T-0042, where NX-T-0042 is a local teaching identifier with no claim to correspond to any real product or code. A reference terminology mostly needs classes. A record of what was done to whom mostly needs instances. An ontology that serves both must carry both and keep them apart.

Relations have types and directions

"Related to" is not a relation an ontology can use. A relation must be typed (binds, is part of, occurs before) and, for most types, directed. "Agent A binds Target P" does not entail "Target P binds Agent A" unless the ontology says that binds is symmetric, and whether it should say so is a modeling decision with consequences. Relation types can themselves be organized: is located in may be declared transitive, so that if a lesion is in a lymph node and the node is in the mediastinum, the lesion is in the mediastinum. Each such declaration is a commitment, and each can be wrong for a particular use.

Identity and change

An ontology must also decide what makes something the same thing over time. Is the tumor on a baseline scan and the tumor on a follow-up scan one entity observed twice, or two entities in a relation of continuity? Is a radiopharmaceutical the same product after a change of manufacturing site? There is no neutral answer; there are answers that serve purposes, and they can coexist if their scopes are stated.

Biomedical relevance

Biomedicine was the first large field to build formal ontologies at scale, not because medicine is unusually orderly but because it is the opposite. Clinical language uses the same word for a disease and its manifestation, for a procedure and its result, for a drug substance and a product that contains it. A hospital system that stores "pneumonia" needs to know whether a query for "lung disease" should find it, whether "pneumonia of left lower lobe" should count, and whether a record coded before a classification change still means the same thing. A controlled vocabulary gives each of these a name. Only an ontology can say how they relate.

SNOMED CT, the largest clinical terminology in routine use, illustrates the transition. Its logical model consists of concepts, descriptions, and relationships, and the relationships include both the hierarchy (is a) and defining attributes such as finding site and associated morphology, expressed in a description logic so that a concept's position in the hierarchy can be computed from its definition rather than asserted by hand. Whether this makes SNOMED CT an ontology in the full sense is taken up in SNOMED CT and the NucLex niche. The point here is that a terminology of that size could not be maintained without ontological machinery.

The Gene Ontology and the Foundational Model of Anatomy show the other direction: ontologies built first as accounts of a domain and only secondarily as sources of codes. Their experience produced the methodology in Arp, Smith, and Spear's 2015 handbook: start from an upper ontology that fixes the most general kinds, define each domain class by genus and differentia, type and direct every relation, and keep classes separate from the terms that name them. The methodology is contested, but it is the most explicit account available of what building a biomedical ontology involves.

Nuclear medicine relevance

Everything in this section is a synthetic example. The agents, targets, and identifiers are invented for teaching. Identifiers of the form NX-T-nnnn are local and correspond to no product, code system, or regulatory record.

A department keeps a list of the radiopharmaceuticals it uses:

• NX-T-0001, "Agent A," a diagnostic agent labeled with fluorine-18.
• NX-T-0002, "Agent B," a therapeutic agent labeled with lutetium-177.
• NX-T-0003, "Agent C," a diagnostic agent labeled with gallium-68.

The list is a vocabulary. It has identifiers, labels, and (we may suppose) definitions and a person responsible for changes. It can be used to code a record. Now ask the questions a physician, a radiochemist, and an informatician actually ask.

The physician asks: which agents in our list bind the same target as Agent B, so that a diagnostic scan can select patients for the therapy? The list cannot answer. It does not know what a target is.

The radiochemist asks: if we change the supplier of the precursor for Agent B, is it still Agent B? The list cannot answer. It has no account of what makes an entry the entry it is; it has only a label.

The informatician asks: a record was coded NX-T-0002, but the administered activity field is empty. Is that a possible record? The list cannot answer. It has no constraints.

Now rebuild the list as an ontology fragment.

Classes: Radiopharmaceutical; Radionuclide with subclasses Fluorine-18, Lutetium-177, Gallium-68; Ligand; Molecular target; Administration with subclasses Diagnostic administration and Therapeutic administration.

Relations: has radionuclide, has ligand, binds (Ligand to Molecular target), administers, has route, has administered activity.

Constraints: a Radiopharmaceutical has exactly one radionuclide and exactly one ligand; a Therapeutic administration administers only Radiopharmaceuticals whose radionuclide is a therapeutic radionuclide (a subclass we declare to contain Lutetium-177 for the purposes of the example); every Administration has exactly one route and exactly one administered activity.

Individuals: NX-T-0001 has radionuclide Fluorine-18, has ligand Ligand L1, and Ligand L1 binds Target P. NX-T-0002 has radionuclide Lutetium-177, has ligand Ligand L2, and Ligand L2 binds Target P. NX-T-0003 has radionuclide Gallium-68, has ligand Ligand L3, and Ligand L3 binds Target Q (a second teaching target).

The physician's question is now answerable by a query over the relations: find every Radiopharmaceutical whose ligand binds the same target as the ligand of NX-T-0002. The answer is NX-T-0001. Nothing has been asserted that a human did not already know, but it has been stated once, in a form a machine can traverse, rather than rediscovered in every reading room.

The radiochemist's question can now at least be posed precisely. Agent B is defined by its radionuclide and its ligand. If the ligand is the same molecular entity after the change of supplier, the ontology says the product is the same radiopharmaceutical; if the ontology also carries a class for the formulation or manufacturing process, it can distinguish same substance from same product. The modeling choice is visible, and so is its consequence.

The informatician's record is now flagged. A Therapeutic administration must have an administered activity, and a constraint on route (which we would add) must be satisfied. The ontology did not make the record correct; it made the record's incorrectness detectable.

The instance and class distinction surfaces here too. "Lutetium-177" above is a class of radionuclide. The particular lutetium-177 in the vial from which NX-T-0002 was drawn on a given morning is an instance, and its activity at calibration time is a quality of that instance, not of the class. A system that conflates them will either attribute an activity to a nuclide kind (meaningless) or lose the fact that two administrations drew on the same vial (sometimes important).

What this example does not do matters as much. It does not say whether Agent A is a good selection tool for Agent B; that is a clinical question answered by evidence. It does not say what Target P is in reality, or that NX-T-0002 corresponds to any real product. It states what kinds the department is committed to and how they relate, so that further questions can be asked precisely. That is what an ontology is for.

Disagreements and limitations

Realism versus conceptualism. Smith and Ceusters argue that terms in a biomedical ontology should refer to universals instantiated in reality, not to concepts in anyone's mind. Critics reply that many clinical entities (a syndrome, a stage, a risk category) are defined by convention and have no counterpart in reality independent of the practice that defines them. The disagreement is not academic. It determines whether an ontology may contain a class for "suspected metastasis" (a state of a reader's judgment) or only for "metastasis" (a state of a patient), and therefore what a record can say.

One ontology or many. The ambition of a single shared ontology conflicts with Heidegger's and Quine's observation, in their different ways, that categories are shaped by practices and theories. A reading room, a radiopharmacy, and a billing office cut the same events differently. The realist response is that reality is one and the cuts can be reconciled; the pragmatic response is to build several ontologies with stated scopes and map between them. NucLex's position, developed in the tools section, is the second, with the first as a regulative ideal.

The limits of formalization. A formal ontology captures what can be stated as classes, relations, and constraints. It does not capture the skill that lets an experienced reader know that an image is artifactual, nor a department's tacit agreement about what "equivocal" means this month. Formalization can make such things visible by failing to capture them, but it cannot replace them. A complete ontology is not a complete understanding.

Scope of this draft. This monograph has not been source-checked. The claims about the coinage of ontologia and about Wolff should be verified against the primary texts, and the characterization of SNOMED CT's logical model against the current release documentation. The synthetic example uses no real product and makes no clinical claim.

Related entries

• Ontology, the lexicon entry with a compact definition and relationships.
• Concept, for the distinction between a concept and the terms that name it.
• Relationship, for the requirement that relations have type and direction.
• Epistemology, for the question of how we know what we claim.
• Taxonomy, for hierarchy as one restricted kind of ontology.
• Description logics, for the formal machinery that makes class and relation axioms computable.
• OWL, for the standard language in which formal ontologies are published.
• Identifier, for why a local teaching identifier is not an external code.

References

1. Aristotle. Categories. In: Barnes J, ed. The Complete Works of Aristotle: The Revised Oxford Translation. Vol 1. Princeton University Press; 1984.
2. Aristotle. Metaphysics. In: Barnes J, ed. The Complete Works of Aristotle: The Revised Oxford Translation. Vol 2. Princeton University Press; 1984.
3. Wolff C. Philosophia Prima sive Ontologia. Frankfurt and Leipzig; 1730.
4. Heidegger M. Being and Time. Macquarrie J, Robinson E, trans. Blackwell; 1962.
5. Quine WV. On what there is. Review of Metaphysics. 1948;2:21-38. Reprinted in: Quine WV. From a Logical Point of View. Harvard University Press; 1953.
6. Gruber TR. A translation approach to portable ontology specifications. Knowledge Acquisition. 1993;5(2):199-220.
7. Guarino N, Oberle D, Staab S. What is an ontology? In: Staab S, Studer R, eds. Handbook on Ontologies. 2nd ed. Springer; 2009:1-17.
8. Smith B. Ontology. In: Floridi L, ed. The Blackwell Guide to the Philosophy of Computing and Information. Blackwell; 2004:155-166.
9. Smith B, Ceusters W. Ontological realism: a methodology for coordinated evolution of scientific ontologies. Applied Ontology. 2010;5(3-4):139-188.
10. Arp R, Smith B, Spear AD. Building Ontologies with Basic Formal Ontology. MIT Press; 2015.
11. Rector AL. Clinical terminology: why is it so hard? Methods of Information in Medicine. 1999;38(4-5):239-252.
12. W3C. OWL 2 Web Ontology Language Document Overview (Second Edition). W3C Recommendation, 11 December 2012. https://www.w3.org/TR/owl-overview/ (access checked 10 October 2026).
13. SNOMED International. SNOMED CT logical model. SNOMED CT Starter Guide. https://docs.snomed.org/snomed-ct-practical-guides/snomed-ct-starter-guide/5-snomed-ct-logical-model (access checked 10 October 2026).