Biomedical ontologies tend to model domain entities as classes, sometimes even such that have the ontological characteristics of instances (particulars) rather than true concepts (universals). Transformation of biomedical ontologies to an alternative representation, with classes meta-modeled using instances, may support their interoperability with linked data vocabularies/datasets and facilitate some reasoning-lean operations on them. We studied a set of best-practice patterns published at the Manchester ontology pattern portal and explored how they would be expressed in terms of instances and what are the implications.
The biomedical domain is probably the first in which description-logic (DL) reasoning
had been used at large scale, leveraging on resources such as SNOMED-OWL.2 Since
traditional DL reasoning tasks put emphasis on the ontology Tbox, the development of
biomedical ontologies is naturally prone to modeling domain entities as classes,
sometimes even for entities that have to some degree the characteristics of ontological
particulars (e.g., it is unlikely or impossible for them to have instances). On the other hand,
linked data (LD) vocabularies, a large amount of which are cataloged within the LOV
portal [
The idea of transforming the ontology Tbox to some kind of Abox representation
is not new. Specifically for SKOS, it has been addressed by the OWL2SKOS Converter
(and associated guidelines) developed by Abdul-Manaf [
Our analysis below is rather in the direction of the former approach. However, it does not primarily aim at taxonomies but at structures involving non-taxonomic properties.
Rather than examining particular ontologies randomly chosen from, e.g., those 800+
at the BioPortal,5 we focused on a library of ontology design patterns, which should,
ideally, represent the gist of how biomedical ontologies are designed (or, will be
designed in the future by modelers who subscribe to such ‘best-practice’ patterns). A
pattern library proposed primarily for the biomedical domain is the ‘Manchester’ one,
available at http://www.gong.manchester.ac.uk/odp/html/. It contains 17 patterns,
roughly categorized to three groups, Extension ODPs, Good Practice ODPs and
Domain Modelling ODPs. We examined the structure of the patterns and explored how
the same structure (leveraging on the accompanying examples) would be expressed in
terms of instance-based meta-modeling. While all patterns are in some way amenable to
transformation to an instance-based version, in this paper we focus on a family of three
patterns, Entity-Quality, Entity-Property-Quality and Entity-Feature-Values, which have
been jointly described in a paper by Egan˜a et al. [
An important question for our research is whether and how we can characterize the entities captured in biomedical ontologies as being inherently universals or particulars. This is a complex topic; in this short paper we only approximate it in pragmatic terms, omitting references to foundational ontology literature where such problems were studied.
Concepts in biomedical ontologies referring to objects that have (to some degree) clear boundaries and macroscopic size, such as organs or medical instruments, usually have a clear universal/particular distinction. This also holds for some processes, such as examinations or surgeries, which can be precisely anchored in time. As an example, the concept of heart is clearly different from a particular heart that instantiates it. A concrete surgery (e.g., a valve replacement), instantiating the general concept of such a kind of surgery, is then performed on a concrete heart. In a clinical context, both the universals and the particulars instantiating them may be worth representing in data. Thus, if we 4https://www.w3.org/TR/shacl/ 5https://bioportal.bioontology.org/ontologies decide to model the generic concept of heart, or of valve replacement, by an individual, this re-modeling should be viewed as true meta-modeling.
On the other hand, other or even the same ontologies may refer to processes that occur at microscopic level, where we do not routinely follow the trajectory of individual objects (such as cells or organelles). The likely purpose of ontologies is then to facilitate the description of generic biological mechanisms. In this situation it would be quite natural to represent the respective concepts as individuals. While their instances exist, these are not clinically relevant, and we would not process data on them in most applications (except, say, some sophisticated studies of rare events using high-definition microscopy). Furthermore, we also have biomedical concepts that refer to some kind of matter. In this case we do not have a clear notion of instantiation at all; whether a molecule (or some measurable amount) of a compound is an instance of this compound is a tough ontological issue. Again, it would be natural to represent the concept as an individual.
While these differences do not affect the way the classes from pattern-based structures are meta-modeled by individuals in our analysis, they will likely have an impact on the preservation of the ontological faithfulness, from the point of view of a human interpreting the resulting structure.
Egan˜a et al. [
The original diagram of the EQ pattern is in Fig. 1. We see that its structure contains three property restrictions, one equivalence axiom with a class union, and one disjointness axiom. All remaining boxes correspond to placeholders for named OWL classes. In the corresponding example diagram,6 entity n is substituted with apical complex, entity category with cell part, has quality with has position, quality value n with apical, and quality with position (only considering entities relevant for the biomedical knowledge statement describing the apical complex position).
The left-hand side of Fig. 2 shows the main part of the proposed meta-modeled version of EQ, namely, its structure of Abox triples, with individuals depicted as ovals. The ‘a’ edge correspons to the rdf:type property (as in the Turtle RDF serialization), and entity naming conventions common in LD vocabularies are applied: camel case and upper-initial naming. Finally, placeholder labels are in italics, while concrete entities such as the hasQuality property (invariant wrt. pattern instantiation), are in normal font. 6See http://odps.sourceforge.net/odp/html/Entity_Quality.html.
The triple-based structure conveys the given knowledge statement in simple terms; yet, there is no major information loss wrt. the original Tbox version of EQ, since the important axioms can be transformed alongside and kept within a secondary Tbox layer: The max-cardinality and universal axioms are left intact, since both EntityCategory and Quality are preserved as classes.
The exact-cardinality axiom becomes void: the ‘max 1’ feature is entailed by the max-cardinality defined for the EntityCategory, which EntityN instantiates, and the ‘min 1’ feature is entailed by the actual hasQuality property assertion. The equivalence axiom right-hand side is transformed from a concept union to an enumeration of individuals (though, with possible impact on the supported OWL profile); the disjointness axiom is then void, too.
EPQ and EFV7 only mildly differ from EQ. Instead of the max cardinality axiom they use domain and range axioms; however, the entities involved are unchanged by the meta-modeling transformation, and thus remain intact. The EPQ pattern meta-model has a property placeholder instead of the hasQuality property; the property substituted for this placeholder will however be a subproperty of hasQuality. The EFV pattern metamodel Abox core is in the right part of Fig. 2; in introduces the specific Feature entity allowing to group different aspects of the feature (such as the state and orientation of the position) together. All of this is straighforwardly meta-modeled.
7Their diagrams are in http://odps.sourceforge.net/odp/html/Entity_Property_Quality. html and http://odps.sourceforge.net/odp/html/Entity_Feature_Value.html, respectively. We do not include them here due to the space limitations.
In contrast to the above patterns, we will also briefly comment on the N-ary DataType Relationship pattern,8 which describes how a combination of data values can be assigned to a single entity. In the accompanying example, this entity is standard water boiling point, which is characterized by the pressure of 1 atmosphere and by the temperature of 100 degrees Celsius. Standard water boiling point, although modeled by a class in the original pattern, is however not an ontological universal but a particular. If we represent this entity by means of an individual and replace the restrictions on the (datatype) properties with mere property assertions, we yield a more ontologically faithful representation: it is thus not meta-modeling from the foundational point of view.
We made an early proof of concept in transforming ontology patterns to an alternative representation relying on individuals meta-modeling the original classes. We believe that such a study can be useful both for bridging between the biomedical ontology and LD vocabulary communities and for making some reasoning-lean operations over biomedical ontologies themselves more feasible. As a side effect, we can also obtain insights on the ontological faithfulness of pattern use cases along the universal-particular dimension.
We are currently finalizing the meta-modeling study for the whole Manchester catalog (17 patterns), and plan to extend this analysis, among other, to the lexico-syntactic patterns from the ontologydesignpatterns.org portal. We would also like to verify which of the Tbox axioms of the original patterns can be emulated by SHACL constraints. Furthermore, we also want to perform a qualitative study of the likelihood of biomedical ontology classes to be populated by instances, and, consequently, estimate which of them might deserve to be converted to (Abox) code lists easily reusable in the linked data context. Finally, we plan to focus on specific use cases, such as that of aligning and comparing fragments of biomedical ontologies, built using the design patterns in a classintensive style, with pre-existing instance-level code lists or ‘wild’ knowledge graphs. Acknowledgment: This research is supported by the IGA VSE project no. F4/33/2018.