If one chooses to follow the ontological approach, it soon becomes obvious that biomedical research almost exclusively deals with classes of biological entities, not with individuals. The statement "E. coli contains the LacZ gene" normally does not refer to a certain E. coli cell in a certain petri dish, it refers to E. coli—cells in general.

This fact can be problematic when we want to employ a knowledge representation language like OWL DL, which is a core standard of the Semantic Web development. OWL DL is based on description logics, which imposes strict restrictions on the ways in which classes and their instances ('individuals' in OWL terminology) can be used and related.

These restrictions do not pose a problem when an ontology commits to a certain system level of reality. For instance, the increasingly popular Biological Pathways Exchange ontology (BioPAX) employs a rather flat hierarchy of classes. Classes like 'Protein' or 'Pathway' as the most specialised classes in the ontology and are supposed to be directly instantiated. In one context, a molecular biologist might instantiate the class Protein with the individual 'Serotonin_Receptor', which is a clear enough description in many cases. In another context, a neuroscientist might instantiate the class Protein with the individual 'Serotonin_Receptor_2A'. Serotonin_Receptor_2A is in fact a subclass of Serotonin_Receptor, but this information cannot be represented in accordance to OWL DL semantics, since both have been defined as individuals.

The restrictions in the use of classes and individuals in OWL DL as compared to OWL Full are supposed to produce an advantageous trade-off between reasoning support and flexibility. However, in the case of biomedical ontologies on the Semantic Web, this trade-off seems less favourable. One of the main visions of the biomedical Semantic Web is the seamless integration of many different levels of description of biomedical reality. If one system level requires the use of instances, while another system level requires the use of classes, this integration is severely hampered by the restrictions of OWL DL.

The resulting ontologies are prone to become a tangled mixture of linguistic descriptions and realist ontological descriptions, which can easily lead to misconceptions, inconsistency and poor maintainability.

Matthias SAMWALD
Section on Medical Expert and Knowledge-Based Systems, Core Unit for Medical Statistics and Informatics, Medical University of Vienna, Vienna, Austria

Link to the full article (.pdf): http://neuroscientific.net/res/semsyn/samwald-classes-vs-instances.pdf