This paper reports our experience with OWL for the Foundational Model of Anatomy (FMA). We show that converting the FMA from Protégé into OWL DL was possible, with most features of the original FMA captured. The conversion relies on translation and enrichment rules, implemented with flexible options. Unsurprisingly, reasoning with OWL proved to be a real challenge, due to the sheer size and complexity of the FMA. As the entire FMA in OWL DL raised inference problems hard to solve in terms of time and memory, an incremental approach was adopted. A number of various smaller versions that Racer could handle were successfully tested. Some inconsistencies were identified and some classes reclassified. The analysis of the results obtained so far shows the benefits of representing the FMA in OWL and, more generally, the usefulness of DLs reasoning techniques for large-scale biomedical ontologies shared on the Web.
As OWL is now the W3C recommended standard for ontologies, converting
framebased ontologies to OWL becomes an important need. Once converted to OWL,
ontologies currently developed with frames become virtually interoperable with other
ontologies and can be used as resources for the Semantic Web. Also of interest is
OWL higher expressiveness and the powerful reasoning services associated to its
underlying description logic. There is now a large and increasing library of ontologies
developed in Protégé [
The method used to automatically convert the FMA from Protégé 2.1 into OWL DL is first presented (§2). The experience of reasoning with OWL and its results are next reported (§3). The choices of conversion and future perspectives for the FMA are then discussed. Lessons learnt from this experiment may be generalized for largescale ontologies of the Semantic Web (§4).
As DLs and frames share the same object paradigm, it might be thought that converting a Protégé ontology into OWL is straightforward and could be achieved by a simple export function mapping Protégé primitives to OWL constructs. But, the export function from Protégé to OWL did not work for the FMA, neither in one step (i.e., directly), nor in two steps (i.e., from database to text then to OWL). Besides, even if it had worked, it would be ineffective, mainly for two reasons:
First, migrating a frame-based ontology to OWL requires not only a syntactic “translation”, but also a semantic “enrichment”. Indeed, property restrictions such as allValuesFrom or someValuesFrom contained in the OWL axioms cannot be directly derived from the original frame representation, where they are not specified. Additionally, classification strongly relies on the classes logical definitions. A reasoner (e.g., Racer) can only automatically classify classes under “defined” classes1 – i.e., classes with at least one necessary and sufficient condition. Necessary and 1 except if a property has a domain (or range) that is a primitive class, which can coerce classes to be reclassified under the primitive class that is the domain or range of the property (§4). sufficient conditions cannot be derived directly from the Protégé model, because in frames, all slots at class with a specified range or value, are considered as a set of necessary conditions. Specifying defined classes is a major “enrichment” of the ontology.
The second reason is that the FMA in Protégé makes an extensive use of
metaclasses2, which are not allowed in OWL DL. Each anatomical entity is modeled
both as a metaclass and as an instance of a metaclass. This was the “technical solution
for enabling the selective inheritance of attributes” in Protégé [
Therefore, we defined our own method of conversion, which aims at providing the desired enrichments and at capturing the knowledge encoded at metaclasses differently.
The migration was achieved from the CLIPS files. The conversion relies on translation and enrichment rules, implemented with flexible options (for conversion rules see the annex available at http://www.med.univ-rennes1.fr/lim/doc_116.pdf).
Translation draws on the structural correspondence between Protégé and OWL constructs. The Protégé class taxonomy defined at meta level is translated into an OWL subclass hierarchy. Template slots defined at the top level are translated into OWL properties with the same features as those specified in Protégé, i.e. same range, inverse, cardinality, etc., simply mapping each of them to the corresponding OWL primitive (Fig. 1). For example, the Protégé single slots ‘has_mass’ or ‘has_boundary’, defined with type SYMBOL, allowed values FALSE TRUE, and cardinality 0 1, are simply translated into an owl:DatatypeProperty, with range datatype Boolean, and declared to be an owl: FunctionalProperty. The Protégé multislot constitutional_part defined with type SYMBOL, allowed parents Physical_anatomical_entity and inverse slot constitutional_part_of, 2 A metaclass is a class whose instances are themselves classes is translated into an owl:ObjectProperty with (rdfs:range rdf:resource="#Physical_anatomical_entity") and inverse (owl:inverseOf rdf:resource="#constitutional_part_of").
Type INTEGER, FLOAT, STRING Type SYMBOL with allowed values T or F Type SYMBOL with allowed values (not T nor F) Type SYMBOL with allowed parents Type INSTANCE with allowed classes
Enrichment, in contrast, introduces new logical features. The enrichment rules
were designed to reflect the original underlying principles of the FMA model. Some
enrichment rules and the rationale behind them are presented next.
− Property restrictions: the choice between universal and existential property
restrictions is mainly based on the distinct role of template and own slots in Protégé.
Template slots “specify which slot each member of a class shall have and what the
restrictions (facets) on the values of these slots shall be” [
with allowed parents or allowed classes Ci with an allowed value
with a specific class C as value assigned with a specific datatype value assigned
OWL property restriction owl:allValuesFrom constraint to the union of all the allowed classes Ci enforced on the property when applied to the class owl:hasValue constraint to the specified value enforced on the property when applied to the class owl:someValuesFrom constraint to the class C, enforced on the property when applied to the class owl:hasValue constraint to the specific value, enforced on the property when applied to the class For example, the multislot bounded_by of the metaclass Organ with allowedparents Surface_of_organ is converted into the universal restriction (∀ bounded_by Surface_of_organ) on the property bounded_by of Organ, that is next inherited by its subclass Heart. But when Heart inherits bounded_by as an own slot assigned with the value Surface_of_heart, it is converted into the existential restriction (∃ bounded_by Surface_of _heart). Similarly, venous_drainage is restricted by a universal restriction inherited from its superclasses, but when Heart inherits venous_ drainage as an own slot assigned with the values Oblique_vein_of_left_atrium, Left_marginal_vein, Coronary_sinus, Posterior_vein_of_left_ventri cle, Unnamed_tributary_of_cardiac_ vein, Anterior_interventricular_vei n, Small_cardiac_vein etc. they are converted into owl:someValuesFrom restrictions specifying the value constraints on the property for the class Heart (Fig.3). − Equivalent class definition: a “defined” class has at least one necessary and sufficient condition. At this preliminary step, one slot p is manually selected, and a class A having values B1,…, Bn assigned to its own slot p, is defined as equivalent to the conjunction of all the existential value restrictions on p to the classes Bi and of metaclass and superclass of A (after some optimization). As aggregated objects are often described in terms of their parts and as meronymic relationships play a particularly important role in the FMA, it was chosen to define anatomical entities in terms of their parts. At this first step, the property “constitutional part” was selected, resulting in 570 defined classes. For example, the class Heart is defined by: Heart ≡ Organ_with_cavitated_organ_parts П (∃ contitutional_part Wall_of_heart) П (∃ Organ_with_cavitated_organ_parts Cavity_of_left_atrium) П (∃ contitutional_part … ) … П(∃ contitutional_part … ) (Fig. 4). The choice of the property ‘constitutional part’ was partly motivated by a size issue: constitutional part is well populated in FMA, compared for instance to ‘custom partonomy’, thus is computationally more significant. But, such a definition is not “semantically” satisfying for all classes, as all the classes cannot be uniformly defined solely in terms of their constitutional parts (the same parts may belong to different structures). But at this step, the priority was to test if Racer classification could be run. Different definitions for the different subtrees, and more complex expressions combining several properties will be next investigated (§ 4).
Metaclasses are converted into ordinary OWL DL classes. Subclass relation between metaclasses and metaclass instantiation are both translated into OWL subClassOf axioms. According to the previous rules, range restrictions of a template slot defined at metaclass are represented by universal property restrictions, while structural own slots with values assigned at class by existential property restrictions. In order to respect “selective inheritance”, own slots such as name, identifiers e.g. UWDAID, author etc., with values assigned at class, are converted to OWL annotations, preventing their propagation to their instances or subclasses. Each entity of the FMA is thus represented by an OWL DL class, its metaclass and instance definitions been merged. Fig. 4 shows the class Heart with its equivalentClass definition and its subClassOf axioms including universal or existential restrictions, derived from the original metaclass and class definitions of Heart in Protégé.
Metaclass (defclass Heart (is-a Organ_with_cavitated_organ_parts)
… )
Instance of Metaclass ([Heart] of Organ_with_cavitated_organ_parts (constitutional_part
Wall_of_heart Cavity_of_left_atrium Cavity_of_right_ventricle Cavity_of_left_ventricle Right_coronary_artery Left_coronary_artery … (bounded_by
Surface_of_heart) (arterial_supply
Right_coronary_artery Left_coronary_artery) … )
Heart in OWL DL <owl:Class rdf:ID="Heart"> <owl:equivalentClass> <owl:Class> <owl:intersectionOf rdf:parseType="Collection"> <owl:Class rdf:about= "#Organ_with_cavitated_organ_parts"/> <owl:Restriction> <owl:onProperty rdf:resource="#constitutional_part" /> <owl:someValuesFrom
rdf:resource="#Wall_of_heart" /> </owl:Restriction> … </owl:equivalentClass> <rdfs:subClassOf> <owl:Restriction> <owl:onProperty rdf:resource="#bounded_by"/> <owl:someValuesFrom rdf:resource="#Surface_of_heart"/> </owl:Restriction> </rdfs:subClassOf> <rdfs:subClassOf> <owl:Restriction> <owl:onProperty rdf:resource="#arterial_supply"/> <owl:someValuesFrom rdf:resource="#Right_coronary_artery" /> </owl:Restriction> … </owl:Class> The other conversion rules are reported in the annex available at http://www.med.univ-rennes1.fr/lim/doc_116.pdf. Aware of the arbitrariness of some of these choices, the enrichment rules were designed and implemented with flexible options. This flexibility permitted to automatically generate various OWL files with different flavors, size and computational complexity. Moreover, these options can be easily modified, which is key to the incremental approach adopted for reasoning (§3.1). 2.2 Results Ignoring laterality distinctions, i.e. classes differing from their parents only by laterality, a subset of about 40,000 concepts and their slot values were extracted for conversion, i.e. 57% of the 70,000 concepts of the original FMA. Applied to this subset, the conversion process described earlier resulted in about 117,000 frames, including 40,000 OWL named classes. More precisely, there are 187 properties and 85 individuals specified in this file. 20 properties correspond to annotation, 19 to datatype and 148 to object properties. There are 107,238 subClassOf axioms (38,772 from taxonomy and 3,378 from metaclass instantiation), 39337 classes where 559 are defined by equivalentClass axioms based on constitutional_part. OWL constructors allValuesFrom, someValuesFrom, hasValue, oneOf, unionOf, FunctionalProperty, SymmetricProperty, InverseOf all occur in the OWL file resulting from the conversion (available at mor.nlm.nih.gov/pubs/supp/2005-owled-cg/FMA-constitutionalPartForNS.owl). It took about 15mn to load the FMA OWL file in Protégé OWL in a Windows XP PC with 4GB memory (1h30 with 512Mb).
Reasoning with OWL proved to be a real challenge, due to the sheer size and complexity of the FMA. As the entire FMA in OWL DL raised inference problems hard to solve in terms of time and memory, an incremental approach was adopted.
We used Racer (Version 1.7) with the OWL files generated by the conversion process to investigate consistency checking and automatic classification. Launched from Protégé-OWL, the classification failed. Running Racer directly from Rice, we experienced problems related to memory limitation (4GB). Since Racer could not handle the entire FMA OWL file (in fact restricted to 2/3 of the whole FMA), as suggested by the Racer authors, we decided to test smaller versions so as to reduce the size and time issues and investigate eventual errors, adding more features incrementally. First, a FMA OWL version with all classes but without any properties was checked to test if the taxonomy alone could be successfully classified. Then, we added equivalent class definitions using only one property to test if the ontology with defined classes could pass Racer. Next, we successively introduced, step by step, object properties, annotation properties, datatype properties, and finally object properties used for attributed slots. When properties are introduced in partial versions, the conversion rules described previously are applied. For example, a small version where the object property bounds and its inverse bounded_by are introduced, includes for each class having these properties specified, the subclass axioms containaing the corresponding existential and universal restrictions of the properties bounds and bounded_by. 3.2 Results Racer passed the first test: the classification of the FMA OWL version without any properties was successful, taking about 25 minutes with 512Mo on a Pentium 4. Then, the classification with “defined” classes described by the conjunction of the existential restrictions on the constitutional_part or custom_partonomy property as necessary and sufficient condition was also successful. Next, various versions were generated with all classes but containing a limited number of properties. Depending on the properties introduced, the tests were successful or not. Some results are summarized below: Reasoning with Racer (version 1.7) was successful for the following partial versions: • Ontology with only the class hierarchy defined but without any property. • Ontology with defined classes (based on constitutional_part). • Ontology with defined classes (based on constitutional_part and constitutional_part_of). • Ontology with defined classes and annotation properties. • Ontology with defined classes, annotation properties, and all datatype properties. • Ontology with defined classes and primitive classes with restrictions on the property branch_of in subClassOf axioms. • Ontology with defined classes and primitive classes with restrictions on the property arterial_supply in subClassOf axioms. • Ontology with defined classes and primitive classes with restrictions on the property 2D_part in subclass axioms. • Ontology with defined classes and primitive classes with restrictions on the property bounds and its inverse bounded_by in subclass axioms. • Ontology with defined classes and primitive classes with restrictions on properties dimension and has_physical_state in subclass axioms. • Ontology with primitive classes with restrictions on attributed slot location and all slots used in location (e.g., related_object, etc.). • Ontology with primitive classes with restrictions on attributed slot attributed_part and all slots used in attributed_part (e.g., related_part, etc.) Reasoning with Racer (version 1.7) failed for: • Ontology with defined classes and annotation properties, added with primitive classes with restrictions on all the object properties in subclass axioms. • Ontology with defined classes and primitive classes with restrictions on all object properties in subclass axioms. • Ontology with primitive classes with restrictions on the property branch_of and on its inverse in subclass axioms. • Ontology with Subclass axioms with restrictions on the property continuous_with, declared symmetric. The reasons for failure are not easy to analyze. For instance, Racer was successful with equivalent class axioms and subclass axioms with restrictions on the property bounds and its inverse bounded_by, but failed for subclass axioms with restrictions on the property branch_of and its inverse branch, while it was successful with branch_of alone (without its inverse). This experience shows that the sheer size of the FMA is not the only issue. The results of reasoning with OWL are related to several factors. The complexity of the generated OWL ontology, due to the OWL constructors used (e.g. oneOf), the presence of inverseOf axioms or “global” axioms, and the algorithms implemented and optimization techniques of the reasoners, are certainly critical issues for the FMA.
Although problems with computational resources occurred for reasoning with the whole FMA in OWL DL, Racer could handle various less complex versions, which enabled to detect inconsistencies in the original FMA and to reclassify some classes.
No inconsistencies were found in the first versions, but when datatype properties were added several inconsistencies were identified. 113 classes were identified as unsatisfiable by Racer because of opposite boolean values: − Inconsistencies from conflicts between metaclass and class definitions in Protégé. A class assigned with a boolean value in its own slot and which inherits the opposite value from its superclasses, is unsatisfiable in OWL. For example, Zone_of _cell is unsatisfiable (hence, all its subclasses) because its own slot has_mass was assigned false at instance (converted to the restriction has_mass:false) while this single-slot had value true at its superclass Material_physical_anatomical_entity (converted to has_mass:true). Other inconsistencies were revealed from the inconsistency of the metaclass and class definitions of an entity. A class A subclass of B and instance of C in FMA, where B and C have opposite values for a boolean datatype property, e.g. has_mass, is unsatisfiable in OWL. For example, Compartment_subdivision is defined as a subclass of Anatomical_cluster, which is a subclass of Material_physical_anatomical _entity (has_mass:true). On the other hand, Compartment_subdivision is an instance of Anatomical_space, which is a subclass of Nonmaterial_physical_anatomical_entity (has_mass:false). − Inconsistencies from global and local conflicting domain or range. rdfs:range (resp. domain) restrictions are global. Thus if p has class A’ as domain and B’ as range, and A has a property p with range B, then B must be a subclass of B’ and A must be a subclass of A’. Conflicting definitions of global and local ranges or domains lead to inconsistencies in OWL. For example, Surface_of_wrist is unsatisfiable because ‘2D_part’ has an existential restriction to Anatomic_snuff_box which is a subclass of Material_physical_anatomical_entity (has_mass:true), while the range of “2D_part” is Non-material_physical_anatomical_entity (has_mass:false). These inconsistencies exhibit modeling errors in the original Protégé FMA.
Racer also reclassified some classes. In the ontology including defined classes based on the constitutional_part property, 286 classes of the asserted hierarchy were moved within the inferred hierarchy, and some classes were identified to be equivalent. For example, as the two sibling classes Wall_of_biatrial_part_of_heart and Wall_of_biventricular_part_of_heart, have the same constitutional parts3 in the original FMA, they became equivalent for this definition. However, the equivalence did not hold anymore when adding other restrictions to these definitions. For example, adding restrictions on the property constitutional_part_of, enables to differentiate the two classes, as they are parts of different wholes: Wall_of_biventricular_part_of_heart is a constitutional_part_of Biventricular_part_of_heart, while Wall_of_biatrial_part_of_heart is a constitutional_part_of Biatrial_part_of_heart. Thus, although most of the reclassifications were related to the class definitions in terms of their constitutional parts, it nevertheless shows the power of reasoning with OWL DL.
In conclusion, the results obtained so far show the benefits of OWL DL for the FMA. First, checking the logical consistency of the FMA enabled to find errors that would have probably been missed otherwise. Second, automatically computing the classification hierarchy is another advantage for such a large ontology. As the FMA has been under development at the University of Washington since 1994 and is still evolving, such services are useful for quality assurance purposes. 4
Converting a large part of the FMA from Protégé into OWL DL was possible. [
(1) First, to model each anatomical entity as a “set of sets”, (e.g., Vertebra as a
set of different types of vertebrae: cervical, thoracic, lumbar, themselves sets of other
sets e.g., first,.., fifth lumbar vertebra). A first order language as OWL DL cannot
capture this feature. However, the use of the representation of an anatomical entity as
a “set of sets” is quite limited in Protégé. In fact, the “members of each of these
collections are represented in Protégé as subclasses of Vertebra” [
(2) The metaclass construction has another purpose, “to enforce slot value
restrictions” [
(3) Metaclasses were also intended to specify multiple values specific to each class e.g., specifying that a Vertebra has parts Body_of_vertebra, Vertebral_arch, Bone_of_vertebra, etc. In OWL this can be captured by several restrictions such as (∃ part_of Body_of_vertebra) П (∃ part_of Vertebral_arch) П (∃ part_of Bone_of_vertebra)etc.
(4) Finally, metaclasses are used for specifying metadata such as name, author, authority, UWDAID, etc. Assigning vlaues to these “non structural” own slots at metaclass instantiation prevents them from being propagated to their instances or subclasses. In OWL this can be done thanks annotations. In conclusion, thanks to OWL’s higher expressiveness, most intended meanings of the Protégé metaclasses can be captured, with the exception of “set of sets”, which does not represent a significant loss in our opinion, considering their use in Protégé.
As far as we know, the NCI Thesaurus was one of the largest file in Protégé OWL so far. But it is much smaller and exhibits less complexity than the FMA in OWL. The NCI Thesaurus contains 53,000 frames, including 34,000 classes, 100 properties and 9,000 conditions, while the original FMA contains 70,000 concepts and the converted subset 117,000 frames, including 40,000 OWL classes, 187 properties and about 110,000 axioms. NCI was converted to OWL Lite, while the FMA is represented in OWL DL. No defined class, no hasValue, allValuesFrom restrictions, nor unionOf or enumerated classes oneOf are specified in the NCI, while they all occur in the FMA OWL file.
The size and complexity of the FMA in OWL make it a real challenge for DLs
systems. It showed that, with the current state of the art of DL inference technology, it
might generate inference problems that are hard to solve in terms of time and space
resources. Indeed, the main problem was computational. Some optimizations were
achieved to reduce the complexity. For example, it was necessary to step down the
number of disjunctions generated by the conversion for the domain of properties,
which caused Racer – would have any inference system – to run into space problems
Interestingly, after optimization, two classes remain in the domain of location
instead of 1,618 originally [
In the future, we would like to improve the current conversion process and to remove some of its limitations: − First, we suggest adding disjointness axioms between sibling primitive classes.
Ideally, a classification satisfies the so-called “jointly exhaustive and pairwise
disjoint” rule. The inconsistencies reported §3 are mainly based on opposite
values of a boolean datatype property and their propagation, but disjointness
axioms will most probably lead to identifying more inconsistencies in the FMA.
− Second, we propose using qualified cardinality restrictions. We converted
structural own slots values by existential property restrictions, mainly for two
reasons. On the one hand, the assumption that if a class A has a slot p filled with
values B1, B2 … Bn in Protégé (e.g., constitutional part), it means that for every
individual of A, p has at least one value of each class Bi. On the other hand we
were faced to the expressiveness limitation of OWL, which does not support
qualified cardinality restrictions. However, for example defining restrictions “has
Part someValuesFrom B1” and “hasPart someValueFrom B2” is weaker
than “hasPart exactly one B1 and one B2”, as it does not prevent from having
several parts of the same Bi. If OWL was extended with qualified cardinality
restrictions, more precise definitions might be provided.
− Thirdly, we suggest completing our current class definitions by closure axioms
[
lobe and left lower lobe” ) (regional_part
Upper_lobe_of_left_lung
Lower_lobe_of_left_lung) …) …) ([Right _lung] of Lung (definition “Lung which consists of the right upper
lobe, middle lobe and right lower lobe”) (regional_part
Upper_lobe_of_right_lung Middle_lobe_of_lung
Lower_lobe_of_right_lung)
− A major issue is the specification of the defined classes. Several possible options
might be considered [
At this first stage, the conversion aimed at capturing the Protégé FMA model as faithfully as possible, in order to evaluate its original properties. In the future, in addition to the above proposals, we suggest to introduce some changes in the model. For example, the OWL classes used for the Protégé attributed relations might be specified by n-ary relations in an external base related to the ontology. New classes might be introduced such as Venous_drainage, Arterial_Supply for improving consistency and factorizing reasons. Enumerated classes might be approximated otherwise etc.
An interesting point of discussion is about the choice of OWL DL versus OWL
Full for large-scale ontologies such as the FMA, and more generally for domain
versus application Web ontologies. The main lessons learnt from this experience is
that a possible option for large-scale domain ontologies such as the FMA, designed as
“sharable reference ontologies”, is to represent them in OWL DL with only primitive
classes, but a library of optional usual class equivalent definitions been provided
together. As each particular application, may have different needs, it will remain on
the user responsibility to select predefined definitions from the library or to build his
own definitions, so as to refine and customize the ontology according to his own
needs. For example, the brain MRI images application [
Converting the whole FMA from its original frame-based representation into the first order language OWL DL was possible, while capturing most features of the original model in Protégé. Reasoning with OWL proved to be a real challenge, because of the sheer size and complexity of the FMA in OWL. The entire FMA raised computational problems hard to solve in terms of time and space resources, but after some optimizations, various smaller versions were successfully tested with Racer. Several inconsistencies were revealed in the original modeling of the FMA. Some classes of the asserted hierarchy were reclassified; some classes were identified to be equivalent. Although most of them were related to the class definitions in terms of their constitutional parts, it nevertheless shows the power of reasoning with OWL DL. Thus, the results obtained so far demonstrate the advantages of OWL over frames for large-scale domain ontologies such as the FMA and help suggest future additional possible improvements of the FMA. This experiment is only a first step, the conversion rules are still being improved and refined. The resulting ontologies are being tested with RacerPro™ 4 and other reasoners may also be used. Issues with DL reasoner scalability should be carefully investigated. 4 http://www.franz.com/products/racer/ This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH), National Library of Medicine (NLM), while Songmao Zhang and Christine Golbreich were visiting scholars at the Lister Hill National Center for Biomedical Communications, NLM, NIH. Our thanks to Volker Haarslev and Ralf Möller for their valuable advice and help on DLs and Racer, to Cornelius Rosse, José Mejino and Todd Detwiler for the FMA, and to other helpful recommendations.