Biomedical ontologies with many types of relations allow a domain of interest to be modeled with a high level of semantic specificity and expressivity. However, the complexity of modeling with many relations can introduce inconsistencies and errors into the representation. Using the Foundational Model of Anatomy as a case study, we identified and analyzed the musculoskeletal content to show both consistencies and inconsistencies in the use of relations for modeling. We also share our early work in addressing the problem of consistency in ontology modeling through use of ontology-specific design patterns.
1.1.
The Foundational Model of Anatomy (FMA) is a theory of anatomy expressed in the form of an
ontology. As a theory, the FMA asserts that anatomy can be represented using a series of organizing
units that represent different levels of granularity. These units include “Cell”, “Portion of tissue”,
“Organ”, and “Organ system”. The theory accounts for all structures and spaces produced by
coordinated expression of an organism’s genes, as well as immaterial boundary of entities [
The FMA is intended to serve as a knowledgebase for software applications requiring knowledge of human anatomy and as a reference ontology for construction of application-specific anatomy ontologies. However, inconsistencies have crept into the modeling of the FMA, resulting in similar structures within the body represented in different ways. Causes of these inconsistencies include changes in modeling schemes used by the Structural Informatics Group at the University of Washington during the 20-year development of the FMA and inter-author variation. In addition, the last decade of development was funded through projects focused on specific regions of anatomy—which provided little opportunity to address consistency and completeness of the whole-body model. These issues of inconsistency and incompleteness compromise the ability of the FMA to serve as a knowledgebase for the next generation of intelligent biomedical informatics applications. 1.2.
The size and complexity of the FMA have made it a subject of study for groups working on quality
assurance methods for ontologies. General auditing techniques, including examining class and part
relationships, have been performed by several teams [
The FMA class “Organ system” (FMA ID 7149) is defined as “Anatomical structure, each instance of which has as its direct parts instances of one or more organ types which are interconnected with one another by zones of continuity.” This definition emphasizes the structural (rather than functional) modeling of the FMA. “Musculoskeletal system” (FMA ID 7482) is defined through its constitutional parts, which are “Entire musculature” and “Skeletal system”. (For purposes of this work we are not including the neural network and vasculature of the musculoskeletal system, which has very little modeling.). “Skeletal system” has constitutional parts “Skeleton” and “Set of all skeletal ligaments”.
Because the musculoskeletal system consists of a large number anatomical structures of a limited number of types that display a great deal of repetition in how they relate to one another, it serves as an excellent test case for examining consistency in a complex biomedical ontology. The purpose of this work is to demonstrate how we identified the musculoskeletal content of the FMA, to present our analysis highlighting consistencies and inconsistencies in the modeling, and to share our early work in addressing the problem of consistency in ontology modeling.
This study uses the latest version of the FMA (version 5.0.0, created April 2019, retrieved from http://sig.biostr.washington.edu/share/downloads/fma/release/latest/fma.zip). Manual inspection of the FMA was performed using Protégé Desktop [8] version 5.5.0. In addition, a Resource Description Framework (RDF) triple store database was built using the Apache Jena 3.12.0 framework on a MacOS 10.15.7 (Catalina) laptop using OpenJDK 13.0.2. This allowed us to query the triple store using the SPARQL query language via the Jena command line tool tdb2.tdbquery.
Our analysis proceeded in five steps:
Step 1: The class hierarchy of the entire FMA ontology was manually inspected to identify classes relevant to the musculoskeletal system. A list of “superclasses of interest” was created that consisted of classes for which all subclasses represent parts of the musculoskeletal system. We anticipate these classes (such as “Zone of bone organ”, and “Articular capsule”) will be useful for describing modeling schemes for the musculoskeletal system. Classes that do not have children (such as “Joint of sternum”) or do not appear to be fully classified in the FMA (such as “Subdivision of compartment of palm of hand”) were not included in this list. Classes for anatomical points, lines, and conduits were not included. An additional three classes (“Skeleton”, “Musculoskeletal system”, and “Skeletal system”) do not have children but were included in the list of superclasses for modeling purposes.
Step 2: Two types of SPARQL queries were performed to identify the relations (known as object properties in OWL) used to construct triples that make assertions about instances belonging to classes relevant to the musculoskeletal system. The first query identified triples for which the subject is a subclass of a “superclass of interest”. The second query identified triples for which the object is a subclass of a “superclass of interest”. Using the sets of triples returned from the queries, counts of the use of each type of relation were tabulated.
Step 3: Lists of relations were examined for their relevance to modeling the musculoskeletal system. The subset of relations describing how parts of the musculoskeletal system relate to one another were designated as “relations of interest”.
Step 4: SPARQL queries were performed to identify triples that contain a subject (or object) that is a subclass of a “superclass of interest” and relation that is a “relation of interest”. The object (or subject) of these triples was then classified using the “superclasses of interest” list. Queries are provided in a Gitlab repository at https://gitlab.com/roggsky/ontologyauditor.git.
Step 5: The types and frequencies of the classified triples were examined to identify consistent and inconsistent use of relations, their subject classes, and their object classes.
Manual inspection of the FMA identified 55 superclasses of interest for representing the musculoskeletal system. These classes and their placement in the class hierarchy is shown in Table 1. Our SPARQL queries identified a total of 69 relations used to form triples with these classes. Table 2 shows the 29 relations of interest for this work, as well as those relating musculoskeletal structures to those of other organ systems (n=14), describing spatial relationships between anatomical structures (n=18), and describing developmental anatomy (n=7).
Our approach to identifying and analyzing the musculoskeletal content of the FMA highlights the usefulness of this type of audit for assessing consistency and detecting errors.
Infrequent subject-relation-object as tool for error detection: As shown in Table 4, we detected a number of invalid subject-relation-object combinations. Most of these combinations occurred with low frequency, suggesting that low-frequency combinations could serve as a flag for verification by a curator. We note that not all triples with unusual subject-relation-object combinations are incorrect— they may simply represent anatomical configurations that occur less frequently. For example, of 262 triples using attaches_to, 259 of those have a “Zone of bone organ” class as the object, while the remaining three use a “Subdivision of skeletal system” as the object. These triples describe the attachment of the proximal tendon of the temporalis to a region of the cranium spanning more than one bone organ, so this explains the unusual classification. As another example, the combination “Skeletal ligament” origin_of “Muscle organ” was detected only once in our audit, but the triple with this classification (“Ligamentum nuchae” origin_of “Serratus posterior superior”) is anatomically valid. Several additional triples, including “Ligamentum nuchae” origin_of “Trapezius” could be modeled following this scheme.
Inconsistent use of relations: We found two different methods used for modeling the relation of muscles to the structure they attach to. One method employs the relation attaches_to (inverse: receives_attachment_from). The other uses has_origin (inverse: origin_of) and has_insertion (inverse: insertion_of). The latter method is intended to denote which bone does not move during muscle contraction, but this is complicated by the fact that many muscles can change their functional origin and insertion based on the joint that is being mobilized during a given contraction. Although not explored in this audit, an additional relation (muscle_attachment) is used in a 23 cases to model a ternary relationship among a muscle organ, its site of origin or insertion, and the region of the muscle involved in the attachment. This lack of standardized modeling for the attachment of muscles highlights that inconsistent use of relations in the FMA makes it difficult to use as a computable knowledgebase.
Our audit also points to several additional issues:
Lack of definitions for relations: The fuses_with relation was introduced into the FMA while modeling craniofacial anatomy to describe time-based relationships between anatomical structures during development [9]. However, three triples use this relation to represent how fibers of the distal tendon of biceps femoris intermingle with those of the fibular collateral ligament. This is not necessarily an error (because the relation does not have a definition), but the term “fuses with” is not generally used outside of embryological development (attaches_to is a more appropriate relation for this example).
Constitutional parts of the Musculoskeletal system are incomplete: FMA modeling aims to represent exhaustive partitions of structures when they are described using either the constitutional_part or regional_part relations. Tracing the constitutional parts of “Musculoskeletal system” through the part hierarchies accounts for the muscle organs, bone organs, cartilage organs, and skeletal ligaments organs. But the synovial sac organs and skeletal membrane organs are missing from this partonomy.
Updates needed to legacy naming and modeling: We encountered a few labels of classes that reflect early modeling. For example, “Region of bony part of bone” should be interpreted as “Region of bony part of bone organ” (the term “bone” has been deprecated and replaced with wording that clarifies whether the bone organ or a portion of bone tissue is referred to). The relations part and part_of are no longer to be used in constructing triples (current modeling uses either regional_part or constitutional_part), but we found these relations used in 15 triples.
To address issues of consistency in large biomedical ontologies such as the FMA, we are in the early stages of developing a method to represent knowledge of how domain-specific entities relate to one another in a machine readable format. Our approach is similar to the use of design patterns for ontologies, which are schema that serve as reusable solutions to common modeling tasks in ontology development [10]–[12]. However, the patterns we are developing are specific to the FMA.
These ontology-specific patterns are specifications of subgraphs that are machine readable and have parameterized properties which layer “closed world” constraints onto the “open world” assumptions of OWL. In the context of the FMA, they place restrictions and requirements on relations between types of anatomical structures. Our process for developing these patterns begins with analyzing the FMA to understand variations in modeling, as shown in this paper. We then construct the patterns in the form of diagrams based on existing modeling in the FMA, the intent of the FMA, and expert anatomical knowledge. Figure 1 shows a diagram of a preliminary pattern for the class “Synovial joint”.
We envision four ways in which this type of pattern can be applied to the development and use of ontologies: • Guided content creation. When the authors have provided a pattern for classes of a given type (such as muscle organs), editors creating new classes of that type (for example, Biceps brachii) can use the pattern as a template for how that class is to be represented in the ontology. This would help enforce consistent modeling. • Guided content retrieval. If content has been modelled uniformly, patterns can be used to facilitate content retrieval because they specify the relations that may be of interest to users (such as retrieving tendons related to a muscle organ via the constitutional_part property). • Guided error checking. If patterns were not used consistently to guide content creation, then patterns can be used to find areas of content that violates patterns. Depending on how the patterns are specified, they could find not only incomplete modeling, but also the use of relations or types of classes that are prohibited. • Documentation. Users of complex ontologies need to make sense of the modeling schemes employed, and simply browsing an ontology may not reveal these schemes. Providing humanreadable diagrams of patterns will allow viewers to better understand the ontology's structure, particularly when they cannot directly ask the author(s).
The FMA provides an excellent case study for the development of ontology-specific patterns. As demonstrated in this work, the musculoskeletal system consists of a number of organ types that relate to one another in predictable ways. Additional organ systems, such as the nervous system, circulatory system, are also good candidates for describing with patterns. Patterns could be applied to the FMA to describe modeling of other repetitive structures, such as anatomical cavities, their walls, and their contents.
Our approach to developing ontology-specific patterns is intended to be generalizable to any ontology, and could be useful for ensuring consistency in other biomedical ontologies that use a large number of relations, such as SNOMED CT and the Cell Ontology.
We thank José (Onard) Mejino and James Brinkley of the Structural Informatics Group at the University of Washington for supporting M.C. during her doctoral and postdoctoral studies when this work was initiated. [5] L. Luo, L. Tong, X. Zhou, J. L. V. Mejino, C. Ouyang, Y. Liu, Evaluating the granularity balance of hierarchical relationships within large biomedical terminologies towards quality improvement, Journal of Biomedical Informatics, vol. 75, pp. 129–137, Nov. 2017, doi: 10.1016/j.jbi.2017.10.001. [6] L. Luo, J. L. V. Mejino, G.-Q. Zhang, An analysis of FMA using structural self-bisimilarity, Journal of Biomedical Informatics, vol. 46, no. 3, pp. 497–505, Jun. 2013, doi: 10.1016/j.jbi.2013.03.005. [7] I. J. Kalet, J. L. V. Mejino, V. Wang, M. Whipple, J. F. Brinkley, Content-specific auditing of a large scale anatomy ontology, Journal of Biomedical Informatics, vol. 42, no. 3, pp. 540–549, Jun. 2009, doi: 10.1016/j.jbi.2009.02.006. [8] M. A. Musen, The Protégé project: A look back and a look forward, AI Matters, vol. 1, no. 4, pp.
4–12, Jun. 2015, doi: 10.1145/2757001.2757003. [9] J. F. Brinkley et al., The Ontology of Craniofacial Development and Malformation for translational craniofacial research, American Journal of Medical Genetics Part C: Seminars in Medical Genetics, vol. 163, no. 4, pp. 232–245, Nov. 2013, doi: 10.1002/ajmg.c.31377. [10]V. Presutti, A. Gangemi, Content ontology design patterns as practical building blocks for web ontologies, in Conceptual Modeling - ER 2008, vol. 5231, Q. Li, S. Spaccapietra, E. Yu, and A. Olivé, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2008, pp. 128–141. URL: http://link.springer.com/10.1007/978-3-540-87877-3_11 [11]A. Gangemi, Ontology Design Patterns for Semantic Web Content, in: The Semantic Web – ISWC 2005, vol. 3729, Y. Gil, E. Motta, V. R. Benjamins, and M. A. Musen, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2005, pp. 262–276. URL: http://link.springer.com/10.1007/11574620_21 [12]Ontology Design Patterns, http://ontologydesignpatterns.org/wiki/Main_Page.