|
|
AMIA Annu Symp Proc. 2006; 2006: 804–808. | PMCID: PMC1839588 |
Copyright This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media for any purpose. Enhancing Biomedical Ontologies through Alignment of Semantic Relationships: Exploratory Approaches Lowell Vizenor, Ph.D.,1 Olivier Bodenreider, M.D., Ph.D.,1* Lee Peters, MS,1 and Alexa T. McCray, Ph.D.2 1 National Library of Medicine, Bethesda, Maryland 2 Harvard Medical School, Boston, Massachusetts Objective This paper investigates several methods for aligning Metathesaurus relationships with their counterparts in the UMLS Semantic Network. Unlike the categorization link defined between Metathesaurus concepts and Semantic Network types, no such correspondence exists between the relationships at these two levels of the UMLS. Methods The first approach attempts to elicit the semantics of Metathesaurus relationships through an examination of their relata at different levels: concept, high-level ancestors and semantic types. The second approach examines the frequency of association between a given Semantic Network relationship and the actual relationships observed in the Metathesaurus between the concepts categorized by these semantic types. Results : A total of 139 relationships are present in the Metathesaurus. Using the methods described in this paper, 80 (58%) could be aligned with Semantic Network relationships. The remaining relationships are vocabulary internal, used, for example, for vocabulary management or to indicate strictly lexical relationships. The work reported here is a first step in the attempt to build a more comprehensive ontology of biomedical relationships. Two complementary yet independent Unified Medical Language System ® (UMLS ®) knowledge sources are the Metathesaurus ® and the Semantic Network. The Metathesaurus is a large repository of interrelated concepts coming from one hundred biomedical vocabularies. The Semantic Network, by contrast, is a small, manually curated high-level network of 135 semantic types and 54 semantic relationships. The UMLS editors assign categorization links, which thereby relate the two structures. More precisely, every Metathesaurus concept is assigned to at least one semantic type, independently of its hierarchical position in a source vocabulary. The rationale for this two-level structure is to provide a uniform semantics to the concepts “regardless of the particular structure of the source vocabulary” [ 1]. In addition to the over one million Metathesaurus concepts there are also a number of Metathesaurus relationships—most of which come from the individual source vocabularies. In this case, however, the UMLS does not directly link Metathesaurus relationships to Semantic Network relationships. One consequence of this is that Semantic Network relationships cannot be used in a straightforward way to validate Metathesaurus relationships. For example, as shown in , one auditing method for the UMLS is to simply check the compatibility between a relationship asserted between two concepts in the Metathesaurus (R M) and the possible relationships defined in the Semantic Network (R SN) between the semantic types of these two concepts. Intuitively, R M is expected to be equivalent to or subsumed by R SN. However, since no equivalence or subsumption relations are defined between relationships across the two levels of the UMLS, validation on a large scale is not easily accomplished [ 2, 3]. The objective of this study is to explore methods for establishing such relations (e.g., R M equivalent to R SN, R M more specific than R SN). This work is part of a broader and on-going project of enriching the Semantic Network through the development of an ontology of biomedical relationships. We develop a number of related methods for aligning Metathesaurus relationships with Semantic Network relationships—a preliminary step in the development of such an ontology of biomedical relationships. The general framework of this study is that of an ontology alignment. However, in contrast to most existing ontology alignment methods [ 4], our primary focus is the alignment of relationship (not concepts) across ontologies. Because the two knowledge sources compared in this study represent knowledge at widely different levels of granularity, it is difficult to use existing tools. This study also represents an attempt to align the two UMLS knowledge sources under investigation in [ 5]. There, however, the authors used sets of concepts (and not relationships) to align concepts and semantic types. To our knowledge, the only other comparison of relationships across biomedical ontologies is a study between relationships in the Foundational Model of Anatomy and GALEN [ 6]. In this work, the authors identified patterns of relationships between equivalent concepts across ontologies and derived correspondences between relationships from their frequency of association. This associative technique is one of several approaches we are using in this paper. Some progress has been made recently in defining the relationships used in biomedical ontologies. Smith and colleagues [ 7] provide formal definitions for ten such relationships ( Is_a, Part_of, Located_in, Contained_in, Adjacent_to, Transformation_of, Derives_from, Preceded_by, Has_participant, and Has_agent), which as a whole represents a small ontology of biomedical relationships that is currently part of the OBO ontology library [ 8]. There exist other ontologies of relationships as part of ontologies such as GALEN [ 6] and the UMLS Semantic Network [ 9]. Most vocabularies integrated into the UMLS contribute thesaural relationships (e.g., parent/child, broader/narrower than) to the Metathesaurus. In addition, some vocabularies contribute specified relationships such as isa and location_of. Relations in the Metathesaurus are represented bidirectionally. In practice, each relation (C1, rel, C2) is mirrored by a relation (C2, rel’, C1), where rel’ is the inverse of rel. Relationships present in several vocabularies include isa, location_of, ingredient_of, manifestation_of and mapped_to). The semantics of the Metathesaurus relationships are implicit; that is, no definitions are given for the relationships used. A total of 4,328,245 direct relations involving specified relationships are represented in the 2005AC version of the UMLS. Our first step was to establish the list of all relationships used in the Metathesaurus paired with their inverses, because no such list is provided as part of the UMLS distribution. The semantics of the relationships in the UMLS Semantic Network are explicit. Each of the Semantic Network relationships has an inverse, a textual definition, and a list of semantic types that are linked by the relationship. The relationships are organized in a hierarchy which comprises five high-level categories: functionally related to (e.g., treats), physically related to (e.g., contains), spatially related to (e.g., adjacent_to), temporally related to (precedes) and conceptually related to (e.g., analyzes). There are about 7,000 relations defined in the Semantic Network. Our methods for aligning relationships can be summarized as follows. The first, Metathesaurus-centric approach consists of eliciting the semantics of Metathesaurus relationships by examining their relata at different levels: concept, high-level ancestors and semantic types. The second, Semantic Network-centric approach examines the frequency of association between a given Semantic Network relationship and the actual relationships observed in the Metathesaurus between the concepts categorized by these semantic types. Metathesaurus-Centric Approach 1) Manual elicitation We created two random samples of a maximum of 50 relations per Metathesaurus relationship to be evaluated by the authors. (50 represents a manageable number of relations to be examined manually and random selection ensures the representativity of the sample). The authors carefully studied the way these relationships are used within a given terminology to determine the meaning of the relationship within that terminology. In those cases where the usage was clear, it was possible to link the Metathesaurus relationship to a Semantic Network relationship and to identify the type of relation between them (e.g. semantically equivalent, more specific than or more general than). For example, the relata of the Metathesaurus relationship causative_agent_of appeared to be infectious agents and pharmacologic substances on the one hand and disorders on the other. This relationship so understood corresponds to the Semantic Network relationship causes, which is defined as follows: “Brings about a condition or an effect. Implied here is that an agent, such as for example, a pharmacologic substance or an organism, has brought about the effect” [ 10]. 2) Abstraction at the level of high-level concepts In order to characterize the relata, we compute the lowest common ancestor for the domain and range of each Metathesaurus relationship in a given source. To this end, we construct a graph of all the ancestors of the relata on the side the domain and on the side of the range, respectively. Second, for each ancestor, we record the frequency (number of concepts from the domain/range having this concept as their ancestor) and distance between this concept and the original domain/range concept. Finally, we identify the ancestor in the graph for which the frequency is maximal and the depth minimal. Consider the relationship access_instrument_of in SNOMED CT. The lowest common ancestors are Endoscope (domain) and Procedure by method (range). In practice, this automatic approach allows us to understand the prototypical relationship access_instrument_of as being a relationship between endoscopes and medical procedures. In some cases, the prototypical relationship is uninformative, because the lowest common ancestor is the root of the terminology. In other cases, there exists so much dispersion that the semantics of the relationship cannot be elicited by this method. 3) Abstraction at the level of semantic types The relata can also be characterized at a higher level by their semantic types. For each Metathesaurus relationship, we compute the distribution of the semantic types of the concepts on the side of the domain and on the side of the range, respectively. For example, all 1,600 Metathesaurus relations involving the SNOMED CT relationship access_instrument_of have Medical Device as the semantic type of the range concept. Domain concepts, in contrast, are essentially divided between Therapeutic or Preventive Procedure (898 cases) and Diagnostic Procedure (699 cases). In the remaining 3 cases, the semantic type of the domain concept is Health Care Activity. In order to find the equivalent of access_instrument_of in the Semantic Network, we need to examine the possible Semantic Network relationships defined between Therapeutic or Preventive Procedure and Diagnostic Procedure on the one hand and Medical Device on the other. In this case, we find uses, defined as “Employs in the carrying out of some activity. This includes applies, utilizes, employs, and avails.” [ 10] Semantic Network-Centric Approach Semantic types are high-level categories to which all Metathesaurus concepts are linked. It is therefore possible, starting from a given semantic type, to obtain its extension, i.e., the set of concepts that have been assigned this semantic type. From a given Semantic Network relation (T1, relSN, T2), we extract E1 and E2, the extensions of T1 and T2, respectively. We sthen examine what relations (C1, relM, C2) are represented in the Metathesaurus, pairwise, between one concept C1 from E1 and a concept C2 from E2. We obtain a set of Metathesaurus relationships {relM1, relM2, …} along with frequency information for each relationship in this set. Of these, the Metathesaurus relationships relM associated with the original Semantic Network relationship relSN with a high frequency constitute the candidates for identifying counterparts of relSN in the Metathesaurus. For example, the Semantic Network relationship ingredient_of is defined between the domain semantic type Substance and the range semantic type Clinical Drug. Additionally, ingredient_of is inherited downwards along the isa hierarchy of semantic types. The union of the extensions of the semantic types on the side of the domain and range of ingredient_of contains 260,505 and 160,994 Metathesaurus concepts, respectively. The Semantic Network relationship ingredient_of is frequently associated with several Metathesaurus relationships, including active_ingredient_of, dose_form_of and ingredient_of. Of note, other Metathesaurus relationships linking substances to drugs (e.g., metabolizes and has_contraindication) are also associated with the Semantic Network relationship ingredient_of, although with a lesser frequency. In order to illustrate this general methodology, we take an extended example of a single Metathesaurus relationship and show how each method is used to help determine its semantics and identify a correspondence with Semantic Network relationships. The SNOMED CT relationship finding_site_of is used throughout this example. The first step in determining the semantics of finding_site_of is to look at the way this relationship is used in SNOMED CT. In this case, we study the list of fifty randomly selected relations. Here is a sample of the domain and range pairs for finding _site_of: - - Endocrine structure / External endometriosis
- - Gallbladder structure / Malignant tumor of gallbladder
- - Skin structure / Epithelioma basal cell
- - Stomach wall structure / Gastromalacia
In this case, we determined that finding _site_of is a specification of the Semantic Network relationship location_of, defined between anatomical structures and disorders. In other words, finding _site_of is more specific than location_of. Ideally, we would like to show that our conclusion is consistent with the source terminology and the UMLS as a whole. To this end, we abstract the relationship to the level of high-level concepts by computing the lowest common ancestor for the domain and range of each Metathesaurus relationship in the source under investigation. As expected, all fifty domain concepts of finding_site_of in SNOMED CT are descendants of Anatomical structure. The lowest common ancestor to the fifty range concepts is the root of the SNOMED CT vocabulary, indicating that the range concepts belong to several distinct hierarchies in SNOMED CT. In fact, while Gastromalacia and Malignant tumor of gallbladder are under the top-level concept Clinical finding, Epithelioma basal cell is subsumed by Morphological abnormality, itself under the top-level concept Body structure. This finding is not inconsistent with our previous claim that finding_site_of is a specification of location_of, but it also provides no positive evidence to support it directly. We now abstract to the level of Semantic Types. On the side of the domain of finding_site_of, the most frequent semantic types are Body Part, Organ, or Organ Component, Body System, Body Location or Region, Body Space or Junction and Tissue. Analogously, Injury or Poisoning, Disease or Syndrome, Finding, Congenital Abnormality and Neoplastic Process are the most frequent semantic types on the range side. The only Semantic Network relationship defined between the domain and range semantic types above is location_of. Moreover, out of the 63,655 pairs of Metathesaurus concepts related by finding_site_of, 99.5% have their semantic types related by location_of (when any relationship is defined between their semantic types at all). At the level of abstraction of the semantic types, we find strong evidence to support the correspondence between the Metathesaurus relationship finding_site_of and the Semantic Network relationship location_of. Additional evidence can be found when examining this correspondence using the Semantic Network-centric approach. We create the set of Metathesaurus concepts categorized by the semantic types in the domain and range of the Semantic Network relationship location_of, respectively. The relationships existing in the Metathesaurus between these sets of concepts include finding_site_of, predominantly, but also procedure_site_of, and the Metathesaurus relationships location_of and isa. This relative dispersion does not contradict our prior finding, but rather indicates that other Metathesaurus relationships (e.g., procedure_site_of) represent a specialization of the Semantic Network relationship location_of. The presence of isa in association with location_of is somewhat unexpected as there is an important ontological distinction between diseases (processes) and anatomical structures (entities). However, this is explained when we note that abnormal anatomical structures are often also considered diseases (e.g., Bladder fistula isa Bladder disease). A total of 139 relationships are present in the Metathesaurus. Most of the 45 English vocabularies that include relationships have just a small number. The largest number of relationships are contributed by SNOMED CT (62), LOINC (15) the National Drug File – Reference Terminology (15), the University of Washington Digital Anatomist (8), and RxNorm (7). 116 are unique to a specific vocabulary and 23 are found in at least two vocabularies, e.g., component_of is found in SNOMED CT, PDQ, and LOINC. Using the methods described above, we were able to align 80 (58%) of the Metathesaurus relationships with Semantic Network relationships. In some cases this alignment is at a coarse level of granularity, e.g. metabolic_site_of is more specific than functionally_related_to, in other cases, the relationships are roughly equivalent to each other, e.g., focus_of and issue_in, and in 27 cases there exists an identical relationship in the Semantic Network, e.g., affects, process_of, ingredient_of. The remaining 59 Metathesaurus relationships fall into a number of additional categories. Some are lexical relationships (e.g., british_form_of, permutated_term_of, xml_form_of and suffix_of), some are mapping relationships (e.g., see_from and uniquely_mapped_from), and others are used strictly for vocabulary management purposes (e.g., classifies, moved_from, replaces). These relationships are not associated with any particular pair of semantic types and do not converge towards any particular Semantic Network relationships. They are also not useful candidates for inclusion in an ontology of biomedical relationships. The methods employed in this study combine manual and automated techniques. In some cases, even using multiple methods, it is difficult to discern the underlying semantics of the Metathesaurus relationships. The methods described here provide a good characterization of the meaning of a relationship, but they are not a substitute for an explicit definition. The developers of individual vocabularies have invested a good deal of effort in linking pairs of terms with these specific relationships. Their usefulness for vocabulary-specific applications would be greatly enhanced if their semantics were made explicit. Clearly defined relationships would also make it possible for these originally vocabulary-specific relationships to be fully and accurately integrated with other vocabularies, thereby further broadening their usefulness. We have explored a number of methods for aligning two UMLS knowledge sources. We have conducted this work, first, to improve the usefulness of the vocabulary-specific relationships in the context of the UMLS. Second, we intend for the methods described here to be a first step in identifying and classifying biomedical relationships beyond the existing 54 Semantic Network relationships. The goal is to use the augmented set of relationships as the basic building blocks for a broader and more comprehensive ontology of biomedical relationships. This work is important since it will go some way toward ensuring that the ontology of biomedical relationships will remain relatively stable as new terminologies are added to the UMLS and as changes are made to existing terminologies. There are already some efforts toward developing an ontology of relationships that are specific to biomedicine [ 6, 11, 12] and there is ongoing research on the underlying ontological principles [ 7, 13– 15]. The work reported here is an effort to contribute to these investigations. An ontology of biomedical relationships promises to add logical and ontological rigor to biomedical ontologies, to bring existing terminologies and ontologies more closely in-line with one another, and to serve as a resource for the construction of future vocabularies. In this study we have employed a number of methods to determine the semantics of Metathesaurus relationships and to align relationships between the Semantic Network and the Metathesaurus. These methods should be seen as complementary. In the ideal case, all the methods would point to a single semantic interpretation of a given relationship. In practice, some methods work better than others for some cases. In some cases, the various methods might even produce inconsistent results, due to inconsistencies in the vocabularies under investigation. The hope is that, in combination, these methods will provide, first, a comprehensive strategy for guiding the alignment of Metathesaurus relationships and Semantic Network relationships, and, second, will serve as a good starting point for the development of a comprehensive ontology of biomedical relationships. This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH), National Library of Medicine (NLM). This work was done while Lowell Vizenor was a visiting fellow at the Lister Hill National Center for Biomedical Communications, NLM, NIH. 1. McCray AT, Nelson SJ. The representation of meaning in the UMLS. Methods Inf Med. 1995;34(1–2):193–201. [PubMed] 2. Cimino JJ. Auditing the Unified Medical Language System with semantic methods. J Am Med Inform Assoc. 1998;5(1):41–51. [PubMed] 3. McCray AT, Bodenreider O. A conceptual framework for the biomedical domain. In: Green R, Bean CA, Myaeng SH, editors. The semantics of relationships: an interdisciplinary perspective. Boston: Kluwer Academic Publishers; 2002. pp. 181–198. 4. Noy NF. Tools for mapping and merging ontologies. In: Staab S, Studer R, editors. Handbook on Ontologies: Springer-Verlag; 2004. pp. 365–384. 5. Bodenreider O, Burgun A. Aligning knowledge sources in the UMLS: methods, quantitative results, and applications. Medinfo. 2004;11(Pt 1):327–31. 7. Smith B, Ceusters W, Klagges B, Kohler J, Kumar A, Lomax J, et al. Relations in biomedical ontologies. Genome Biol. 2005;6(5):R46. [PubMed] 11. Mungall CJ. Obol: integrating language and meaning in bio-ontologies. Comparative and Functional Genomics. 2004;5(6–7):509–520. [PubMed] 12. Rosse C, Mejino JL., Jr A reference ontology for biomedical informatics: the Foundational Model of Anatomy. J Biomed Inform. 2003;36(6):478–500. [PubMed] 14. Schulz S, Hahn U. Part-whole representation and reasoning in formal biomedical ontologies. Artif Intell Med. 2005;34(3):179–200. [PubMed] 15. Smith B, Grenon P. The cornucopia of formal-ontological relations. Dialectica. 2004;58(3):279–296.
|
