The National Library of Medicine (NLM) and the Food and Drug Administration (FDA) Center for Drug Evaluation and Research (CDER) are collaborating on a research project to extract adverse drug reactions from the biomedical literature. More specifically, this investigation leverages the indexing of MEDLINE citations to extract associations between co-occurring drug entities and clinical manifestations in the context of adverse events.
The biomedical literature is indexed with the Medical Subject Headings (MeSH) vocabulary. For data mining purposes, however, adverse drug reactions are usually analyzed in reference to other standard vocabularies, namely the Anatomical Therapeutic Chemical (ATC) drug classification system for drug entities, and the Medical Dictionary for Regulatory Activities (MedDRA) for clinical manifestations. Toward this end, drug entities have to be mapped from MeSH to ATC, and manifestations from MeSH to MedDRA. This paper focuses only on the drug entities.
Drug entities include not only individual drugs (e.g., atorvastatin), but also drug classes (e.g., statins). In previous work, we have mapped individual drugs between RxNorm (which includes MeSH drugs) and ATC (Bodenreider and Taft, 2013;Winnenburg and Bodenreider, 2012). In contrast, no mapping is available between pharmacologic classes in MeSH and in ATC. Moreover, unlike individual drugs, whose names are relatively standardized across vocabularies, pharmacologic classes exhibit greater variability, not only in their names, but also in granularity. For example, the drug lisinopril is classified as Angiotensin-Converting Enzyme Inhibitors in MeSH, but as ACE inhibitors, plain in ATC.
The objective of this study is to investigate various ontology matching techniques for aligning pharmacologic classes between MeSH and ATC. Such methods are expected to facilitate the curation of a mapping by experts. To our knowledge, this work represents the first effort to map pharmacologic classes between MeSH and ATC using a sophisticated instance-based alignment technique.
The general framework of this study is that of ontology alignment (or ontology matching). Various techniques have been proposed for aligning concepts across ontologies, including lexical techniques (based on the similarity of concept names), structural techniques (based on the similarity of hierarchical relations), semantic techniques (based on semantic similarity between concepts), and instance-based techniques (based on the similarity of the set of instances of two concepts). An overview of ontology alignment is provided in (Euzenat and Shvaiko, 2007).
The main contribution of this paper is not to propose a novel technique, but rather to apply existing techniques to a novel objective, namely aligning pharmacologic classes between MeSH and ATC. To this end, we use lexical and instancebased techniques, because the names of pharmacologic classes and the list of drugs that are members of these classes are the main two features available in these resources.
Lexical techniques for ontology matching compare concept names across ontologies. When synonyms are available, they can be used to identify additional matches. Matching techniques beyond exact match utilize edit distance or normalization to account for minor differences between concept names.
As part of the Unified Medical Language System (UMLS), linguistically-motivated normalization techniques have been developed specifically for biomedical terms (McCray, et al., 1994). UMLS normalization abstracts away from inessential differences, such as inflection, case and hyphen variation, as well as word order variation. The UMLS normalization techniques form the basis for integrating terms into the UMLS Metathesaurus, but can be applied to terms that are not in the UMLS. For example, the ATC class Thiouracils (H03BA) and the MeSH class Thiouracil (D013889) match after normalization (ignoring singular/plural differences).
Lexical techniques typically compare the names of concepts across two ontologies as provided by these ontologies. However, additional synonyms can be used, for example, synonyms from the UMLS Metathesaurus. In other words, we leverage cosynonymy similarity for matching pharmacologic classes. In this case, although the ATC class Anticholinesterases (N06DA) and the MeSH class Cholinesterase Inhibitors (D002800) do not match lexically, both names are cosynonyms, because they are found among the synonyms of the UMLS Metathesaurus concept C0008425.
Also called extensional techniques, instance-based techniques compare classes based on the sets of individuals (i.e., instances) of each class. Many biomedical ontologies consist of class hierarchies, but do not contain information about instances. Here, however, individual drugs (e.g., atorvastatin) are the members -not subclasses -of pharmacologic classes (e.g., statins). In other words, pharmacologic classes have individual drugs as instances, not subclasses.
Several methods have been proposed to implement instancebased matching. (Isaac, et al., 2007) decompose these methods into three basic elements: (1) A measure is used for evaluating the association between two classes based on the proportion of shared instances. Typical measures include information-based measures (e.g., Jaccard similarity coefficient) and statistical measures (e.g., log likelihood ratio). (2) A threshold is applied to the measures and pairs of classes for which the measure is above the threshold are deemed closely associated and mapping candidates. (3) Hierarchical relations in the two ontologies to be aligned can also be leveraged by deriving instance-class relations between instances of a given class and the ancestors of this class. In other words, in addition to asserted classes (i.e., the classes of which individual drugs are direct members), we also consider inferred classes (i.e., the classes of which asserted classes are subclasses). For example, the class asserted in MeSH for the drug atorvastatin is Hydroxymethylglutaryl-CoA Reductase Inhibitors (i.e., statins), whose parent concepts include Anticholesteremic Agents. Therefore, the class Anticholesteremic Agents is an inferred pharmacologic class for atorvastatin.
As part of the EU-ADR project, (Avillach, et al., 2013) extracted adverse drug reactions from the biomedical literature and mapped MeSH drugs to ATC through the UMLS. How-ever, their mapping was limited to individual drugs and did not include pharmacologic classes.
Lexical techniques are a component of most ontology alignment systems (Euzenat and Shvaiko, 2007). While there have been attempts to map individual drugs from ATC to concepts in the UMLS and MeSH through lexical techniques, (Merabti, et al., 2011) note that these techniques are not appropriate for the mapping of pharmacologic classes.
While instance-based techniques are also available in many systems, the applicability of this technique is limited, because there is often no available information about instances as part of the ontologies to be aligned. For example, most biomedical terminologies and ontologies are simple class hierarchies. The instances of these classes are present in electronic medical record systems and clinical data warehouses, but typically not distributed along with the ontologies. One exception in the biomedical domain is the Gene Ontology (GO) (Ashburner, et al., 2000), for which the gene products annotated to GO terms can be considered instances of the corresponding classes. (Kirsten, et al., 2007) have aligned GO terms across the three hierarchies of GO through the gene products to which they are co-annotated.
To our knowledge, our work is the first attempt to align pharmacologic classes with instance-based techniques (i.e., beyond name matching), and the first application of aligning pharmacologic classes in ATC and MeSH. Moreover, while most ontology alignment systems mainly consider matches between equivalent classes, we are also interested in identifying those cases where one class is included in another class.
The ATC is a clinical drug classification system developed and maintained by the World Health Organization (WHO) as a tool for drug utilization research to improve quality of drug use (ATC, 2013). The system is organized as a hierarchy that classifies clinical drug entities at five different levels: 1st level anatomical (e.g., A: Alimentary tract and metabolism), 2nd level therapeutic (e.g., A10: Drugs used in diabetes), 3rd level pharmacological (e.g., A10B: Blood glucose lowering drugs, excluding insulins), 4th level chemical (e.g., A10BA: Biguanides), and 5th level chemical substance or ingredient (e.g., A10BA02: metformin). The 2013 version of ATC integrates 4,516 5 th -level drugs and 1,255 drug groups (levels 1-4).
The Medical Subject Headings (MeSH) is a controlled vocabulary produced and maintained by the NLM (NLM, 2013). It is used for indexing, cataloging, and searching the biomedical literature in the MEDLINE/PubMed database, and other documents. The MeSH thesaurus includes 26,853 descriptors (or "main headings") organized in 16 hierarchies (e.g., Chemical and Drugs). Additionally, MeSH provides about 210,000 supplementary concept records (SCRs), of which many represent chemicals and drugs. Each SCR is linked to at least one descriptor. While most chemical descriptors provide a structural perspective on drugs, some descriptors play a special role as they can be used to denote pharmacological actions in drug descriptors and SCRs.
MeSH 2013 is used in this study.
RxNorm is a standardized nomenclature for medications produced and maintained by the U.
The UMLS is a terminology integration system created and maintained by the National Library of Medicine (NLM) (NLM, 2013). The UMLS Metathesaurus integrates over 150 terminologies, including MeSH, but not ATC. Synonymous terms across terminologies are grouped into concepts and assigned the same concept unique identifier. The Metathesaurus provides a comprehensive set of synonyms for biomedical concepts and is often used for integrating terminologies beyond its own. NLM provides an application programming interface (API) for accessing UMLS data programmatically. Version 2012AB of the UMLS is used in this study.
Our approach to aligning pharmacologic classes between MeSH and ATC based on their instances is depicted in In our alignment work, we excluded the 14 ATC groups of level 1 (anatomical classification), because they are too broad classes. We also excluded 164 of the 1,241 ATC groups (2 nd -4 th level) corresponding to drug combinations, because combination drugs are often underspecified in ATC.
Similarly, in MeSH, we excluded the top-level descriptors of the Chemicals and Drugs hierarchy (i.e., D01 -D27), as well as the top-level of the pharmacological action descriptors (Pharmacologic Actions, Molecular Mechanisms of Pharmacological Action, Physiological Effects of Drugs, and Therapeutic Uses).
We mapped all 1,077 eligible ATC classes (2 nd -4 th level) to MeSH descriptors in the Chemicals and Drugs [D] tree using the UMLS Terminology Services (UTS). More precisely, we used the ExactString and NormalizedString search function of the UTS API 2.0 to establish mappings from the names of the ATC classes to UMLS concepts. We used normalization only when the exact technique did not result in a mapping. We then mapped the UMLS concepts to MeSH descriptor IDs.
Mapping ATC drugs to RxNorm ingredients. In previous work we have mapped ATC single-ingredient drugs to Ingredients (IN) and Precise Ingredients (PIN) in RxNorm using a lexical approach with additional normalization steps (Winnenburg and Bodenreider, 2012). We used these mappings to establish the alignment of ATC and RxNorm drugs in this study.
We considered the hierarchical relations from 5 th level drugs to their 4 th level chemical groups as asserted drug class membership. We inferred membership between 5 th level drugs and groups of level 3 and 2 through transitive closure. For example, temafloxacin (J01MA05) is a member of the chemical group Fluoroquinolones (J01MA -asserted), the pharmacological group QUINOLONE ANTIBACTERIALS (J01M -inferred), and the therapeutic group ANTIBACTERIALS FOR SYSTEMIC USE (J01 -inferred).
Table 1. Asserted and inferred MeSH classes for the drug temafloxacin (C054745) with type of relationship to the drug and tree numbers in MeSH.
We associated each RxNorm ingredient (IN or PIN) with its corresponding MeSH supplementary concept record (SCR) or main heading (MH). In turn, we associated these drugs with their asserted classes. For an SCR, we considered its pharmacological actions, as well as the MeSH heading(s) mapped to. For a MH, we considered its pharmacological actions, as well as its direct ancestors. These constitute the asserted classes. We inferred membership between the drugs and higher-level descriptors in the MeSH hierarchy. For example, as shown in Table 1, the SCR temafloxacin has Anti-Bacterial Agents as pharmacological action and Fluoroquinolones as main heading mapped to. Form these asserted classes, we infer membership to Anti-Infective Agents (from Anti-Bacterial Agents) and to Quinolones, Quinolines, and Heterocyclic Compounds, 2-Ring (from Fluoroquinolones).
Based on the asserted and inferred class membership of drugs in ATC and MeSH we conducted a pairwise comparison of all ATC against all MeSH classes. For each pair of ATC class (A) and MeSH class (M), we computed the Jaccard coefficient.
In order to reduce the similarity of pairs of classes with a small number of shared members, we used a modified version of the Jaccard coefficient, JCmod, as suggested in (Isaac, et al., 2007),
where A ∩ M represents the number of drugs common to A and M, and A ∪ M the total number of unique drugs in both classes.
The Jaccard coefficient measures the similarity between the two classes, but does not reflect whether one class is included in the other. Because of the difference in granularity between classes in ATC and MeSH, we introduce a simple metric for detecting whether the drugs that are not shared by both classes are primarily in one of the two classes. This "one-sidedness" coefficient is calculated as follows:
0, for a = 0 and m = 0 |a-m| / a+m, otherwise.
where a and m are the number of drugs specific to the ATC class and the MeSH class, respectively. Thus, a "onesidedness" coefficient close to 0 indicates that the drugs that are not shared by the two classes are evenly distributed between the ATC and MeSH class. In contrast, a coefficient close to 1 indicates that only one of the classes contains most of the drugs that are not shared by the other.
In order to select the best equivalent or inclusion mappings between ATC and MeSH, we characterize each pair of ATC and MeSH classes with respect to Jaccard similarity and one-sidedness. Low one-sidedness indicates equivalence and high one-sidedness indicates inclusion. High Jaccard similarity indicates strong overlap between the two classes. Based on preliminary analysis, we selected of a threshold of 0.5 for the one-sidedness metric. Similarly, we selected of a threshold of 0.5 and 0.25 for Jaccard similarity for equivalence (low one-sidedness) and inclusion (high one-sidedness), respectively. The lower threshold for Jaccard similarity for inclusion was determined empirically. As shown in Table 2, each pair of ATC and MeSH classes is characterized as an equivalence mapping (EQ+), an inclusion mapping (IN+), or not a mapping (EQ-and IN-).
For the 1,077 eligible ATC groups of level 2-4, we were able to retrieve 226 mappings to descriptors from the Chemicals and Drugs [D] tree in MeSH. We have 18 mappings for therapeutic classes (2 nd level), 42 for pharmacological classes (3 rd level), and 161 for chemical classes (4 th level). We ignored mappings for the broad anatomical classes (1 st level). Of the 221 mappings, 96 are to pharmacological ac- tions (functional perspective) in MeSH, whereas 125 are to other descriptors at various levels of the MeSH hierarchy (structural perspective).
Of the 1,077 eligible ATC groups, 874 (81%) could be associated with at least one descriptor or pharmacological action in MeSH. We identified a total of 933 associations for the 874 ATC groups (multiple associations per ATC group possible). As shown in Table 2, based on the onesidedness metric, we characterized 323 associations as equivalence and 610 as inclusion. Of the 323 equivalence associations, 113 (35%) exhibit high Jaccard similarity and are selected as equivalence mappings (EQ+). Of the 610 inclusion associations, 230 (38%) exhibit high Jaccard similarity and are selected as inclusion mappings (IN+). The other associations (EQ-and IN-) are not deemed strong enough to denote mappings. In summary, we were able to characterize as a mapping (EQ+ and IN+) 343 (37%) of the associations between ATC and MeSH classes. It should be mentioned that we were not able to obtain mappings to MeSH classes for 203 ATC classes, because they only contain drug instances that could not be mapped to drugs in MeSH.
Table 2. Characterization of the associations between ATC and MeSH classes based on Jaccard similarity and score for one-sidedness. The numbers in grey fields indicate the associations that are not strong enough to denote mappings.
As illustrated in Table 3, from the 221 lexical mappings between ATC and MeSH classes, we could confirm 61 with our instance-based approach (30 as equivalence mappings, 31 as inclusion mappings). For 19 of the lexical mappings we found an association with low Jaccard similarity (IN-/ EQ -), and for 141 of the lexical mappings we did not find any association through the instance-based alignment (mainly due to the lack of any mapping for the drug instances in these classes). Finally, the instance-based approach produced 282 additional drug class mappings that were not detected by the lexical approach, whereas 633 (571 + 62) ATC classes could neither be mapped by the lexical nor the instance-based approach.
True positive for equivalent instance-based mappings. We identify an equivalence mapping between the 4 th -level ATC group Fluoroquinolones (J01MA) and the MeSH descriptor Fluoroquinolones (D024841). The two classes share 14 drugs.
The ATC group has one extra drug (moxifloxacin), and the MeSH descriptor has 2 (flumequine and besifloxacin). Jaccard similarity is high (0.82) and the one-sidedness score is low (0.33), because the 3 drugs that are not in common are not all on the same side. This mapping is also identified by the lexical technique (exact match).
We identify an inclusion mapping between the 4 th -level ATC group Fluoroquinolones (S01AE) and the MeSH descriptor Fluoroquinolones (D024841). Although the two classes are seemingly identical, our mapping is identified as an inclusion, with 7 drugs in common, 1 drug specific to the ATC class and 9 drugs specific to the MeSH class. In fact, the ATC class is not the same general class for anti-infective agents as in the example above (J01MA), but rather the specific class of fluoroquinolones for ophthalmic use (S01AE).
The fluoroquinolones used for eye disorders are a subset of all fluoroquinolones and the ATC class S01AE is appropriately characterized as being included in the MeSH class for fluoroquinolones. This example also illustrates a false positive for the lexical mapping, since it is generally assumed that lexical mappings are equivalence mappings.
Many ATC and MeSH classes share only one or very few drugs, making it difficult to assess equivalence or inclusion. For example, the 4 th -level ATC group Silver compounds (D08AL) and the MeSH descriptor Silver Compounds (D018030) share only one drug (silver). The modified version of the Jaccard coefficient has a score of 0.45 in this case, which is below our threshold of 0.5 for equivalence.
During this failure analysis, we discovered that some MeSH drugs did not have a pharmacological action assigned to them as we expected. For example, while pyrantel is listed as Antinematodal Agents, oxantel is not. We are investigating whether the pharmacological action for this SCR should be inferred from the descriptor to which it is mapped (Pyrantel in this case). Because of these missing pharmacologic actions, the 3 rd -level ATC group ANTINEMATODAL AGENTS (P02C) fails to be mapped to the MeSH pharmacological action Antinematodal Agents (D000969), the Jaccard similarity being just below the threshold (0.49).
Despite the use of UMLS synonymy and normalization, the lexical alignment fails to identify a mapping between the 3 rd -level ATC group POTASSIUM-SPARING AGENTS (C03D) and the MeSH pharmacological action Diuretics, Potassium Sparing (D062865). In contrast, the instance-based alignment identifies an equivalence mapping with very high Jaccard similarity (0.99). This finding is consistent with the conclusions of (Merabti, et al., 2011).
We have identified several causes for discrepancies between lexical and instancebased alignments. As mentioned earlier, some ATC classes only contain drugs that cannot be mapped to MeSH through RxNorm, which we used to bridge between the two. Sometimes, the best instance-based mapping is to another class than the class found by the lexical technique. Finally, some drugs entities and biologicals (e.g., vaccines) are less well standardized than common drugs. For this reason, the instance-based alignment is unable to map these classes, when simple lexical techniques can.
This exploratory investigation has several limitations, which we plan to address in future work.
Evaluation. This exploratory investigation focuses primarily on the methodology and feasibility of the alignment, and does not include a formal evaluation. Since ATC and MeSH pharmacological actions are being integrated into RxNorm, we will use the alignment created by RxNorm experts as the gold standard to evaluate our methods.
Perspective. Our perspective in this investigation is ATCcentric, because we consider the best MeSH mapping for each ATC class, but not the best ATC mapping for each MeSH class. One future goal is to explore both directions using the same methodology.
Bias towards equivalence mappings. Because we restrict our exploration to the MeSH class with the best Jaccard similarity for each ATC class (which we subsequently categorize as equivalence or inclusion), and because of the differential threshold for Jaccard similarity between equivalence (0.5) and inclusion mappings (0.25), we potentially fail to consider a good inclusion mapping (e.g., with a similarity score of 0.39 [> 0.25]), when the best MeSH class is a bad equivalent mapping (e.g., with a similarity score of 0.41 [< 0.5]).
To our knowledge, our work is the first attempt to align pharmacologic classes with instance-based techniques, distinguishing between equivalence and inclusion relations, as well as the first application of alignment between pharmacologic classes in ATC and MeSH. Our instance-based approach to aligning pharmacologic classes has yielded 343 mappings, and has the prospect of effectively supporting the creation of a mapping of pharmacologic classes between ATC and MeSH. This exploratory investigation needs to be evaluated in order to adapt the thresholds for similarity.
This work was supported by the Intramural Research Program of the NIH, National Library of Medicine and by the Center for Drug Evaluation and Research of the Food and Drug Administration. The authors want to thank Rave Harpaz and Anna Ripple for useful discussions.
The findings and conclusions expressed in this report are those of the authors and do not necessarily represent the views of the FDA.