Potential drug-drug interactions (PDDI) are a signi cant source of preventable drug-related harm. One contributing factor is that there is no standard way to represent PDDI knowledge claims and associated evidence in a computable form. The research we present in this paper addresses this problem by creating a new version of the Drug Interaction Knowledge Base, with scalable, interlinkable repositories for PDDI evidence and PDDI knowledge claims.
A challenging area of focus for patient safety is the management of potential drug-drug interactions (PDDIs). These are de ned as co-prescription or coadministration of two drugs known to interact, which potentially exposes the patient to adverse drug events [9]. PDDIs are a signi cant source of preventable drug-related harm: according to a recent review, clinically important events attributable to PDDI exposure occur in 5.3% to 14.3% of inpatients, and are responsible for 0.02% to 0.17% of the 129 million emergency department visits that occur each year [12].1 Unfortunately, most drug information sources disagree substantially in their guidance about speci c PDDIs [1, 16, 13, 2]. Addressing this is urgent as United States healthcare organizations consider PDDI screening in their strategies to achieve the e ective use of electronic health records. 1 http://www.cdc.gov/nchs/fastats/ervisits.htm
There are both technical and social factors underlying the disagreement that exists across drug information sources [14]. As Figure 1 shows, evidence that might be relevant for establishing PDDI knowledge claims is distributed across several sources including product labeling, the scienti c literature, and case reports. Each source provides complementary evidence that editors of drug information resources (public sources [2] or proprietary sources such as Micromedex, Epocrates, and Medscape) must synthesize. A major social factor underlying disagreement is that drug information editors have di erent criteria for assessing evidence. Fortunately, two di erent conference series have brought leading drug information editors to discuss a standard set of methods for assessing evidence [14, 10].
Pre$market)studies
Post$market)studies
Clinical) experience
Rarely(reported(in Reported(in
Reported(in
Rarely(reported(in Product)labeling
Source(for
Scientific) literature
Rarely(reported(in
Source(for Drug)Compendia) )synthesize)PDDI)evidence)into) knowledge)claims but • May)fail)to)include) important) evidence • Disagree)if)specific) evidence)items) can)support)
or)refute)PDDI)knowledge)claims
A major technical factor yet to be addressed is that there currently does not exist a standard way to represent PDDI knowledge claims and associated evidence in a computable form. As a result, drug information editors resort to ad hoc information retrieval methods that can yield di erent sets of evidence to assess [10]. The research we present in this paper addresses this problem by creating scalable, interlinkable repositories for both PDDI evidence and PDDI knowledge claims. This paper describes our new approach. In Section 2, we outline requirements. In Section 3, we discuss the technical details. In Section 4, we present a benchmark analysis that tests the ability of the new approach to scale. After discussion, we conclude the paper.
In prior work, we created the original Drug Interaction Knowledge Base (DIKBold) [3, 4]. The DIKB2 is an evidence-focused knowledge base designed to support pharmacoepidemiology and clinical decision support. It contains quantitative and qualitative knowledge claims about drug mechanisms and pharmacokinetic drug-drug interactions for over 60 drugs.
Prior work on the DIKB-old focused on development of an evidential approach representing the evidence associated with a scienti c claim. The system considers the evidence board as a socio-technical reasoning system that manages both a knowledge base and an evidence base. The knowledge base holds PDDI knowledge claims while the evidence base stores information artifacts that can be used to support or challenge those claims. PDDI knowledge claims may be direct (e.g., \drug X interacts with drug Y"); or inferred from pharmacological properties (e.g.,\drug X inhibits enzyme Q which is important for the clearance of drug Y from the body").
In prior work [3, 4], all evidence was collected and entered by an evidence board consisting of an informaticist and a minimum of two drug-experts. The board used the following process to manage the evidence and knowledge base components: 1. All members of the board select drugs of interest. This determines the set of PDDI knowledge claims to be investigated. 2. The informaticist conducts a systematic search for evidence that might support or refute the pre-determined PDDI knowledge claims. 3. Retrieved items are ltered by applying study inclusion criteria. 4. Evidence items that meet inclusion criteria are entered into the system where they are linked to speci c PDDI knowledge claims and any evidence use assumptions (knowledge claims that must be true for the evidence to hold). 5. A truth value for each knowledge claim is determined based on belief criteria.
Experience with the DIKB-old revealed a great need for improvements to the system that would make this process more e cient. First, a substantial amount of time was spent on reconciling and integrating information from various sources (Figure 1). Decision rationales were not recorded in a computable form and the evidence board did not have a process in place to keep up with relevant new evidence. Furthermore, the DIKB-old was ontologically informal, failed to adopt common biomedical ontology terms3, and did not distinguish drug and enzyme classes from individuals. This hindered automated reasoning that integrated external knowledge sources and resulted in treating PDDIs the same as observed 2 When we do not need to distinguish between the old (`DIKB-old') and new (`new
DIKB') versions, we simply mention `DIKB'. 3 For example, the DIKB-old used the predicate `substrate of' to represent the metabolic process of xenobiotic catalysis. However, this predicate was de ned without reference to the formally de ned biological process (e.g, such as that provided by the Gene Ontology). drug-drug interactions. In the new system we wished to resolve these issues. We also wished to retain the ability to compute with a logical representation of drug mechanism knowledge claims, using a rule-based theory of how to infer PDDIs from metabolic mechanistic knowledge of how drugs interact [4].
We summarize these as three requirements for the new DIKB: R1 Create a maintainable structure that supports evidence entry of data, methods, and materials from multiple sources on an ongoing basis.
R2 Create computable, logical representations of drug mechanism knowledge claims.
R3 Link to biological processes while also carefully distinguishing between a drug drug interaction (an actual occurrence in a patient) and a potential drug drug interaction (an information content entity that may exist because of an observation or inference).
Our approach to addressing these criteria are as follows: Addressing R1 We adopt the emerging Micropublications (MP) [6] model for literature integration using `argument graphs' to represent published claims as formal assertions linked to primary data and resources.
Addressing R2 We extended the MP ontology to add two new properties, MP:formalizedAs/MP:formalizes, to enable natural language claims to be linked to useful logical formalizations.
Addressing R3 To stress that potential drug drug interactions are information artifacts, we use a new ontology called DIDEO [5] which has several advantages. DIDEO: (a) Reuses identi ers from existing ontologies (e.g., CHEBI, PRO) that represent biological entities and processes; (b) Di erentiates between the representation (statements about drugs and drug-drug interactions) and the represented (actual drug-drug interactions); (c) Prevents unwanted existential import (further explained in Section 3.3 below); and (d) Distinguishes between the type of a drug or enzyme and portions of a speci c drug or enzyme, by using punning.
3.1
Create a maintainable structure that supports evidence entry of data, methods, and materials from multiple sources We used micropublications to create a structure that supports evidence entry of data, methods, and materials from multiple sources [15]. We now represent PDDI knowledge claims and supporting evidence as queryable RDF statements4 constructed using the Micropublication ontology (MP) [6]. PDDI knowledge claims 4 Queryable at http://purl.org/net/nlprepository/swat-4-med-safety-sparqlendpoint and evidence were transformed from the DIKB-old model into the new one using Python scripts. Drug identi ers were converted to ChEBI identi ers to enable the use of DIDEO. So far, the mapping has been completed for 70% of the drugs that had data from clinical studies or mechanistic experiments. We envision that additional DIKB micropublications could be created by multiple parties, including evidence boards, and potentially the original authors, as we describe in Section 5.
Figure 2 shows the generic form of a DIKB micropublication graph using the example erythromycin - simvastatin interaction. Notice that MP has rigorously de ned ontology classes that support the DIKB evidence curation process discussed above. In MP, the primary object of interest is the claim. A claim is supported by methods, materials, and data: { MP:Claim, a text string representing a scienti c claim. { MP:Method, representing a scienti c method. { MP:Materials, for materials, such as study participants and drugs. { MP:Data, such as the area under the concentration curve (AUC). These are used for entering evidence and later, the evidence is used to determine truth values for the claims that the evidence supports.
obo:CHEBI_48923 mp:qualifiedBy obo:CHEBI_9150 mp:qualifiedBy
obo:DIDEO_00000000 mp:qualifiedBy MP 1 mp:argues
Claim
1 mp:supports
Data 1 mp:supports erythromycin increases the AUC of simvastatin
mp:supports Method
1 mp:supports
Materials 1 http://dx.doi.org/ 10.1016/ S0009-9236(98)90151-5
The process for managing the evidence base and knowledge base described in Section 2 includes assessing the truth value of each PDDI knowledge claim using belief criteria. Operationally, the evidence board uses labels from a taxonomy of evidence types5 to tag each evidence item as it is entered into the evidence base. The board then decides which evidence types are credible for speci c types of PDDI knowledge claims: this speci es a belief criterion.
As an example, the evidence board might decide that, to support a claim that a drug is a substrate of an enzyme, only clinical drug-drug interaction studies 5 http://purl.org/net/drug-interaction-knowledge-base/evidence-typesand-inclusion-criteria are admissible. This would become a belief criterion for all `substrate of' claims. To implement a belief criterion in the new DIKB, the evidence base is queried to nd all PDDI knowledge claims that have at least one supporting evidence item meeting the criterion. The resulting claims are assigned the value of `True'. 3.2
PDDI knowledge claims mention speci c entities such as drugs, drug metabolites, enzymes, and biological pathways whose relationships with each other are more generally modeled in a rule-based theory that infers PDDIs [4]. Sources external to the DIKB provide additional formalized knowledge about these entities. For example, the Gene Ontology provides cellular location and molecular function for the enzyme CYP3A4; this is relevant when the evidence board seeks information about gene expression and about enzyme metabolization.
The spans of unstructured text in MP:Claim are not inherently computable entities, and the semantic quali ers (MP:qualifiedBy) cannot specify the order (i.e. separate the object drug from the precipitant drug). Therefore, we extended the MP ontology to add two new properties, MP:formalizedAs/MP:formalizes, that enable natural language claims represented as MP:Claim to be linked to their logical representation.
RDF is also the language chosen for the formalization of MP:Claim resources, so that a single query language (i.e., SPARQL) can be used to retrieve information from the whole evidence base. We chose to represent the logical form of knowledge claims using OWL for two reasons. First, OWL provides classes and properties that enable the representation of logical statements in RDF. Second, logical statements written in OWL can be checked for logical consistency and new inferences by a reasoner such as Hermit [7].
We chose to formalize claims using the Nanopublication (NP) [8] ontology because: 1. NP provides a class called NP:Assertion that can hold any RDF graph, including logical statements written in OWL. 2. OWL logical statements stored as an NP:Assertion can be integrated into full nanopublications that combine the NP:Assertion, the provenance of the assertion, and the provenance of the nanopublication into a single publishable and citable entity.
A nanopublication represents the logical structure of a claim as an RDF graph. Like micropublications, nanopublications are publishable and citable entities. Their citability and use of provenance enable us to make the evidence review process transparent and auditable. The uptake of nanopublications by the wider community suggests that nanopublication is a relevant publishing mechanism for reconsumption by others.6 Unlike micropublications, nanopublications have no 6 One measure of uptake is the variety of authors of papers using nanopublications; see the bibliography at http://nanopub.org/wordpress/?page_id=638. Another is the size and geographic distribution of current nanopublication datasets: see [11] Table 1 and Figure 3, respectively. explicit evidence structure and do not support claim con ict. They are therefore complementary to micropublications, which provide these missing features. 3.3
Reasonable extrapolation is an important way to infer a PDDI. In contrast to drug-drug interactions (DDIs) that are based on observing an actual drug-drug interaction in some patient, inferred PDDIs based on reasonable extrapolation might not actually occur in reality. Since we do not know whether a PDDI occurs, we cannot assume the existence of any instance of a drug interaction. To model this correctly, we di erentiate between actual drug interactions and statements about PDDIs using the DIDEO ontology [5].
erythromycin molecularly6 decreases6activity
CYP3A4 catalyzes6a6Phase6I6or6 Phase6II6enzymatic6 reaction6involving simvastatin obo:CHEBI_48923 obo:RO_0002449 obo:PR_000006130 obo:DIDEO_00000096
obo:CHEBI_9150 inhibits<catalyzes6 metabolism6 of
obo:DIDEO_00000090
Let us again consider erythromycin and simvastatin. Previously (Figure 2) we described a clinical study establishing this PDDI. We now reconsider this drug pair (e.g. in the absence of that clinical evidence), to show how a computable structure and OWL inferencing allow the new DIKB to compute a PDDI by reasonable extrapolation, as shown in Figure 3.
Assume that credible evidence also establishes that erythromycin inhibits the catalysis of CYP3A4 which is important to the metabolic clearance of simvastatin. We represent this relationship in the evidence base by creating OWL statements representing the following two knowledge claims (Figure 3): \erythromycin moleculary decreases activity of CYP3A4" and \CYP3A4 catalyzes a Phase I or Phase II enzymatic reaction involving simvastatin". These claims are linked to the relevant evidence support using MP as discussed above, and formalized by two di erent NP:Assertion resources using RDF/OWL. If the evidence support meets the belief criteria speci ed by the evidence board, full nanopublications are created for the formalized claims. These are then passed to an OWL reasoner that would infer a relation between erythromycin and simvastatin called inhibits-catalyses metabolism. This inference generates a new entity, namely an individual PDDI, based on the inferred statement, as shown on the left side of Figure 4. This individual resides in the knowledge base. The knowledge base contains other individuals based on inferred statements; it also contains entities that represent individual PDDIs based on data items created by clinical studies, clinical observation or physiological experiments. The key point here is that the inferred individuals do not refer to any class of speci c drug-drug interactions, and so they do not imply the existence of at least one member of the class (known as existential import).
Evidence(Base "erythromycin,inhibits,CYP3A4." "CYP3A4,catalyzes,a,Phase,I,or, Phase,II,enzymatic,reaction, involving,simvastatin."
OWL inference
Knowledge(Base "Statement,A,is,about, {erythromycin}." "Statement,A,is,about,{simvastatin}." "Statement,A,is,about,{CYP3A4}." "Statement,A,is,a,PDDI,statement."
Statement,A transformation
We also distinguish between references to a type and references to a class of collection of things. As an example, consider the two statements: \simvastatin treats dyslipidemia." and \simvastatin was used in the treatment of patient John Doe." The latter sentence refers to a collection of speci c portions of simvastatin, whereas the former sentence refers to the entire type. In the context of PDDIs this allows us to di erentiate speci c portions of a drug used in for example a case study from the drug as a type. Traditionally, OWL represents classes and individuals, but not types. However, OWL 2 enables the representation of types using `punning'.7
Punning means that we use a URI that is assigned to a class (for instance the class `simvastatin') and also assign it to an individual. For instance, the ChEBI and PRO URIs are used to pun individuals that are intended to represent the type of drug or enzyme. To represent the result of the inference shown in Figure 3, we create a new individual which we will be able to link to both the drugs and the enzyme involved. As is shown in Figure 4, we use the punned URI for the ingredients, drug products and enzymes involved, to express that we are talking about the types, not about actual portions of them. 7 http://www.w3.org/TR/owl2-new-features/#F12:_Punning
To evaluate the feasibility of using the model in practice, we investigated how long querying will take. As a benchmark, we execute a set of 31 queries8 across 7 di erent sizes of MP graphs9, from a single Virtuoso endpoint. Results are shown in Figure 5.
To generate G0, the rst MP graph, the original DDI data is extracted from the DIKB-old and reshaped into the MP model, along with its supporting evidence. We include DIKB-old data with assertions that are typed as `inhibits', `substrate of', `increase AUC', making G0 2.0 MB with 16,670 triples and 3498 entities. We also generated 6 additional graphs, with Gi containing 2i (i = 1 to 6) copies of the G0 graph. Claims, methods, materials, and data in these MP graphs have the same relationships as in G0 so that all items are evenly distributed in the testing graphs. One limitation is that, although this is a real-world graph, we scale by making multiple copies of the same entities and relationships. 5
Drug information resources are created by groups, which may include an information specialist, drug specialists, etc. Currently, multiple di erent editorial 8 http://purl.org/net/drug-interaction-knowledge-base/micropublicationqueries 9 http://purl.org/net/drug-interaction-knowledge-base/published-NP-and
MP-graphs boards undertake this process entirely independently. Evidence from scienti c studies, new drug applications, and product labeling is rst searched for, and then evaluated. Complex cases are reviewed by multiple people in order to resolve issues, such as to determine whether inclusion criteria are met.
Currently, each group must independently spend a substantial amount of time on reconciling and integrating information from various sources. This slows the process, because multiple di erent systems, including literature alerts, spreadsheets, and external databases, must be combined in order to evaluate whether a PDDI should be recorded. One new potential is for some of the work of centralizing information to be shared, even across di erent proprietary sources. Even for commercial entities, shared data resources can be bene cial: For instance, the Open PHACTS public-private partnership has received commercial funding to pre-process and integrate certain aspects of pharmaceutical data into nanopublications [17], which saves individual drug companies time and resources in preparing proprietary data.
Maintainability Maintainability will also be a key factor in ensuring the success of the new DIKB. We have already transformed existing annotations into the new format. We are starting to construct a user-friendly annotation process, so that evidence board members and their delegates can annotate knowledge claims and primary data. Our ongoing research will suggest optimal approaches for publishing PDDI claims and evidence in a structured form. • Scientific4literature • Product4label • Other…
Annotating4 while4publishing
Primary4data
Argument4graphs
Authors4 annotate44 during4 the4publication4 process4 444
Evidence4 board4 annotates while4 creating4a4new4drug4information4 sources
Annotating4 published4 documents
Primary4data
Argument4graphs
Micropublication evidence4 item Claim • Drug4X4interacts4with4drug4Y • Drug4X4inhibits4enzyme4Q
• Data Support • Materials • Methods • Scientific4literature Reference • Product4label • Other…
Structured publishing is of increasing interest, and the two main sources of PDDI evidence (product labels and scienti c papers) are already moving toward more structured representations. All drug product labels in the United States are written by drug companies and submitted to the U.S. Food and Drug Administration using the Structured Product Labeling (SPLs) standard10. Scienti c publishers may in the future deeply interlink scienti c papers and data; this is a proposed service of Elsevier, the largest scienti c publisher11. For authors to annotate their own work with DIKB-compatible micropublications, and for publishers to redistribute micropublications, thus seems conceivable.
If either authors or drug editorial boards share DIKB-compatible micropublications, other groups who need to process the same evidence would also bene t because they would not need to separately integrate and reconcile the information artifacts that support PDDI knowledge claims. This greatly improves maintainability.
Shared Representation In the contested area of drug compendium information, centralizing collective information has another bene t: it can help make visible disagreements and discrepancies between these sources, which can be di cult to resolve. Currently, there is no way to answer questions such as: did compendium A and compendium B consider the same evidence? Only the nal assertions{ lists and severity rankings of drug-drug interactions{are exposed; the underlying evidence is not retrievable or comparable, without directly approaching evidence boards, one by one.
We see the potential for deeper structures, that represent intermediary steps of the process in a shareable format. The evidence base of the new DIKB is a shared representation to centralize and unify the information, and to support tasks such as commenting on evidence, rejecting evidence, and asking for opinions about evidence. This would also allow, for instance, citing a particular statement, indicating a disagreement with it, or marking some evidence as more credible than others. We believe that a computable, shared representation could enable compendium editors to more easily integrate and cross-check di erent sources of information { as well as to compare di erent potential interpretations, amongst di erent editorial groups. 6
As we have discussed, the new DIKB has several advantages compared to the DIKB-old. It is designed to meet 3 key requirements: maintainability, computability, and the ability to link to biological processes while retaining logical consistency.
In future work we will test whether the DIKB does in fact simplify the editor's task of searching and synthesizing clinically relevant PDDI information. In particular, we aim to show how clinicians and others interested in PDDIs 10 http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/
Guidances/ucm072317.pdf 11 http://libraryconnect.elsevier.com/articles/best-practices/201302/research-data-driving-new-services can use the DIKB to quickly identify the rationale for any given assertion of knowledge. We plan to compare existing tools to new search tools derived from the new DIKB, in terms of the completeness of information retrieval by drug experts using the tools.
One major advantage of the DIKB is that its evidence base serves as a shared representation, recording the evidence assessed and the results taken into account. To further test the relevance of recording this information explicitly, in the future we will explore an architecture enabling multiple `possible worlds' to be described: multiple knowledge bases, generated from di erent belief criteria.
We hope that in the future some of the bookkeeping work of compedium creation can be shared{ideally amongst authors annotating their own materials{and certainly amongst editors of di erent compendia. We will advocate for authors to annotate publications of various sorts with DIKB-compatible micropublications once appropriate tools can be made available. This route has great potential to reduce much of the bookkeeping work conducted by the evidence board because PDDI knowledge claims and the data supporting those claims would be annotated by the individuals who wrote the submission. It would also help evidence avoid becoming `stale' because the new evidence would be linked directly to knowledge claims as they are published.
Since our approach is compatible with other ontologies in the Open Biomedical Ontologies (OBO) Foundry through the reuse of identi ers, the import of external information is simpli ed compared to the previous approach. With the new DIKB architecture, as gene and protein knowledge is formalized in external sources, that knowledge can be pulled in and queried, in ways that may meet future use cases for other audiences.
The research leading to these results has received funding from the European
Union Seventh Framework Programme (FP7/2007-2013) under grant agreement
no 246016 for the ERCIM \Alain Bensoussan" Fellowship Programme; from a
grant from the National Library of Medicine (1R01LM011838-01); and training
grant 5T15LM007059-29 from the National Library of Medicine/National
Institute of Dental and Craniofacial Research. We thank Carol Collins, Amy Grizzle,
Lisa Hines, Harry Hochheiser, and John R Horn.
3. Boyce, R., Collins, C., Horn, J., Kalet, I.: Computing with evidence: Part I. Journal
of Biomedical Informatics 42(6), 979{989 (2009)
4. Boyce, R., Collins, C., Horn, J., Kalet, I.: Computing with evidence: Part II.
Journal of Biomedical Informatics 42(6), 990{1003 (2009)
5. Brochhausen, M., Schneider, J., Malone, D., Empey, P., Hogan, W.R., Boyce, R.D.:
Towards a foundational representation of potential drug-drug interaction
knowledge. In: Drug Interaction Knowledge Representation at ICBO (2014)
6. Clark, T., Ciccarese, P., Goble, C.: Micropublications: a semantic model for claims,
evidence, arguments and annotations in biomedical communications. Journal of
Biomedical Semantics 5(
Automated Reasoning 53(3), 245{269 (2014) 8. Groth, P., Gibson, A., Velterop, J.: The anatomy of a nanopublication. Information
Services and Use 30(