Drug-drug interaction discovery and demystification using Semantic Web technologies

Abstract

Objective: To develop a novel pharmacovigilance inferential framework to infer mechanistic explanations for asserted drug-drug interactions (DDIs) and deduce potential DDIs.

Materials and Methods: A mechanism-based DDI knowledge base was constructed by integrating knowledge from several existing sources at the pharmacokinetic, pharmacodynamic, pharmacogenetic, and multipathway interaction levels. A query-based framework was then created to utilize this integrated knowledge base in conjunction with 9 inference rules to infer mechanistic explanations for asserted DDIs and deduce potential DDIs.

Results: The drug-drug interactions discovery and demystification (D3) system achieved an overall 85% recall rate in terms of inferring mechanistic explanations for the DDIs integrated into its knowledge base, while demonstrating a 61% precision rate in terms of the inference or lack of inference of mechanistic explanations for a balanced, randomly selected collection of interacting and noninteracting drug pairs.

Discussion: The successful demonstration of the D3 system’s ability to confirm interactions involving well-studied drugs enhances confidence in its ability to deduce interactions involving less-studied drugs. In its demonstration, the D3 system infers putative explanations for most of its integrated DDIs. Further enhancements to this work in the future might include ranking interaction mechanisms based on likelihood of applicability, determining the likelihood of deduced DDIs, and making the framework publicly available.

Conclusion: The D3 system provides an early-warning framework for augmenting knowledge of known DDIs and deducing unknown DDIs. It shows promise in suggesting interaction pathways of research and evaluation interest and aiding clinicians in evaluating and adjusting courses of drug therapy.

BACKGROUND AND SIGNIFICANCE

Drug-drug interactions (DDIs) are often complex processes that depend on many clinical and other factors.¹ While some DDIs are intentionally sought for clinical benefit,² unintended DDIs constitute a major cause of adverse drug effects (ADEs), which represent a global health burden due to the costs of consequential hospitalization, morbidity, mortality, and health care utilization.^3,4 In 2013, ADEs accounted for 711 232 cases of serious illness in the United States, with 117 752 of those cases resulting in death.⁵ Unfortunately, on the present course the number of DDI-related ADEs can only be expected to increase, due to a continual rise in the number of drugs being prescribed to individual patients, which has been shown to increase the risk of experiencing major DDIs.^6,7 There are several factors that work against early identification of potential DDIs, with one obvious factor being the limited availability of DDI information before drugs reach the market.^8,9 Because clinical trials are designed to study the safety and efficacy of new medications,¹⁰ patients taking drugs that might interact with the trial medication are often explicitly excluded from studies.

In addressing this shortcoming, pharmacovigilance methods have shown promise in investigating potential DDIs. These methods require germane DDI information that may be embedded in text form in scientific articles, the US Food and Drug Administration (FDA) Adverse Event Reporting System, electronic health records (EHRs), and drug information sources. Despite comprehensive approaches, pharmacovigilance studies have captured only a portion of available DDI information.

Why potential DDIs are still challenging to discover

One of the primary limitations in identifying potential DDIs is the single-focus approach of current pharmacovigilance research endeavors. For example, although Preissner et al.¹¹ developed a comprehensive database for cytochrome P450 enzymes along with their roles in causing DDIs, the limited focus of the study resulted in neglecting less-common and yet important pathways.^12,13 Of additional concern is the failure to detect multipathway interaction DDIs (MPIs) that occur as a result of 2 or more interactive mechanisms, such as between statins and cyclosporine.¹⁴ Recent works used electronic health care data combined with public data sources to prioritize DDIs, but these focused on only 10 clinical adverse events and were limited to 345 drugs.¹⁵ Another profound challenge in understanding DDIs is the lack of reporting of their mechanistic causes. For instance, although Jiang et al.¹⁶ developed and evaluated a Semantic Web–based approach for mining severe DDI-induced ADEs, their system is unable to provide mechanistic explanations for the interactions. There remains a pressing need to study DDI mechanisms beyond focusing solely on drug metabolism and single-pathway interactions.¹⁷ Indeed, there is a need for pharmacovigilance research focused on modeling both pharmacokinetic and pharmacodynamic DDI mechanisms.¹⁸ However, the disparate nature of the sources of information currently available for DDI discovery poses a significant challenge. Ayvaz et al.¹⁹ found a surprisingly small overlap of asserted DDIs across 14 data sources.

This article describes a prototype pharmacovigilance framework termed D3 (drug-drug interaction discovery and demystification), which aggregates DDI information sources and infers plausible, mechanistic explanations for asserted DDIs. The framework uses Semantic Web technologies to integrate community-created biomedical resources. The objectives of this study were to infer potential pathways causing known DDIs and deduce potential DDIs based on both common and uncommon pathways.

MATERIALS AND METHODS

Data sources

The scientific goal of the D3 framework is to support mechanistic study of DDIs at multiple interaction levels. These levels comprise known contributors to DDIs: pharmacokinetic, pharmacodynamic, pharmacogenetic, and multipathway interaction.

The D3 framework integrates 12 biomedical resources: the Unified Medical Language System (UMLS), a terminology-integration system produced by the National Library of Medicine²⁰; DrugBank, a database containing detailed information about drugs and their targets²¹; The Pharmacogenomics Knowledgebase (PharmGKB), a database of pharmacogenetics knowledge²²; the National Drug File–Reference Terminology (NDF-RT), a reference terminology to describe drugs with respect to pharmacokinetics, pharmacodynamics, physiology, and disease knowledge²³; the National Cancer Institute thesaurus (NCIt), a vocabulary for medical and translational research with a focus on cancer²⁴; Gene Ontology (GO), an ontology to describe gene products in terms of their molecular functions, biological processes, and cellular components²⁵; Universal Protein Resource (UniProt)/Swiss-Prot, an integrated knowledge base about proteins²⁶; National Center for Biotechnology Information (NCBI) Gene, a comprehensive gene-specific database with a focus on completed sequenced genomes from the National Center for Biotechnology²⁷; the Gene-GO file from NCBI, a text file that includes gene–gene ontology relationship information; side effects resource (SIDER), a database of drug indications and side effects that was text-mined from structured product labels²⁸; LInked-Drug-Drug-Interactions (LIDDI), a nanopublication-based Resource Description Framework (RDF) dataset that integrates 8 DDI sources, including EHRs at Stanford University²⁹; and a DDI source by Ayvaz et al.

The D3 framework also encompasses 15 DDI databases as sources of asserted DDIs for use in inference and in confirmation during evaluation. They are: Crediblemeds,³⁰ Drug Interaction Knowledge Base,³¹ PK-Corpus,³² National Library of Medicine (NLM) Corpus,³³ DDI-Corpus-2011,³⁴ Office of the National Coordinator for Health Information Technology (ONC)-HighPriority,³⁵ DDI-Corpus-2013,³⁶ Twosides,³⁷ DrugBank, ONC-Non-interruptive,³⁸ KEGG DDI,³⁹ SemMedDB,⁴⁰ OSCAR,⁴¹ NDF-RT, and LIDDI.

Constructing D3

We use Semantic Web technologies to represent, integrate, store, and query the biomedical resources.³⁰ Data are represented using the RDF, a graph-like formalism amenable to automated reasoning.³¹ We use Uniform Resource Identifiers (URIs) to uniquely identify resources. To integrate data, we map resources to UMLS Concept Unique Identifiers (CUIs). We use Jena to construct, store, and query the D3 knowledge base.^42,43 Data are queried using the RDF-friendly Protocol and RDF Query Language (SPARQL).⁴⁴

Extracting data from UMLS into D3

We use the UMLS Metathesaurus dataset version 2015AA from the UMLS knowledge server, hosted as a local MySQL database. We use 4 tables from Metathesaurus: MRCONSO to retrieve entity names, MRREL to extract semantic relationships (predicates), MRSAT to identify reference for CUIs, and MRSTY to examine the quality of mappings by ensuring the use of correct semantic types.⁴⁵ For all UMLS terminologies, we use MySQL queries to obtain their corresponding UMLS CUIs. Using the UMLS, we generate an RDF network consisting of NDF-RT, NCIt, GO, UniProtKB/Swiss-Prot, and NCBI Gene data.

Integrating external sources into the D3 knowledge base

We augment the mechanistic knowledge in D3 with additional non-UMLS sources including DrugBank, LIDDI, SIDER, and a DDI source by Ayvaz. All sources are mapped to the UMLS using database cross-references except LIDDI, which provides UMLS CUIs, and NCBI (gene-GO relations), which has GO and NCBI identifiers.

Mapping through cross-references

RDF versions of PharmGKB, DrugBank, and SIDER sources were obtained from Bio2RDF (May 10, 2015, release 3).⁴⁶ Mappings between source identifiers and UMLS CUIs are obtained by examining cross-reference identifiers in the source data and querying the MRSAT table to find a candidate UMLS CUI and check for the correct semantic type by querying the MRSTRY table. For example, BE0001032 is a DrugBank identifier for ABCB1, which is linked to P08183, a UniProt identifier. This identifier is linked to C0376622, a UMLS CUI for ABCB1 with a semantic type of gene or genome (T028). Finally, we use a SPARQL/update⁴⁷ to replace the DrugBank identifier with the CUI-containing URI.

Adding semantic relationships

To simplify semantic querying in D3, we assign a basic set of semantic relations between the semantic types. For instance, we follow previous work by collecting 4 relationships in NDF-RT (may_treat, may_prevent, may_diagnose, and induces) into 1 causal relationship (has_indication) between drug-disease pairs.^48,49 The result of this integration yields 32 classes (genes, drugs, diseases, etc.) and 116 relationships for 35 158 drugs with 6741 unique ingredients and more than 100 000 interaction pairs. Figure 1 illustrates the aspirin resource in the D3 knowledge base after enriching it with all available properties from external biomedical resources.

Figure 1.

A graphical representation of aspirin resource, which enriches by integrating needed properties from different biomedical sources in the D3 knowledge base. Each box contains the original source of information that was mapped to D3 through UMLS.

D3 inferential query-based framework

The D3 query-based inferential framework attempts to provide a mechanistic explanation for any pair of drugs for which it is queried. Figure 2 illustrates an overview of the approach. D3 first determines whether the pair of drugs have been asserted to have a DDI in any of the 15 DDI databases it encompasses. If so, it utilizes the resources it has incorporated to infer and return which, if any, of the pathways of which it has awareness might account for the interaction using 9 different mechanisms, discussed in detail below (D3 interactions). If not, the D3 framework attempts to deduce whether a DDI might exist between the queried pair of drugs based on its incorporated resources and the 9 different mechanisms. To accomplish these tasks, the D3 framework utilizes the SPARQL query language for RDF graphs to produce conclusions only if requisite conditions are satisfied. The SPARQL query syntax can be understood as finding a path in RDF graphs by utilizing a set of semantic relationships. For example, D3 can query the knowledge base to identify a metabolic inhibition as follows: Find a drug that is a substrate of an enzyme and another drug that is an inhibitor of this enzyme.

Figure 2.

The flowchart of the D3 framework with an example of an interaction between digoxin and rifampin. Names of drugs involved in an interaction are resolved to UMLS CUIs, and are then used to query the database of known DDIs and explanations for DDIs.

D3 interaction inferences

Pairs of drugs are analyzed against 9 different interaction mechanisms divided into 4 levels: pharmacokinetic (protein binding, metabolic inhibition, metabolic induction, transporter inhibition, and transporter induction); pharmacodynamic (additive-enhancement and competition); pharmacogenetic (single-nucleotide polymorphisms [SNPs] that may alter drug exposure); and MPIs. The 9 D3 inferences based on these 9 mechanisms are discussed in Table 1.

Table 1.

D3 framework interaction inferences explained

Interaction inference	Level	Clinical definition	Conditions for interaction	Fact sources	Clinical example
Protein binding	Pharmacokinetic	When 2 drugs have high affinity (>70% of the drug is bound) to bind to the same protein	x,z bind_to y x,z >70%	DrugBank	Nitazoxanide/warfarin50
Metabolic induction		When 1 drug induces the metabolic processing of the other	x metabolized_by y z induces y	DrugBank, NCIt NDF-RT, and UniProtKB/Swiss-Prot	Rivaroxaban/rifampin51
Metabolic inhibition		When 1 drug prevents the metabolic processing of the other	x metabolized_by y z inhibits y	DrugBank, NCIt and NDF-RT, and UniProtKB/Swiss-Prot	Warfarin/erythromycin52
Transporter induction		When 1 drug induces the transport of another, altering its elimination	x transported_by y z induces y	DrugBank	Carbamazepine/talinolol53
Transporter inhibition		When 1 drug inhibits the transport of another, altering its elimination	x transported_by y z inhibits y	DrugBank	Quinidine/ digoxin54
Multiple pathway	Multiple pathway	When both drugs share at least 1 enzyme and 1 transporter	x metabolized_by y x transported_by y2 z metabolized_by y z transported_by y2	DrugBank, NCIt NDF-RT, and UniProtKB/Swiss-Prot	Cyclosporine/ statins14
Competitive pharmacological	Pharmacodynamic	When both drugs act on the same pharmacologic target	x targets y z targets y	NCBI Gene, Gene-GO, NCIt, GO, and DrugBank	Propranolol/ albuterol55
Additive pharmacodynamic		When both drugs share the same mechanism of action, or have same side effects profile	x share_moa_with y z share_moa_with y OR x has_side_effects y z has_side_effects y	NCIt, NDF-RT, UniProtKB/Swiss-Prot, and SIDER	Glyburide/ metformin56
Pharmacogenetic	Pharmacogenetic	When 2 drugs that share a common metabolic pathway or a transporter are given to a patient with a polymorphism in the drug metabolizing gene or transporter known to alter drug exposure	x associated with _SNPs y z associated with _SNPs y	PharmGKB	Metoprolol/ paroxetine57

Nine different inferences are utilized by the D3 framework to explain asserted DDIs and deduce potential DDIs. In the table, x is a drug (object), z is a drug (perpetrator), and y is a target common to x and z, which can be a protein, enzyme, transporter, mechanism of action, side effect, or SNP.

Evaluation

We evaluated the D3 framework in 4 ways:

1. Assessment of overall inferential coverage of DDI databases.

   1. For each of the 15 DDI databases, count the number of DDIs asserted by that database that satisfy at least 1 of the 9 D3 inferences, and divide by the total number of DDIs asserted by that database to determine the recall rate for the database in terms of some mechanistic explanation.
   2. For the union of the 15 DDI databases, count the number of DDIs asserted by the union that satisfy at least 1 of the 9 D3 inferences, and divide by the total number of DDIs asserted by the union to determine the recall rate of the union in terms of some mechanistic explanation.
2. Assessment of individual inferential coverage of DDI databases.

   1. For each of the 15 DDI databases for each of the 9 D3 inferences, count the number of DDIs asserted by that database that satisfy that inference, and divide by the total number of DDIs asserted by that database to determine the recall rate of the database per the inference.
3. Standard recall, precision, and F-measure.

   1. For a set of 100 asserted DDIs selected at random from the Crediblemeds and ONC-HighPriority DDI databases (chosen for the high regard in which they are held), count the number of asserted DDIs that satisfy at least 1 of the 9 D3 inferences as true positives and the remainder as false negatives.
   2. For a set of 100 drug pairs known not to interact, selected based on a professional pharmacological review by 1 of the authors (AA), count the number of deduced DDIs that satisfy at least 1 of the 9 D3 inferences as false positives and the remainder as true negatives. It should be noted that the lack of a database of noninteracting drug pairs is a major challenge in evaluating DDI deduction performance.
   3. Using these counts, determine standard recall, precision, and F-measure.
4. Evaluate in the context of existing literature.

   1. D3 was evaluated against findings in existing literature for a well-studied pair of drugs that exhibit a DDI (ibuprofen and aspirin).
   2. In an ongoing study, D3 was used to deduce a list of drug pairs with potential interactions via a semantic rule–based model. Irinotecan-levofloxacin was 1 of the most significant pairs deduced to interact.

RESULTS

Low overlap of asserted DDIs among DDI databases

As reported by others,^15,19,58 overlap of asserted DDIs among popular DDI databases is generally low. We examined the overlap between each of the 15 DDI databases by computing an overlap index as the number of the DDIs asserted by both databases divided by the size of the smaller of the 2 databases. We found the cumulative distribution of the overlap indexes of pairs of databases based on weighting the overlap index of each database pair according to the total number of DDIs asserted by both sources. This formula was used to provide a fair means of evaluating overlap based on source sizes. From this distribution, we found that >50% of the pairs of databases had 16% or less overlap and >80% had ≤38% overlap, indicating a high level of diversity of asserted DDIs across the 15 DDI databases encompassed by D3 (Supplementary Material 1).

D3 provides a high recall rate in terms of some mechanistic explanation for the asserted DDIs in the 15 DDI databases it encompasses

We examined how well the D3 framework can suggest at least 1 mechanistic explanation for each of the DDIs asserted within the 15 DDI databases D3 encompasses (see Table 2). In these terms, D3 demonstrated generally excellent recall rates for the 15 DDI databases.

Table 2.

D3 recall rates for the 15 DDI databases it encompasses in terms of some mechanistic explanation

DDI source	Source type	Number of DDIs	Recall rate (%)
Crediblemeds	Clinical	82	100
DIKB	Bioinformatics	560	99
PK-Corpus	Text	165	94
NLM Corpus	Text	246	92
DDI-Corpus-2011	Text	569	89
ONC-HighPriority	Clinical	1150	85
DDI-Corpus-2013	Text	1282	84
Twosides	Bioinformatics	63 333	79
DrugBank	Bioinformatics	11 840	75
ONC-Non-interruptive	Clinical	2079	75
KEGG DDI	Bioinformatics	26 328	65
SemMedDB	Bioinformatics	3536	62
OSCAR	Clinical	7753	55
NDF-RT	Clinical	9883	50
LIDDI	Bioinformatics	4734	33

We also examined the individual recall rates of the 9 D3 inferences per DDI database to provide a breakdown of the concentration of interactions asserted within each database (see Table 3). Certain mechanisms were found to be clearly more prominent than others for the asserted DDIs. Whether this was due to the focus of the research into DDIs or the relative rarity of certain types of interactions is unclear. It should be noted that the recall rates shown in Table 3 were determined without mutual exclusivity for mechanistic explanation for individual asserted DDIs, meaning that a DDI with more than 1 mechanistic explanation would count toward all relevant inferences. These results indicated to us that multiple pathways for individual DDIs might be considerably more common than we had first predicted.

Table 3.

D3 recall rates for the 15 DDI databases it encompasses for each of the 9 D3 inferences in terms of mechanistic explanation

DDI Source	Protein binding (%)	Metabolic induction (%)	Metabolic inhibition (%)	Transporter induction (%)	Transporter inhibition (%)	MPIs (%)	Competitive (%)	Additive (%)	Pharmacogenetic (%)
CredibleMeds	37	38	90	13	35	35	0	98	22
DIKB	49	41	99	18	35	36	1	96	28
PK-Corpus	38	44	84	31	48	50	1	85	26
NLM-Corpus	33	28	52	32	48	40	1	83	14
DDI-Corpus-2011	25	33	55	15	30	23	1	78	12
ONC-HighPriority	22	19	60	8	24	22	2	59	11
DDI-Corpus-2013	28	29	52	15	28	24	1	73	11
Twosides	22	12	29	4	13	9	1	77	5
DrugBank	23	18	47	8	16	14	3	53	6
ONC-Non-interuptive	27	7	18	3	16	7	1	58	4
KEGG DDI	20	15	31	5	12	10	2	48	4
SemMedDB	7	11	24	7	22	12	5	47	4
OSCAR	14	7	20	3	9	6	2	43	2
NDF-RT	18	14	32	7	14	13	2	38	4
LIDDI	12	4	13	1	4	3	1	31	3

Precision, recall, and F-measure of D3 show its usefulness for studying DDIs in terms of mechanisms of interaction

D3 demonstrated an 85% recall rate, 61% precision rate, and 70% F-measure in terms of standard definitions for these measurements. We believe that the recall rate was negatively impacted by the limited number of inferences utilized compared with the potentially large number of possible inferences for providing mechanistic explanations of DDIs. We believe that the precision rate was negatively impacted by some of the inferences having an overly broad scope and thus being overly sensitive to their conditions (or perhaps overly insensitive to individual circumstances of specific drug pairs). We believe that further refinement of these inferences in future work will show promise in raising both the recall and precision rates in terms of including more DDIs that currently fall out of scope and being more sensitive to circumstances of drug pairs that might preclude an otherwise indicated interaction. In its current form, we believe that the D3 framework based on the mechanisms of interaction it currently considers is fully capable of providing reasonable suggestion of mechanisms of interaction for asserted DDIs, deducing potential interactions between pairs of drugs that do not have asserted DDIs, and evaluating DDIs deduced using other methods to determine whether they are without probable merit.

Inferred pathways for asserted DDIs are significant for explaining DDIs: a case study of aspirin and ibuprofen

Aspirin and ibuprofen have been used for decades to relieve pain and other symptoms. A study was published by the MacDonald group that showed that co-administration of both drugs in cardiovascular patients led to increased all-cause mortality and cardiovascular mortality compared with patients who had received aspirin only.⁵⁹ In 2006, the FDA warned about interactions between these 2 agents.⁶⁰ Moreover, the mechanisms of interaction of aspirin and ibuprofen are very complex, occurring at multiple pharmacokinetic and pharmacodynamics levels.⁶¹ Due to the complexity of the mechanisms and their potential to lead to serious ADEs, we evaluated our framework’s usefulness for studying interactions between these 2 drugs. When we submitted a query to the D3 framework for the 2 drugs, it reported an asserted interaction from 5 different DDI databases, NDF-RT, Twosides, DrugBank, KEGG, and OSCAR, and then provided 7 inferred mechanisms of interaction, which are outlined in Table 4.

Table 4.

D3 inferred 7 mechanisms of interaction between aspirin and ibuprofen

Mechanism of interaction	Returned by D3
Pharmacokinetic (protein binding)	Albumin21
Pharmacodynamic (additive)	Cyclooxygenase inhibitors62
Pharmacokinetic (metabolic inhibition)	CYP2C9, PTGS2, PTGS163
Pharmacokinetic (metabolic induction)	CYP2C1964
Pharmacokinetic (transporter inhibition)	SLC22A665
MPIs	CYP2C9, ABCB1, SLC22A6, CYP2C19, CYP2C864,65
Pharmacogenetic	Rs2041766

The results shown in Table 4 demonstrate that the D3 framework can enumerate potentially relevant DDI mechanisms along with their provenances. Our work confirms that D3 can infer mechanisms of asserted DDIs and be applied with confidence to deduce potential interactions involving less-studied drugs.

Potential drug interactions via D3: a case study of irinotecan and levofloxacin

We examined possible interactions between irinotecan and levofloxacin, as both medications are used clinically, with significant changes in irinotecan and the concentrations of its metabolites being potentially lethal. Irinotecan, approved by the FDA in 1996, has been an effective chemotherapeutic medication for treating colon cancer.⁶⁷ Irinotecan induces apoptosis by inhibiting the topoisomerase I enzyme and, as a result, inhibiting DNA replication and transcription.⁶⁸ Unfortunately, patients treated with chemotherapy medications are at high risk for many different kinds of infections. According to clinical practice guidelines regarding cancer, levofloxacin has good evidence supporting its use for primary infection prevention in patients undergoing chemotherapy.⁶⁹ In light of this, irinotecan and levofloxacin were used as a query to the D3 framework, which returned 4 potential mechanisms of interaction: metabolic inhibition (irinotecan is a CYP3A4 enzyme substrate and inhibitor, and levofloxacin is a CYP3A4 enzyme inhibitor), transporter inhibition (irinotecan is a P-gp transporter substrate, and levofloxacin is a P-gp transporter inhibitor), MPIs (both drugs share the CYP3A4 enzyme and the P-gp transporter), and pharmacogenetic interactions (a specific SNP, rs1045642, that is associated with altered P-gp activity). To provide precise evidence, we queried the D3 knowledge base to find a drug that was already known to interact with irinotecan and had a similar pharmacokinetic profile to levofloxacin (an inhibitor of both the CYP3A4 enzyme and the P-gp transporter). D3 returned indinavir. From this, we hypothesized that there is a possible interaction between irinotecan and levofloxacin in 2 important pharmacological respects. The first is based on P-gp transport. Indinavir has been shown to interact with irinotecan, as it increases the level or effect of irinotecan.⁷⁰ Since levofloxacin and indinavir share the same transporter action (both are P-gp inhibitors),⁶² we hypothesized that levofloxacin and irinotecan may be interacting via their inhibition of P-gp. The second reason is based on the CYP3A4 enzyme. Levofloxacin is a CYP3A4 inhibitor, and CYP3A4 plays an important role in irinotecan metabolism.⁷¹ Therefore, when CYP3A4 inducers or inhibitors are administered, irinotecan concentrations may change significantly.

DISCUSSION

Significance

In this study, we presented a novel pharmacovigilance framework that uses Semantic Web technologies to infer potential mechanisms for DDIs. Our D3 framework analyzes pairs of drugs for pharmacokinetic, pharmacodynamic, pharmacogenetic, and multiple pathway interactions using 12 biomedical resources. We demonstrated that aggregation and integration of these biomedical resources into 1 consistent knowledge base offers the possibility of inferring potential interactions between drugs. Such functionality offers more informed explanations for potential interactions than what is typically offered in EHR systems, such as OSCAR, and enables clinicians and clinical researchers to monitor for potential interactions in their patient populations.

The D3 framework demonstrated reasonable recall rates (per Table 2) across the 15 DDI databases it encompasses, in terms of suggesting some explanation for their asserted interactions. The low level of overlap in terms of asserted DDIs between these sources coupled with the reasonable recall rates D3 achieves indicates that D3 is able to successfully integrate diverse knowledge bases to reliably provide at least some mechanistic explanation for a wide variety of asserted DDIs. D3 was also reasonably demonstrated to indicate noninteraction between pairs of drugs known not to interact, with a precision rate of 61% paired with a recall rate of 85% with an equal balance of asserted DDIs. Determining precision rate proved challenging, due to the lack of a substantial database of drug pairs known not to interact from which to draw knowledge and evaluation samples. We believe that variations in recall rates across the 15 DDI databases in terms of suggestion of some mechanism of interaction show specialization among sources regarding certain types of interactions. As the system operates on collective, generalized knowledge, DDI cases with relatively unique characteristics compared with the current collection of knowledge may be completely overlooked by the system.

This work differs from previous studies focused on target-based DDIs (eg, Drug Interactions Ontology, DIO),⁷² metabolic-based DDIs (eg, the Drug Interaction Knowledge Base, DIKB),³¹ or a limited set of mechanisms of interaction (eg, Drug-Drug Interactions Knowledge, DINTO)⁷³ in 3 important respects. First, D3 can be used to evaluate DDIs through consensus, as it integrates 15 available DDI databases. Second, D3 provides mechanistic explanations for DDIs using a multiresource, integrated knowledge base. Finally, D3 provides a wide range of coverage with regard to mechanism of interactions, with 9 different pathways for studying and inferring explanations of DDIs.

Limitations

The method we present in this study has several limitations that need to be addressed in future work. First, availability of drug and DDI information is constantly growing, which could improve our ability to identify putative DDI mechanisms. Second, while we examined 9 different mechanisms of interaction, other mechanisms – such as synergistic drug effects, occurring when the pharmacological effect of 1 drug increases the pharmacological efficacy of another, and antagonistic pharmacodynamic drug effects, occurring when the pharmacological effect of 1 drug reduces the pharmacological effect of another – have not yet been included in D3. Including a wider set of mechanisms may help to explain DDIs that have low corroboration in clinical sources such as OSCAR. Third, D3 is not able to predict the severity of interactions, since it does not take into account the dose-dependent nature of drug effects and their interactions. In fact, determining the severity of any interaction based on mechanistic information is challenging due to limited experimental data. However, incorporating pharmacometric models may help to provide more detailed insight into essential or patient-specific concentrations.^74,75 Fourth, D3 identifies a set of possible mechanisms that are not ranked by their contribution to the interaction. Mechanistic ranking may be of value in the context of drug development or clinical utility. Finally, while the D3 framework showed reasonable recall, it might still miss the correct mechanistic explanations. However, this framework will open the horizon for researchers to incorporate more details to come up with better predictions that can be used efficiently by clinical researchers.

CONCLUSION

A novel semantic pharmacovigilance framework was developed to deduce potential DDIs and their mechanisms at the pharmacokinetic, pharmacodynamic, pharmacogenetic, and multipathway interaction levels. The framework utilizes 9 inference rules for inferring mechanistic explanations of DDIs and deducing potential DDIs. We evaluated the recall rate of D3 for each of the 15 DDI databases it encompasses in terms of providing some mechanistic explanation for the DDIs asserted in the database. Additionally, we evaluated the recall rate of D3 for each of the 15 DDI databases it encompasses for each of the 9 inference rules in terms of providing mechanistic explanations. We applied the D3 framework to infer mechanistic explanations for the known DDI between aspirin and ibuprofen. Finally, we utilized D3 to deduce a potential interaction between irinotecan and levofloxacin. Overall, the D3 framework addresses the inherent difficulties in aggregating DDI-relevant information from disparate sources and automates inferences of explanations for DDIs for pharmacovigilance studies.

FUNDING

This project was founded by the Deanship of Scientific Research (DSR) at King Abdulaziz University, Jeddah, under grant no. G-685-611-38. The author, therefore, acknowledges DSR with thanks for technical and financial support.

COMPETING INTERESTS

None

SUPPLEMENTARY MATERIAL

Supplementary material is available at Journal of the American Medical Informatics Association online.