The potential of translational bioinformatics approaches for pharmacology research

Abstract

The field of bioinformatics has allowed the interpretation of massive amounts of biological data, ushering in the era of ‘omics’ to biomedical research. Its potential impact on pharmacology research is enormous and it has shown some emerging successes. A full realization of this potential, however, requires standardized data annotation for large health record databases and molecular data resources. Improved standardization will further stimulate the development of system pharmacology models, using translational bioinformatics methods. This new translational bioinformatics paradigm is highly complementary to current pharmacological research fields, such as personalized medicine, pharmacoepidemiology and drug discovery. In this review, I illustrate the application of transformational bioinformatics to research in numerous pharmacology subdisciplines.

Introduction

Bioinformatics is a vibrant research field that has now demonstrated its value in unravelling the novel genomic 1 and proteomic 2 underpinnings of numerous complex diseases. The success of bioinformatics rides with the present day wave of ‘big omics’ datasets associated with rapidly evolving biotechnology advancements 3,4. The fundamental driving force of bioinformatics is its powerful nature of discovery. While this discovery approach is not new to clinical pharmacology research, both pharmacogenomics and pharmacoepidemiology research (i.e. pharmacovigilance) 5 are now being increasingly driven by bioinformatics-associated discovery. Moreover, accessibility to enormous data sources presents tremendous opportunities for translational bioinformatics research in pharmacology. Unlike previous review papers, which emphasized either bioinformatics resources 6,7 or computational methodologies 8, here I will focus on emerging applications of bioinformatics and their impact on clinical pharmacology research.

Prospects

Knowledge representation, a fundamental step in connecting big data sources

Clinical pharmacology research-associated big data sources include observational (e.g. electronic medical/health records, the FDA adverse event reporting system 9, etc.) and drug knowledge (e.g. DrugBank 10, PubMed 11, PharmGKB 12, LINCS 13) bases and databases. Extensive details of these data sources are well described in Russ Altman’s review paper 7. Using big data, novel pharmacology research starts with communication between these databases, which in turn depends on standardized data annotation schemes. For example, the observational medical outcomes partnership (OMOP) recently attempted to integrate multiple observational databases through a common data model (CDM) 14 by normalizing the dictionaries of medications, diagnoses and laboratory tests across all its (OMOP) databases. The OMOP has also provided an annotation scheme to characterize the time-dependent nature of medications, laboratory tests and diagnoses. Furthermore, based on the CDM, OMOP has further standardized the implementation and reporting of drug/adverse drug event (ADE) association analyses.

Although ADEs have been standardized through MedDRA (a medical terminology database for adverse drug events, www.meddra.org), whose utility in drug labelling has been well accepted, MedDRA still lacks sufficient annotation schemes to link an ADE with its molecular mechanism(s) 15. Within the informatics research domain, the formulation of a terminology dictionary is now evolving towards the construction of an ontology. Ontology deals with questions concerning what entities exist or can be said to exist, how such entities can be grouped or related within a hierarchy, and how they can be subdivided according to their similarities and differences. In pharmacology research, a properly constructed ontology will facilitate the proper presentation of drugs and ADE data, either from large clinical databases or from the literature. Recently, the Ontology of Adverse Events (OAE, a biomedical ontology in the domain of adverse events) by He et al. 16 further strengthened the link between ADEs and anatomic regions, hence adding the capability of integrating molecular data into ADEs. In addition, the OAE recently defined causal adverse events (i.e. drug-induced ADEs), representing a much-needed improvement over MedDRA.

While pharmacokinetic (PK) and pharmacodynamic (PD) studies and models are also essential components of pharmacology research, the interactions between PK/PD and observational studies remain limited. Recent work, however, on PK and drug–drug interaction (DDI) ontology 17 has laid a solid foundation for annotating in vitro, in vivo and clinical terminologies for PK and DDI preclinical and clinical studies. In particular, these ontology advancements have defined PK models and parameters, in vitro PK experimental design and results, clinical PK study design and results, and established criteria for defining DDIs. Thus, ontology is an essential step in connecting the OAE to the OMOP, and consequently, connecting PK to ADEs.

All of the above-mentioned data and database standardization efforts are categorized into a field known as knowledge representation, a fundamental system used to connect large data sources. More extensive discussion of biomedical ontology development can be found in a review by Musen et al. 18. All these newly developed pharmacology-related ontologies are publically available for knowledge representation. The primary challenge or roadblock to their utility to the broader pharmacology research community is proper training. The lack of proper tools for researchers, who have limited informatics knowledge to browse these ontologies, is the biggest gap in training.

Data mining of high dimensional drug–ADE associations

Traditional pharmacovigilance studies have focused on associating single drugs with single ADEs 5. However, more recent successful studies have significantly expanded the dimension of associations. For example, Duke et al. investigated drug interactions, using a local version of the OMOP database at Indiana University 19, to identify successfully six novel drug interaction pairs that significantly increased myopathy risk above a mere additive risk from two single drugs. A notable example was a newly identified interaction between loratadine and simvastatin. In this drug interaction study, myopathy cases and controls were matched by the index time of the cases. This data mining analysis adjusted for covariates such as age, gender and the number of co-medications in a logistic regression.

In another example of multiple drug–ADE discovery, Tatonetti et al. further expanded association analysis between drugs or drug interactions and adverse events to assess all the FDA approved drugs and ADEs 20. Using the FDA’s Adverse Event Reporting System (FAERS) as a training set and Stanford’s medical records as the validation set, they identified 47 associations of drugs and drug interaction effects. Notable examples included interactions between antidepressants and thiazide diuretics that resulted in increased incidence of prolonged QT. Additionally, Tatonetti et al.’s data mining algorithm adapted a propensity score approach to adjust for confounding from all co-medications 20. This paper illustrates the tremendous value of data mining for correcting many false positive associations.

To detect associations between any combinations of drugs and any combinations of adverse events, an association rule mining was implemented using the FAERS database. This association rule refers to the frequency of a drug–ADE association in a health record database. It claimed that 67% of associations were clinically validated by domain experts 21. Computational efficiency of association rule mining was recently further improved by Xiang and Li et al. 22.

The combination of knowledge representation and data mining provides an extremely powerful translational bioinformatics approach for investigating drug–ADE associations. This new approach tremendously scales up association analyses between all FDA approved drugs and all ADEs, and is more discovery driven. It is also complementary to traditional pharmaco-epidemiology research into drug–ADE associations, which rely on stringent epidemiological study design and sensitivity analyses. Similar to the previously described challenges to knowledge representation, training and proper computational tools remain underdeveloped.

The value of text mining in clinical pharmacology research

The contributions of text mining to clinical pharmacology research arise mainly from two sources, medical records and research literature. The definitions of diseases or ADE phenotypes using medical records require specific text mining tools. Wang et al. demonstrated the feasibility of investigating drug-induced ADEs by a natural language processing (NLP) algorithm that could identify and extract ADEs from clinical notes with moderate recall and precision 23. Additionally, using known ADEs from seven drugs, they demonstrated the feasibility of establishing ADE pharmacovigilance from their NLP algorithm. Another example, by Xu et al., is to develop NLP algorithms to extract medication data (i.e. drug names and dose information) from clinical notes, with significant high performance 24. Using this NLP system, they successfully reproduced the pharmacogenetic effects of VKORC1 and CYP2C9 on warfarin weekly dosing.

Another application of text mining is literature extraction of gene–drug and drug–drug interactions with a two-fold research goal: (1) curating gene–drug and drug–drug interaction data from the literature and (2) discovering novel gene–drug and drug–drug interactions. To that end, Garten & Altman developed a text mining approach that could effectively extract pharmacogenomics relationships from full text articles 25. More recently, they further expanded gene–drug interaction text mining into drug-drug interaction text mining 26. To develop ‘gold standards’ and true positive training sets for drug interaction text mining, Li et al. 17 and a Spanish group 27 independently set up a corpora of drug interaction abstracts and sentences from both PubMed and FDA drug labels. Li’s group further developed an ontology of pharmacokinetics and drug interactions, successfully differentiating clinical from in vitro pharmacokinetic drug interaction studies. Using the machine learning algorithms developed from this corpus, Duke et al. mined all the published in vitro DDI studies on the enzyme cytochrome P450 to predict 13 193 novel drug interaction pairs 19.

As diverse disciplines and varied studies are involved 6, pharmacology evidence is often not available across in vitro experiments, clinical pharmacokinetic and clinical pharmacodynamic studies, which creates knowledge gaps. These knowledge gaps are the primary driving force for translational pharmacology research. Text mining, if successfully developed, will be a tremendously powerful approach to identify knowledge gaps for all currently used drugs. As text mining has emerged only recently in pharmacology research, it is limited by its own underdeveloped methodological research. Its theory and application tools need to mature before its further transition to general usage in pharmacology research.

Integrating molecular data with health data

The eMERGE (Electronic Medical Records and GEnomics, emerge.mc.vanderbilt.edu) network is the first national consortium developed to investigate and catalogue genomic effects on disease/drug responses 28, as defined using observational databases coupled with biobanks. In a feasibility study, Ritchie et al. investigated associations between five disease phenotypes (atrial fibrillation, Crohn’s disease, multiple sclerosis, rheumatoid arthritis and type 2 diabetes) and various genetic variants 29. These disease phenotypes were defined by a mixture of coded variables, and their features were extracted from clinical notes using NLP algorithms. Previous genotype–phenotype associations were replicated (P < 0.05) in 8/14 cases, at an odds ratio of > 1.25 29. More recently, Denny et al. extended a biobank-based genome-wide association study (GWAS) to 86 reported disease phenotypes in the National Human Genome Research Institute GWAS catalogue, successfully replicating 66% of prior GWAS associations 30.

Using a chemo-informatics approach, Pouliot et al. investigated whether drug-screening data from the publically available PubChem BioAssay database could accurately predict ADEs 31. Subsequently, in mining the FAERS, they successfully identified nine ADEs, grouped by system organ classes, based on information from the PubChem Bioessay database. Schelleman et al. further demonstrated drug interactions between sulfonylureas and fribrates/statins 32, finding that fenofibrate combined with gemfibrozil increased glyburide-induced hypoglycaemia, based on Medicaid and Medicare claimant data. However, follow-up in vitro experiments predicted that fenofibrate and gemfibrozil only moderately increased glyburide drug exposure, suggesting that CYP inhibition was unlikely to cause this particular adverse drug interaction. In other work, Peng et al. investigated multiple drug targets, using structure-based protein-ligand interactome prediction 33. By scoring the interactions of 1000 FDA-approved drugs docked to 2500 human protein pockets, they predicted even more interactions among them, especially drug ‘off-targets’. Using additional compound physicochemical properties such as flexibility, hydrophobicity and size, they predicted and ranked drugs having similar predicted off-target profiles. In particular, the therapeutic effects of losartan and ergotamine on lung cancer were further evaluated and validated using an electronic medical records database at Indiana University, demonstrating the feasibility of revealing new drugs for lung cancer 33.

The integration of molecular data and health records data is well accepted, and has become widely used in current pharmacology research. The previously described knowledge representation and data mining approaches will bring about the discovery of new genetic risk factors for drug effects. On the other hand, it also currently suffers from the same challenges in training and availability of proper tools.

Network-based system pharmacology

Network-based system pharmacology approaches are becoming a new paradigm for drug discovery, as described by Csermely et al. in a comprehensive review of model databases, visualization tools and computational methods 34. Here, I highlight a few significant recent papers on this topic. Besnard et al. developed a computational scheme for designing drugs against profiles of multiple possible target proteins, demonstrating that an approved acetylcholinesterase inhibitor could evolve into brain-penetrable ligands with selective profiles for G-protein-coupled receptors 35. Moreover, among 800 predicted ligand-target predictions, 75% were confirmed experimentally 35.

Using another network approach, Li et al. showed strong connections between gene–gene relatedness and specific disease phenotypes, suggesting a method for discovering disease-associated genes 36. Interestingly, using this network approach, this group further investigated the active compounds of a Chinese herb for treatment of type II diabetes 37. Almost in parallel, Gottlieb et al. used a similar network approach to decipher drug–drug similarities and drug–disease associations, for eventual drug repurposing and drug response prediction 38,39. Most recently, Gottlieb & Altman constructed molecular pharmacology networks for integrative linkage of drug target genes, disease genes and pharmacogenetic genes 40.

Network-based system pharmacology is still in its infancy. Although its impact on pharmacology research is currently limited, its research potential for pharmacodynamic effects and modelling is immense. The current challenge to network-based system pharmacology is that its theory and methodology have not yet met the needs of pharmacology research. In other words, current systems pharmacology models have not well-predicted either in vitro or clinical drug responses.

Emergence of big data sources

Social media is emerging as a valuable data source for assessing drug safety. Early work by Benton et al. showed that it was feasible to recapture hormone therapy ADEs using breast cancer medical ‘message boards’ 41. Subsequently, Mao et al. further demonstrated that discontinuation of hormone therapies could also be captured by medical message boards 42. Concurrently, White et al. used online health-seeking searches to investigate drug interaction signals, successfully identifying a synergistic hyperglycaemia ADE, based on joint searches for paroxetine and pravastatin 43. Recently, this line of research was further evaluated by the FDA 44.

Another emerging resource is cell-based perturbation data sources, with one salient example being LINCS (Library of Integrated Cellular Signatures), an NIH-supported research network that has thus far investigated cellular responses (‘signatures’) of nearly 100 cell lines perturbed by 20 000 compounds and 6000 siRNAs 13. Cellular response signatures are measured using various omics platforms. Such data sources will be extremely valuable to investigate the molecular mechanisms of pharmacodynamic effects.

Although the immense impact of these data resources is on the horizon, the immediate challenge is the lack of user-friendly interfaces, such that a biomedical research investigator with limited informatics skills can browse these data sources.

Conclusions

Translational bioinformatics has shown great potential for contributing to two aspects of pharmacology research. First, based on its nature of large scale discovery, it expedites the translational process of pharmacology research. For example, using traditional pharmacology research approaches, it took 8 years of progression from drug metabolism and in vitro interaction studies to clinical pharmacology studies to understand tamoxifen pharmacogenetics and its interaction with selective serotonin receptor inhibitors 45. By contrast, using translational bioinformatic methods, it took our group only 3 years to establish myopathy-associated drug interaction epidemiology data, combined with molecular mechanisms, by mining numerous drug interaction signals 19. As text mining and data mining methodologies mature, we believe this process will become even more efficient. Secondly, translational bioinformatics is riding the wave of ‘big data’, and can investigate the same pharmacology problem from many different angles, using different data sources. Our myopathy-associated DDI study is again one such example 19, in which we mined medical records to obtain evidence of pharmacoepidemiological DDIs, while also mining the literature to examine DDI molecular mechanisms. Thus, the strength of translational bioinformatics approaches will only increase, as more and more data sources become more accessible.

The emergence of translational bioinformatic approaches will not replace traditional pharmacology research. As a matter of fact, while translational bioinformatics is more discovery driven, it will be highly complimentary to traditional pharmacology research approaches, which are usually hypothesis-driven. For example, data mining medical records for novel drug–drug adverse interactions still need well-designed pharmacoepidemiology studies to validate causal relationships and/or traditional pharmacology experiments, to determine molecular mechanisms.

The primary challenge of incorporating translational bioinformatics into clinical pharmacology research is training. Pharmacology research by itself has already faced its own challenges in integrating both clinical and basic sciences, and translational bioinformatics adds yet another dimension to this field. Unlike the PK and PD modelling skills required for pharmacometrics, translational bioinformatics requires much broader skills in data mining, text mining, knowledge representation and omics. These requirements call for broader collaboration among different academic units, such as medicine, pharmacology and informatics. Joint training programmes will be absolutely necessary to prepare young scientists to enter this field. We believe, however, that these challenges will be met with innovative solutions, similar to our long history of successful adoption of emerging scientific pursuits.

Proper bioinformatics training for pharmacology research investigators requires user-friendly bioinformatics tools, which will allow them to mine large medical databases and literature text, build network-based systems models to predict drug responses and explore large databases. As many of these translational bioinformatics areas remain in development and many tools are not necessarily available at the moment, an alternative strategy is to seek synergistic collaborations with informaticians. As the research goals and approaches between informaticians and those of traditional clinical pharmacologists and molecular pharmacologists are rather complementary, I fully expect successful synergistic collaborations.

Competing Interests

The author has completed the Unified Competing Interest form at www.icmje.org/coi_disclosure.pdf and declares no support from any organization for the submitted work [LL], no financial relationships with any organizations that might have an interest in the submitted work in the previous 3 years [LL] and no other relationships or activities that could appear to have influenced the submitted work [LL].

This work was supported by DK102694, GM10448301 and LM011945.