Potential Role of Pharmacogenomics in Reducing Adverse Drug Reactions:  A Systematic Review FREE

Several highly publicized reports and policy initiatives have urged greater efforts to reduce the rate of adverse events in medical care.^1- 4 Pharmaceutical agents are one of the most commonly identified causes of adverse events, resulting in significant patient morbidity, mortality, and excess medical care costs.^2,5 A widely cited meta-analysis estimated that more than 2 million hospitalized patients have severe adverse drug reactions (ADRs) annually in the United States even when drugs are appropriately prescribed and administered, and that ADRs ranked between the fourth and sixth leading cause of death in the United States in 1994.⁶ However, there have not been any updated, systematic reviews published since that time.

One possible cause of ADRs is genetic variation in how individuals metabolize drugs. The Human Genome Project heralds new opportunities for using genetic information to individualize drug therapy, called pharmacogenomics.⁷ In fact, pharmacogenomics may be one of the most immediate clinical applications of the Human Genome Project⁸ and may become part of standard practice for "quite a number of disorders and drugs by year 2020."⁹

A primary benefit of pharmacogenomics that has been repeatedly cited in prominent articles is the potential to reduce ADRs.^10- 15 Some ADRs caused by genetic variation—previously considered nonpreventable—may now be preventable. Adverse drug reactions could be reduced by modifying drug selection or dosing in patients with poor ability to metabolize a drug because of genetic variation in their drug metabolizing enzymes or by developing drugs a priori that will avoid metabolic pathways with adverse genetic variability.

Despite the commonly accepted notion that pharmacogenomics will reduce ADRs, there have not been any systematic and quantitative evaluations of the potential role of genetic variability in ADRs. Typically, studies have addressed either the nongenetic causes of adverse events such as human error^2,5 or specific genetic variants associated with drug metabolizing enzymes without linking this to the ADR literature.^14,16- 17 Although several studies have found a direct link between specific genetic variants and ADRs,^18- 19 these are single studies that have not been systematically combined.

The purpose of this study was to evaluate the potential role of pharmacogenomics in reducing the incidence of ADRs and to discuss the clinical and policy implications. Specifically, we conducted 2 systematic literature reviews: one for studies reporting ADRs and the other for studies reporting variant alleles of drug-metabolizing enzymes. The results of the 2 reviews were then "linked" via the enzymes responsible for metabolizing each of the drugs to examine the possible contribution of genetic variability to ADRs.

This is the first study to our knowledge to systematically identify which specific drugs are linked to ADRs and the genetic variability in drug-metabolizing enzymes relevant to those drugs and to evaluate these findings in a clinical and policy context.

We restricted the scope of our analysis as shown in Figure 1. Within the adverse drug event literature, we focused on ADRs. We used the common definition of ADRs as any noxious, unintended, and undesired effect of a drug that occurs at doses used in humans for prophylaxis, diagnosis, or therapy, excluding therapeutic failures, intentional overdose, errors in drug administration, and noncompliance.⁶ Adverse drug reactions are caused by inherent properties of drugs (and are thus often called nonpreventable); therefore, they have the most relevance to our study. In contrast, adverse drug events (ADEs) include preventable events, such as human errors. We focused on drug-metabolizing enzymes, rather than receptors or transporters, since drug-metabolizing enzymes are the predominantly known cause of genetic variability in drug response. We examined both phase 1 and phase 2 drug-metabolizing enzymes. Phase 1 drug-metabolizing enzymes are composed mainly of P450 cytochromes (CYPs) that oxidize drugs while phase 2 enzymes conjugate drugs for subsequent excretion. Genetic variants of phase 1 drug-metabolizing enzymes have been extensively studied whereas existing data on phase 2 drug-metabolizing enzymes are less comprehensive.

Within the genetics and genomics literature, we focused on pharmacogenomics. Although both pharmacogenetics and the broader field of pharmacogenomics address genetic factors responsible for variation in drug response among patients, we use the term pharmacogenomics in this article. For each gene encoding a drug-metabolizing enzyme, variant alleles may exist (called polymorphisms when they occur in more than 1% of the population). Our primary analysis focuses only on drug-metabolizing enzymes with known variant alleles that cause poor metabolism because these are most relevant to ADRs.

Lastly, we linked our 2 literature reviews. Thus, the scope of our study is limited to drug-metabolizing enzymes with known variant alleles causing poor metabolism of drugs frequently identified in ADR studies.

Below we discuss the literature review of ADRs, the literature review of variant alleles and how these were linked. Each review is treated as a separate study due to its complex nature, and details are provided in an online appendix (available in PDF format:

Grahic Jump Location

We conducted pilot studies to assess the study's feasibility and to refine our methods (details are available on request). All inclusion and coding decisions were documented in standardized spreadsheets. Two reviewers (E.O., J.L.) conducted the majority of coding, while others (K.P., D.V., W.S.) reviewed the results and reconciled differences.

We synthesized the literature by applying the same systematic approaches used in meta-analysis.²⁰ However, we did not estimate summary effect sizes. As several studies have noted, it is difficult to estimate accurately the true incidence of ADRs—and even more difficult to estimate incidence based on specific drugs—because of heterogeneity among studies.^1- 2,21 It is also currently not feasible to estimate a summary measure of attributable impact.

Literature Search. Previous systematic ADR reviews have used keyword searches,⁶ but we found in pilot studies that a keyword search would have low sensitivity and specificity. We thus developed a specific MEDLINE search strategy using Medical Subject Headings (MeSH) terms, using 19 articles previously identified as key articles^2,6 (available on request). We included studies from the past 5 years (January 1995-June 2000) to extend a previous meta-analysis.⁶ Our search strategy, resulting in 333 hits, was major exact subject drug therapy-adverse effects or pharmaceutical preparations-adverse effects or medication errors-statistics and numerical data or iatrogenic disease-epidemiology and language English and exact subject human and publication type journal article and not review.

Article Selection. Inclusion criteria were studies that

We excluded studies reporting only on ADEs, hypothesized ADRs, clinical trials of specific drugs, reviews, case reports, and articles with redundant data. We tested the impact of our inclusion criteria on our results by conducting sensitivity analyses based on key study characteristics (online appendix [available in PDF format:

Grahic Jump Location

We excluded 206 studies based on the abstract for a total of 127 potentially eligible studies. Review of full articles' text resulted in a final total of 18 studies.^22- 39 In each screening phase, about one third of excluded studies were conducted outside the United States, one third did not include ADRs, and one fourth were not primary data.

Article Coding. We coded each study for the specific drugs identified. We used the Drug Information Handbook⁴⁰ to add class and therapeutic category information. We included drug-drug interactions in our database because of evidence that such interactions may be due to genetic variability. Since the classification of drug-drug interactions as avoidable ADEs or unavoidable ADRs was inconsistent in the literature, we could not delineate those that might be true ADEs (and thus not caused by genetic variability). When a combination of specific drugs was named as causing an ADR, we coded each drug separately but counted it as a single observation.

Literature Search. We used the same approach in developing a search strategy as for the ADR articles, using 8 well-known articles as key studies (available on request). The period January 1997-August 2000 was chosen because articles from 1997 onward should capture recently identified variant alleles. We only included review articles because they were most relevant to our study and because of redundancy among the large numbers of original reports. We supplemented the review articles with data from an extensive Web site (http://www.imm.ki.se/CYPalleles).

Our search strategy was major exact subject cytochrome P-450—genetics or Pharmaceutical Preparations—metabolism or sulfotransferases—genetics or glutathione transferase—genetics or methyltransferases—genetics or glucuronosyltransferase—genetics or epoxide hydrolases—genetics or arylamine N-acetyltransferase—genetics and exact subject polymorphism genetics and language English and publication type review.

Article Selection. Articles that reviewed variant alleles of drug-metabolizing enzymes and their effect on drug metabolism were included. We excluded articles on gene-environment interaction (ecogenetics or disease risk), methods or techniques for genotyping, and editorials (39 studies were excluded; available on request). Based on reviews of article texts, 22 articles were included.^{13,17,41- 59} Fifty-five percent of the exclusions were because the articles were not review articles.

Web Sources. The literature on variant alleles is rapidly growing and thus publications are often outdated and difficult to summarize because of differences in nomenclature. Therefore, we combined the literature review with data from the Human Cytochrome P450 (CYP) Allele Nomenclature Committee⁶⁰ Web site (available at: http://www.imm.ki.se/CYPalleles, accessed January 2001). This Web site provides more valid data than a typical single site because it is a synthesis of prior work and is authored by an international committee. It is thought to be used by the majority of researchers in this area (M. Ingelman-Sundberg, written communication, January 17, 2001).

Article and Web site Coding. We coded the following information:

Because prevalence data come from multiple and varying sources, we report ranges rather than point estimates.

Individuals possess 2 alleles for each gene encoding a drug-metabolizing enzymes and variations may occur in neither, one, or both alleles. Patients who are poor metabolizers are usually homozygous carriers of 2 nonfunctional alleles whereas heterozygotes are more frequent but often show only moderate impairment of metabolism. Some studies report data on the percentage of individuals who are poor metabolizers while others report the frequency of the variant alleles themselves in a population. We abstracted both types of data from the studies whenever possible.

Data on ADR-associated drugs and variant alleles were linked via the relevant enzymes (Figure 1). We expected that drugs identified in ADR studies would be more likely to be metabolized by enzymes with evidence of genetic variability than drugs not identified in ADR studies. We constructed 2 comparison groups to assess how likely these results occurred by chance by using a random sample of 27 drugs from all drugs sold in the United States⁶¹ and by using a random sample of 27 drugs from the top 200-selling drugs in the United States.⁶² In both samples, we excluded drugs found in the ADR literature review.

We also conducted 2 sensitivity analyses. First, to examine the implications of restricting our sample to the most cited drugs, we compared our results with those reported in a recent review of the ADE literature.² Second, because our primary results are based on the top 27 drugs causing ADRs drawn from a list of 131 drugs, we conducted a reliability test by examining whether we obtained similar results using a random sample of 27 drugs drawn from the same list of drugs cited in ADR studies.

To identify drug-metabolizing enzymes involved in the metabolism of drugs cited in the ADR studies, we used multiple standard drug information sources^{40,61,63- 64} as well as an online source (http://gentest.com) and drug interaction database.⁶⁵ We also conducted MEDLINE searches for each included enzyme using MeSH, the name of the drug, and the word metabolism.

It is important to note that our coding was based on data as reported in the sources reviewed. Therefore, our estimates of the involvement of genetic factors in ADRs are probably conservative because new studies are rapidly finding new genetic variants. For example, enzymes CYP3A4 and CYP3A5 were not identified in the review articles as having nonfunctional variant alleles although recent studies have identified such alleles.^66- 67

The majority of the ADR studies had patient sample sizes of less than 1000 and were based on hospital data, used nonprospective study designs, and included both ADRs and ADEs (online appendix [available in PDF format:

Grahic Jump Location

Grahic Jump Location

About half of these articles were published recently (1999-2000) and focused only on phase 1 P450 enzymes (online appendix [available in PDF format:

Grahic Jump Location

Table 2 lists the enzymes, specific variant alleles, and prevalence of poor metabolizers and variant alleles relevant to the drugs we found in the ADR studies. This table only includes enzymes for ADR-associated drugs and it is limited to variant alleles known to cause poor metabolism (Table 3; online appendix [available in PDF format:

Grahic Jump Location

The prevalence of poor metabolizers and variant alleles shows substantial variability across and within enzymes (Table 2). For example, the NAT2 enzyme shows wide variability of poor metabolizers across racial and ethnic groups (Japanese, 8%-10%; white, 50%-59%; Egyptian, 92%), and some Asian groups show higher prevalence of poor metabolism and/or variant alleles in the CYP2C18, CYP2C19, and CYP2D6 families. Therefore, these groups may be more susceptible to ADRs from drugs metabolized by those enzymes. However, there are many gaps in the available prevalence data.

Linking Drugs Identified in ADR Studies to Variant Alleles. We found that 59% (16/27) of the drugs cited in the ADR studies are metabolized by at least 1 enzyme with a variant allele known to cause poor metabolism. Conversely, only 22% of randomly selected drugs sold in the United States (P = .006, z test) and 7% of randomly selected top-selling US drugs are metabolized by enzymes with this genetic variability (P<.001, z test). Although our study design does not allow for causal inferences, these analyses support the hypothesis that drugs identified in ADR studies would be more likely to be metabolized by enzymes with genetic variability than drugs not identified in ADR studies.

The results for enzymes CYP1A2 and CYP2D6 are particularly interesting (Table 2 and Table 3). The CYP1A2 enzyme has only 1 identified variant allele with poor metabolism, but there is a significant prevalence of poor metabolizers of CYP1A2 substrates among whites. Whereas CYP1A2 is estimated to be the major metabolic pathway for only 5% of all prescribed drugs,⁶⁸ it is involved (at least partly) in metabolizing 75% of the ADR drugs associated with variant alleles. Although CYP1A2 is only a minor enzyme for many of these drugs, these results indicate that CYP1A2 might play a more important role in ADRs than previously identified and therefore an area into which further research may be useful. On the other hand, CYP2D6 is estimated to be the major metabolic pathway for 25% of all prescribed drugs⁶⁸ and is widely suspected of causing ADRs because it has a multitude of known variant alleles. However, CYP2D6 has a slightly lower prevalence of poor metabolizers (3%-10% whites) than CYP1A2. Moreover, we found CYP2D6 to be involved in metabolizing 38% of the relevant ADR drugs. Although this incidence is greater than among randomly selected drugs, it is less than what is observed for CYP1A2. These results may reflect an increasing awareness of CYP2D6 variants as a complicating factor in drug therapy and the selection of non-2D6 drugs if severe adverse effects are likely.⁶⁹

First, our sample appears to be generally representative of reported ADRs. The top 4 most frequently occurring drug categories in our study, accounting for 61% of observations, were cardiovascular, antibiotics, psychiatric, and analgesic. These are the also the most frequent categories reported in the General Accounting Office review.² (Note that there was no available comparison of specific drugs rather than drug categories.)

Second, our results appear to be reliable; that is, we would have obtained the same results if we had chosen another group of drugs from the drugs found in our ADR literature review. Using a sample of randomly selected drugs cited in the ADR studies, we found that 44% are metabolized by at least 1 enzyme with a variant allele known to cause poor metabolism—a proportion that is not significantly different from our results using the most frequently cited drugs in the ADR studies (95% confidence interval [CI], 27%-61% using a finite population correction to adjust the sample variance). These findings suggest that our results are not biased by focusing only on the 27 most frequently cited drugs.

We found that more than half of the drugs cited in ADR studies are metabolized by at least 1 enzyme with a variant allele known to cause poor metabolism. These results suggest that genetic variability in drug metabolizing enzymes is likely to be an important contributor to the incidence of ADRs.

Many recent articles have noted the potential for a decrease in ADRs through the use of pharmacogenomics, but our study differs in several ways from previous work. First, we used a systematic approach to identify drugs associated with ADRs. Second, we conducted a systematic review of variant alleles of drug-metabolizing enzymes. Third, we linked these data sets to assess the potential contribution of genetic variability to ADRs. However, the link between ADRs and genetic variability is complex, and our findings do not imply a causal relationship or that ADR incidence would necessarily be reduced if drug selection and dosing were based on genetic variability. Our findings do indicate, however, that the converse hypothesis—there is no relationship between ADRs and genetic variants—is probably not true, and our findings suggest in what area future research may have the greatest pay-off.

The application of pharmacogenomics information has great potential but also faces substantial challenges. We summarize herein some of the key issues from a clinical, industrial, and societal perspective. We then provide a clinician's checklist and discuss criteria that can be used to evaluate the potential impact of pharmacogenomics information in reducing ADRs.^70- 71

From a clinical perspective, there are obvious potential benefits to individualized drug therapy although there are many issues that must first be addressed.^69,72- 73 Currently, clinical applications, although beginning to emerge, lag behind the available technology. For example, a clinical test for the CYP2D6 enzyme has been developed and may be available in physicians' offices within the next 2 years.⁷⁴ However, currently, genotyping for enzymes is used only in a limited number of primarily academic centers, eg, genotyping for the CYP2D6 enzyme to aid individual dose selection for drugs to treat psychiatric illness,¹⁰ genotyping for thiopurine methyltransferase (TPMT) for treatments for childhood leukemia,⁷⁵ and HER-2 receptor expression levels for herceptin breast cancer therapy.⁷⁶ A relatively more common use of genotyping is for viruses instead of individuals, eg, genotyping of the human immunodeficiency virus.^70,77 Others have noted particularly important uses of genotyping for the near future, for treatment for such diseases as Alzheimer disease, atherosclerosis, and cerebral-vein thrombosis among women who have taken oral contraceptives.⁷

In general, however, clinical practice is not yet prepared to benefit from the genetics revolution. Most health care professionals have not had "even one hour" of instruction in pharmacogenomics as part of their formal training.⁷⁸ Therefore, moving pharmacogenomics from bench to bedside will be a challenging undertaking.

From an industry perspective, there will be both pros and cons to developing drugs a priori that will avoid ADRs. On the one hand, drug development may become more efficient and less costly as a result of the use of genetic information.^69,79- 80 Pharmacogenomics may also allow companies to resurrect drugs that have failed clinical trials by using genetic information to target them to a smaller group than the population. However, small-target populations may also decrease incentives for companies to develop new drugs and present other challenges, such as recruitment into trials.

From a societal perspective, we have only begun to examine the social, economic, and ethical aspects of pharmacogenomics.^70,81- 84 Research to date has focused more on the use of genetic information to predict future risk, such as privacy issues involved in testing for the BRCA gene to determine risk for breast cancer,⁷ than on individualized drug therapy. Only a few studies have begun to address the economic impact of introducing genotyping as a guide for developing individual therapy⁸⁵: whether it should be provided and to whom, how much it would cost,⁴⁴ and whether insurers would cover it.⁸⁶ It will thus be critical to conduct cost-effectiveness analyses and other evaluations of the clinical, economic, and societal effect of pharmacogenomics. For example, race is correlated with many gene patterns and therefore genotyping raises issues about stereotyping and preferential treatment.^8,73 As stated by one observer, "What happens when the patient comes in and says, ‘I hear there's a great new drug for asthma,' and the doctor says, ‘Yeah, but it's only for whites?'"⁸

Even when metabolism of a drug is found to have genetic variability, the ultimate question is: "Does it matter?" It is currently difficult to estimate the attributable impact of genetic variability on ADRs because this would require complex data on many factors that are unknown. However, we have outlined criteria that can be used to evaluate the potential impact of pharmacogenomics information in reducing ADRs from a clinical and societal perspective. These criteria can guide future research, assist clinicians in considering these issues, and serve as starting points for more comprehensive analyses of the cost-effectiveness and cost-benefits of pharmacogenomic information.

The potential effect of pharmacogenomics information from a societal perspective will be a function of medical need, clinical utility, and ease of use (BOX 1). Medical need will be driven primarily by the prevalence of variant alleles in a population, the use of a drug in that population, the severity of the ADR, and the ability to monitor drug toxicity using current technologies. Genetic testing will be clinically appropriate only if there is sufficient evidence to link variant alleles with valid surrogate markers of drug toxicity or patient outcomes. And finally, genetic tests must be easy to use, and clinicians must be able to interpret them in order to gain widespread acceptance.

MEDICAL NEED
Prevalence of ADRs Caused by Drug
The incidence of ADRs and the use of a drug is high enough to warrant use of genetic information.

Prevalence of Poor Metabolizers and Variant Alleles
The prevalence of poor metabolizers and/or variant alleles is high enough to warrant use of genetic information. Depending on the clinical consequences of genetic variation, even a low prevalence may warrant genetically based interventions.

Outcome
The consequences of associated ADRs are severe enough to produce significant changes in clinical or quality of life end points, or lead to significant economic costs.

Monitoring
Current methods for monitoring therapeutic response or evaluating toxic effects are unavailable or inefficient.

CLINICAL UTILITY
Association
Sufficient evidence exists to link the variant allele to clinical response to a drug, and ultimately, patient outcomes (gene penetrance). Moreover, the genotyping assay is predictive for a substantial portion of the patient population, taking into account the most prevalent variant alleles contributing to the disease.

EASE OF USE
Assay
An assay that can rapidly, relatively inexpensively, and reliably detect the variant allele is available.

Clinicians
Clinicians are able to interpret the results and appropriately use the information.
Data are based on Spear,⁸⁷ Phillips et al,⁷⁰ and Veenstra et al.⁷¹

Although much of the information needed to apply pharmacogenomics information to clinical practice is currently unknown, our results do suggest several steps that clinicians can take when prescribing a drug with a high incidence of ADRs (BOX 2).

Check whether the drug is known to be metabolized by a polymorphic drug-metabolizing enzyme. Pay special attention to the prevalence of polymorphic alleles of the relevant drug-metabolizing enzyme in the patient population being treated since prevalence varies considerably among groups.

If genetic variability may be a significant problem:

We highlight the criteria in BOX 1 and BOX 2 by applying them to warfarin sodium, a drug commonly identified in ADR studies. Warfarin is an anticoagulant used in the prophylaxis and treatment of thromboembolic disorders and is metabolized primarily by the CYP2C9 enzyme. Individuals who are deficient in CYP2C9 enzyme activity may require a lower warfarin dose or more frequent monitoring and may be at higher risk for bleeding episodes.

Based on the high incidence of ADRs caused by warfarin, as demonstrated in our study and others,^2,88- 89 the potential effect of interventions to reduce ADRs from warfarin could be high because of the high usage of warfarin, which ranks 29th in US drug sales⁶²; the relatively high prevalence of poor metabolizers; and the severity of outcomes. The CYP2C9 enzyme genotype assays are readily performed at the clinical research level and are being developed for commercial use,¹³ and it is likely that clinicians would be able to interpret the results of such information. However, although several studies have found an association between the CYP2C9 genotype and ADRs, there has not been a definitive study linking CYP2C9 genotype to ADRs and warfarin dose requirements.^90- 91

Therefore, the assessment of warfarin using our criteria suggests that it provides an example of at what point pharmacogenomics information could reduce ADRs, but what is unknown is whether it actually will. Given that warfarin therapy is already individualized by assessing blood coagulation times and that the link between variant alleles and clinical outcomes is uncertain, it remains to be determined whether genotyping of warfarin patients would produce substantial additional gains.

Our analysis should be considered only a first step in examining the association between genetic variability and ADRs. Most drugs have complex metabolic pathways so that multiple variant alleles could be responsible for ADRs. Furthermore, although most drug-metabolizing enzymes exhibit variant alleles, only some of these have been associated with changes in drug effects or adverse effects. Adverse drug reactions may also be a function of variant alleles at independently segregating loci and of environmental exposures. However, our study suggests what types of future studies of these associations may be fruitful. Our study is also limited because much of the data on which it is based are incomplete or of limited quality, and by necessity such a study requires some subjective decisions. Specifically, we were unable to derive quantitative summary estimates of ADR incidence and therefore our estimate of the most common drugs is relatively crude. Also, the review articles did not include recently discovered genetic variants or drugs that are less commonly used (eg, oncology drugs or immunosuppressives). However, we supplemented the review articles with data from an extensive Web site to improve the validity of our data. Finally, our comparisons of ADR-associated drugs to randomly selected drugs might be confounded by several factors. However, our results would have had to be dramatically different to change our primary conclusions, and sensitivity analyses suggest that our conclusions are robust.

The emergence of pharmacogenomics may herald a new era of individualized therapy. Hence, nonpreventable ADRs may become at least in part preventable, as a first step in optimizing drug therapy with genetic information. This study provides empirical evidence that the use of pharmacogenomics could potentially reduce ADRs, a problem of major significance. Our study illustrates the adage, "the sum can be greater than its parts": how 2 bodies of literature can produce additional insights when combined, and our study provides a foundation for future research.

In the future, we may all carry a "gene chip assay report" that contains our unique genetic profile that would be consulted before drugs are prescribed. However, the application of pharmacogenomics information faces significant challenges, and further basic science, clinical, and policy research is needed to determine in what areas pharmacogenomics can have the greatest impact, how it can be incorporated into practice, and what are its societal implications.