Personalised medicine: not just in our genesBMJ 2012; 344 doi: http://dx.doi.org/10.1136/bmj.e2161 (Published 03 April 2012) Cite this as: BMJ 2012;344:e2161
- 1Institute for Clinical Research and Health Policy Studies, Tufts Medical Center, Boston, MA, USA
- 2Department of General Internal Medicine, Lahey Clinic Medical Center, Burlington, MA
- 3Clinical and Translational Science Program, Sackler School of Graduate Biomedical Sciences, Tufts University, Boston, MA 02111, USA
- Correspondence to: D M Kent
- Accepted 13 February 2012
In the 10 years since the completion of the human genome project there has been tremendous scientific progress in genomics. After a brief period of promiscuous false discovery, methodological advances based on agnostic statistical approaches have yielded hundreds of confirmed associations between genes and disease.1 Rapid improvements in “next generation” sequencing mean that people will soon be able to carry their genome on a memory stick at an affordable cost. What scientists have yet been unable to provide, however, has been a compelling reason why anyone would want to do so.
Despite the high expectations for accurate forecasting of disease based on genetic information, and the marketing of direct-to-consumer genetic tests,2 it seems increasingly less likely that genes will have a major clinical effect on the prognosis of common complex diseases. The effects of single nucleotide polymorphisms (SNPs) tend to be small,3 they typically add little information to easily obtainable clinical or phenotypic information,4 and even in combination typically account for only a small proportion of heritability.5 The products of even the most reputable of the direct-to-consumer genetic testing companies (Navigenics, 23andMe, deCODE) can show marked differences in the calculated relative disease risks for individuals.6 Although the reasons common SNPs are so weakly predictive are not wholly settled, if SNPs interact complexly (with themselves or environmental factors) to cause complex diseases or if substantial variation in risk is the result of very uncommon genetic variants, then the use of SNPs to predict disease may prove statistically intractable. History underscores the large gap between statistically detectable associations and clinically actionable information.
However, pharmacogenomics (the study of genetic predictors of response to treatment) still holds the promise for realising personalised medicine. Since the metabolism and actions of drugs depend on particular enzymes and targets under genetic control, pharmacogenomics may yet deliver us from the “one size fits all” approach that has plagued evidence based medicine by revealing the genetic determinants of heterogeneity in treatment effect.
In oncology there are several well known examples of somatic mutations (in the genotype of the tumour) that are strong determinants of drug response. However, few examples exist of pharmacogenomic testing for (inherited) germline variation in SNPs that have entered clinical use.7 To understand better whether germline SNPs will have an important role in tailoring treatment for common chronic diseases, we examine the success of pharmacogenomics in cardiovascular medicine, the subspecialty with perhaps the most mature evidence base.
To identify the most promising associations and assess their evidential support we created an evidence map of literature on cardiovascular pharmacogenomics published from 2005 to February 2011 (see table A on bmj.com for search strategy). We classified the design of the studies according to which phase of clinical translation they informed8: basic biomedical research (candidate gene or genome-wide association scans (GWAS)); clinical research (pharmacogenomic randomised clinical trials, cost effectiveness analyses); clinical application (adoption by clinical practice guidelines). Given the well known limitations of candidate gene studies and the findings of a recent synopsis of the pharmacogenomics literature,9 we defined confirmed pharmacogenomic associations as those supported by meta-analyses with significant summary results, discovered by GWAS, or included in the US Food and Drug Administration labelling of any cardiovascular drug.10
We identified 289 cardiovascular pharmacogenomics studies; 94% pertained to the basic biomedical research phase (90% candidate gene and 4% GWAS) and the remaining 6% to subsequent translational phases. Every major class of cardiovascular drugs has been explored for pharmacogenomics associations, with statins, β blockers, clopidogrel, and warfarin being the most widely investigated. In 80% of primary studies the intended use of the pharmacogenomics test was to improve efficacy of a cardiovascular drug; 15% of studies aimed to optimise dosing of drugs with a narrow therapeutic index (such as warfarin), and 5% focused on avoiding drug related adverse events.
Most basic biomedical research (84%, 229 studies) reported positive findings, resulting in 220 unique positive SNP-drug associations. Of those 220 positive associations, 19 (9%) were confirmed according to our criteria (table⇓): five by meta-analysis, 13 by GWAS, and eight by inclusion in FDA labelling.
None of the 19 confirmed associations has yet been recommended for use in clinical practice guidelines. The most strongly supported pharmacogenomic associations were for clopidogrel and warfarin. Meta-analyses support a strong association between reduced activity CYP2CI9 alleles and phenotypic tests for platelet reactivity11 12 13 14 as well as for higher rates of adverse cardiovascular outcomes in patients taking clopidogrel. However, analyses of “repurposed trials” of clopidogrel—that is, studies that genotyped stored samples of participants in placebo controlled trials enabling a test of the interaction between genotype and treatment15 (see table B on bmj.com)—found no evidence of heterogeneity of effect across different CYP2C19 polymorphisms,16 suggesting that these tests have limited value for guiding treatment decisions.17 As yet, no trials have examined the clinical effect of “test and treat” strategies, and guidelines do not recommend pharmacogenomic testing before prescribing clopidogrel.18 19
CYP2C9 and VKORC1 gene effects on warfarin dosing are perhaps the best established pharmacogenomic associations in the cardiovascular literature and are the only markers for which a test and treat strategy has been assessed in randomised trials. However, a systematic review of these studies found no significant benefit of pharmacogenomic testing on clinical (major bleeding) or surrogate outcomes (time spent in therapeutic international normalised ratio (INR) range),20 and testing for these has also not been adopted into clinical guidelines.21 22 23 24
Meanwhile, successors to both warfarin and clopidogrel have emerged on the market (prasugrel and ticagrelor for clopidogrel; ximelagatran, rivaroxaban, and dabigatran for warfarin). This introduces the question of whether pharmacogenomic based decisions for warfarin or clopidogrel are superior not just to standard methods for determining dose but to routine use of the newer generation drugs, for which no pharmacogenomic information is needed. Notwithstanding the cost of newer drugs, these examples highlight how new drug discovery can potentially render painstakingly developed pharmacogenomic approaches obsolete, even before they are implemented.
The evidence map also revealed other disconcerting patterns. Although the FDA labelling of seven drugs included pharmacogenomic associations for a total of eight genes, five of these have not been confirmed by meta-analysis or GWAS. For example, FDA labelling for atorvastatin indicates that the LDLR gene modifies lipid lowering response, but GWAS across common variants of LDLR25 found no association. The evidential threshold used by the FDA in making labelling decisions is thus unclear.10 17 Moreover, cost effectiveness analyses are often conducted for genetic tests that have unclear or even refuted associations, yet these studies invariably find in favour of pharmacogenomic testing.
Most pharmacogenomics studies sought to examine variation in efficacy and used clinically obtainable phenotypic outcome measures (such as blood pressure or lipid levels) that are surrogates of long term clinical outcomes. However, these phenotypic outcomes are themselves alternative methods for predicting long term response and are easily monitored in routine clinical practice. This limits the value of information gained from pharmacogenomic testing. For example, no pharmacogenomic test based on phenotypic measures can be more accurate in predicting response to antihypertensive drugs than measuring blood pressure itself or to statins than measuring cholesterol change. This partly explains why pharmacogenomic testing for warfarin dosing has had relatively limited clinical impact—direct measurement of anticoagulation (through the international normalised ratio (INR)) becomes the over-riding determinant of dosing after the first few doses, with or without pharmacogenomic testing. More consideration needs to be given to the clinical context where pharmacogenomics would be most helpful when developing tests.26
Bridging the gap between pharmacogenomics and personalised medicine
Although it is perhaps premature to make an overall judgment about the use of pharmacogenomic testing to predict treatment response, our review of cardiovascular medicine suggests a similar pattern to that found for prediction of common diseases. The studies identified abundant SNP associations but no clear case for clinical effect. The influence of SNPs on drug response is often relatively small, and the incremental value of the genomic information on top of readily obtainable clinical information is unclear.
At the time of writing, the FDA website lists 78 different pharmacogenomic associations that are included in drug labels.10 It is not difficult to imagine that in 5-10 years’ time the list will be an order of magnitude longer and that clinicians will have searchable access to their patients’ full genome. If the era of personalised medicine can be realised by compiling a critical mass of pharmacogenomic associations of unclear clinical importance, then we are surely on the right path. However, our evidence map suggests that, for most common chronic diseases, this riot of information will not bring us any closer to tailored treatment. Although it is possible that some yet undiscovered pharmacogenomic associations may provide actionable predictive information, this is likely to be a rare exception. More generally, individual SNPs will provide only a small piece of a much larger puzzle dominated by abundant (though often underused) clinical or phenotypic information.
Notwithstanding these limitations, the goals of personalised medicine are important; the limitations of one size fits all treatment based on the summary results of clinical trials are well established.27 28 29 Ironically, the major methodological barrier to more individualised therapies is not a paucity of predictors of outcomes and effects, it is that patients have so many attributes that potentially affect risk and response to treatment that subgroup analysis is statistically unmanageable. We currently lack a consistent analytical approach that informs how a patient’s multiple attributes combine to affect the fundamental determinants of the desirability of treatment—that is, the individual’s risk of bad outcomes in the absence of treatment versus that individual’s risk if treated.
One method proposed to reduce the dimensionality of subgroup analysis is to use regression models or simple risk scores that combine variables to describe dimensions of risk across which treatment effects are likely to vary.29 30 31 32 33 In this manner, multiple subgroup analyses can be simplified to just a few fundamental dimensions of risk that mathematically determine treatment effect—for example, the risk of the primary outcome, the risk of treatment related harm, and direct modification of the effect of treatment. Though methodological challenges remain, a framework for analysing clinical trials that prioritises the reporting of treatment effect across the range of outcome risk has been proposed,30 34 and there are several examples in cardiovascular medicine of clinically important variation in treatment effect based on risk. For example, the benefits of warfarin in atrial fibrillation, the benefits of aspirin and statins for primary prevention of coronary heart disease, and the benefits of more intensive therapies for acute coronary syndromes have all been shown to be dependent on baseline risk, which requires the simultaneous assessment of multiple clinical risk factors.
Pharmacogenomics has been focused on the discovery of individual SNPs that influence variability in drug response. While emerging methods for bioinformatic and system biology analysis hold promise for gaining insights into the mechanisms of complex interactive systems, we still lack even a basic framework that permits the multiple patient attributes that influence the effect of treatment (be they clinical, genetic, biological, or environmental) to be meaningfully integrated to support personalised decision making. The role of genetic information within this broader context is still to be determined. However, “pharmacogenomic exceptionalism”—the notion that genetic information is uniquely important in determining the risks and benefits of treatments—is clearly unwarranted and counterproductive to the broadly shared goal of tailoring care to individuals.
Cite this as: BMJ 2012;344:e2161
Contributors and sources: GDK has expertise in molecular evidence based medicine and has published widely on meta-analyses of genetic and pharmacogenomic markers of common diseases and treatments. DMK has expertise in methods for studying treatment effect heterogeneity and has a long research interest in the delivery of individualised patient care GDK and DMK co-developed the ideas in the paper and the analysis plan, and cowrote the paper. GDK led the literature search and data extraction, with the assistance of Jennifer Donovan and Abdmoain Abudabrh in development of the evidence map. DMK provided funding and supervision.
Completing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare partial funding of the work in this article from the National Institute of Health (UL1RR025752); no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; and no other relationships or activities that could appear to have influenced the submitted work.