As with research in common disease susceptibility, the path to translation involves a two stage process that first requires the reliable identification of the genetic loci involved, and then research into the healthcare applications of this knowledge, which includes critical appraisal of the performance of genotype as a predictive test. While the extent of the clinical impact of research in both areas is uncertain, the reliable identification of loci involved in drug response (pharmacogenetics) appears to be less advanced than the identification of susceptibility loci for common disease [3]. After more than two decades of research, a continuing expansion in the range and depth of available drug therapies, and the continued promise of ‘personalized medicine,’[4], [5], [6], [7], [8], [9] only four pharmacogenetic tests were mandated as part of the FDA drug approval pre-July 2009, [10] while for another 10 tests recommended by the FDA, clinical utility is not universally agreed [11], [12], [13]. Understanding the reasons for the blocks in development of personalised medicines could help improve efficiency of future research.

The major goal of pharmacogenetic research is development of genotype-based predictive tests of efficacy or toxicity. However, a prerequisite is the reliable identification of the relevant genetic loci. In genetic work, where many hundreds of thousands of hypotheses can be tested, research designs are needed that optimise the detection of true positive (while limiting the potential for false positive) association [17], [18], [19]. Despite some high quality studies, in broad terms, there are several features of the field as a whole that suggest that only a proportion of the positive associations reported are genuine. These include: the small size of most studies coupled with the more frequent evaluation of common rather than rare variants (whose effect sizes would be predicted to be small and which therefore requires large sample sizes for their reliable detection); use of surrogate (usually continuous) outcome measures rather than more clinically relevant binary outcomes; and subgroup analyses with multiple hypothesis testing. Our study may have been limited by analysing only the abstracts of articles satisfying inclusion criteria. However, detailed data (information unlikely to be reported in abstracts) on outcome measures (binary/continuous), gene variants and reported p values were derived from the full text of a subset of 10%, which accurately reflected the span of studies in the database.

Similar problems to those we highlight were recognised in the field of genetics of common disease a decade or so ago. What followed were efforts to systematically and comprehensively collate evidence from genetic association studies, large collaborative meta-analyses, larger primary studies, more comprehensive capture of genetic variation at any given locus, independent replication, and, most recently, whole genome association studies [26]. These developments have contributed to the discovery of many secure genetic associations that are providing new insights into disease pathogenesis, potential therapeutic targets and the possibility of developing predictive tests for disease. Several important and laudable efforts to collate and curate information on the genetic basis of drug response already exist, including those of the Pharmacogenetics Research Network [34]. However, the challenge in identifying primary pharmacogenetic studies is illustrated by our two alternative search strategies. Our comprehensive Medline search was sensitive (yielding >100,000 articles) but non-specific, with a large number of evaluated articles not satisfying our definition of a pharmacogenetic study. However, using a specific search strategy (via the MeSH tool) the majority of articles were missed. We know of no previous attempts to systematically identify all published pharmacogenetic studies in this way but our current analysis suggests that future attempts to do so should adopt an explicit, systematic and comprehensive search strategy such as the one we have used here. The terms “pharmacogenomic” and “pharmacogenetic” have both been used somewhat interchangeably in the literature. For example the Pharmacogenomics Knowledge database (PharmGKB; http://www.pharmgkb.org/resources/forGeneralUsers/pharmacogenetics_pharmacogenomics_and_personalized_medicine.jsp accessed 2009 November 10, archived URL http://www.webcitation.org/5lBBtDJPf) defines pharmacogenetics as “the study of … varying responses to drugs and the determination of the genetic mutations underlying these variations” and pharmacogenomics as “the study of drug response in the context of the entire genome”. However, the Human Genome Project information portal (http://www.ornl.gov/sci/techresources/Human_Genome/medicine/pharma.shtml#whatis accessed 2009 November 10, archived URL http://www.webcitation.org/5lBCB8i5T) defines pharmacogenomics as “the study of how an individual's genetic inheritance affects the body's response to drugs”. These indistinct classifications are exemplified by the U.S. National Library of Medicine's ‘controlled vocabulary’ for indexing articles via MeSH terminology: “pharmacogenomics” is not a MeSH term, on entering it in Medline, all articles indexed with the MeSH term “pharmacogenetics” are displayed.

Other developments that may be helpful include: a greater use of meta-analysis, particularly where four or more independent studies of the same gene have been conducted, perhaps with an online, continuously updated database similar to those established for Alzheimer's Disease, Parkinson's Disease and Schizophrenia [35], [36], [37], [38]. Other improvements might include: primary studies with larger sample sizes; wider use of haplotype tagging single nucleotide polymorphisms (SNPs); studies of rare and structural genetic variants whose effects are predicted to be larger, and which may therefore be more suited for use as predictive tests; and a greater focus on genes influencing drug handling and adverse effects, to fill gaps in knowledge [39].

Important studies with some of these features have been reported since the deadline we set for our literature search. For example, the identification of a SNP in the SLCO1B1 gene, encoding the organic anion-transporting polypeptide OATP1B1, as a susceptibility factor for statin-induced myopathy involved a genome-wide association analysis of 85 individuals with definite or incipient statin myopathy (and 90 controls) from a trial involving over 12,000 subjects [40]. Here, the small size of the genome-wide association study belies the large-scale effort to identify the few subjects who suffer extreme adverse effects. This study provides a paradigm for the identification of genetic loci underlying rare but serious adverse effects of a commonly used drug. Other examples which could be studied in a similar way include heparin-induced thrombocytopaenia (frequency 0.5–2%), oesteonecrosis of the jaw from bisphosphonate treatment (prevalence 4–7% in those receiving intravenous bisphosphonates for hypercalcaemia of malignancy), and angio-oedema from angiotensin converting enzyme inhibitors. Because of the large genetic effect sizes that might be detected with this approach (for example an odds ratio of 17 for statin myopathy in SLCO1B1 CC homozygotes), predictive tests may be more likely to emerge, though the rarity of the adverse effect means that rigorous assessment of the cost-effectiveness of the approach would first be required. Larger scale candidate gene studies[41], [42], [43], [44] are also providing much more secure evidence on loci influencing both drug response and adverse effects that might form the basis of predictive testing for dose adjustment or avoidance of toxic treatments.

As more reliable information begins to emerge on alleles influencing drug response from larger, better designed whole genome and candidate gene studies, focus will need to shift to the critical evaluation of the predictive performance of genetic tests in clinical practice, including studies of cost-effectiveness. These evaluations will require use of different metrics to those conventionally reported in discovery-based genetic studies (such as odds ratios or proportion of variance explained) [45], [46], [47]. Instead, sensitivity and specificity, predictive values and the generation of multivariate models that include genotype will need evaluating [42], [48]. In some cases, the most robust evaluation of the effectiveness of genetic tests may need to come from randomised trials comparing health outcomes among people randomised to pharmacogenetic testing or no testing, together with cost-effectiveness analyses as are now common when evaluating the usefulness of interventions. In concert, these efforts should help realise the promise of personalised medicines with resultant improvements in healthcare. Our recommendations for pharmacogenetic research are summarised below.

Document S1

Table S1