Pharmacogenomics Discovery and Implementation in GWAS Era

Abstract

Clinical response to therapeutic treatments often varies among individual patients, ranging from beneficial effect to even fatal adverse reaction. Pharmacogenomics holds the promise of personalized medicine through elucidating genetic determinants responsible for pharmacological outcomes (e.g., cytotoxicities to anti-cancer drugs) and therefore guide the prescription decision prior to drug treatment. Besides traditional candidate gene-based approaches, technical advances have begun to allow application of whole-genome approaches to pharmacogenomic discovery. In particular, comprehensive understanding of human genetic variation provides the basis for applying GWAS (genome-wide association studies) in pharmacogenomic research to identify genomic loci associated with pharmacological phenotypes (e.g., individual dose requirement for warfarin). We therefore briefly reviewed the background for pharmacogenetic/pharmacogenomic research, with statins and warfarin as examples for the GWAS discovery and their clinical implementation. In conclusion, with some challenges, whole-genome approaches such as GWAS have allowed unprecedented progress in identifying genetic variants associated with pharmacological phenotypes, as well as provided foundation for the next wave of pharmacogenomic discovery utilizing sequencing-based approaches. Furthermore, investigation of the complex interactions among genetic, and epigenetic factors on the whole-genome scale will become the post-GWAS research focus for pharmacologic complex traits.

The clinical response to prescribed drugs is often not consistent across individual patients or populations. At the extreme ends of the drug response spectrum can be severe adverse drug reactions, which have been shown to be one of major health problems responsible for hospital admissions, as well as over 100,000 annual deaths in the United States¹. Understandably, this kind of variation in drug response is of particular concern for those drugs with narrow therapeutic indexes. For example, warfarin, a widely prescribed anticoagulant, exhibits substantial inter-individual variability of dose requirement. Therefore, if not properly monitored, inappropriate warfarin doses may increase patients’ risks for either thrombosis (i.e., dose too low) or serious bleeding (i.e., dose too high). Ideally, being able to identify patients that will respond well, badly, or not at all prior to treatment could greatly improve health care, by potentially maximizing the beneficial effect and minimizing the adverse effect of prescribed medicine.

Previous studies have shown that drug response is likely a complex phenotype or trait that is affected by both genetic and non-genetic factors (e.g., dosage, diet, environment), as well as patient-specific factors such as age, health condition, and socioeconomic status. Genetics (e.g., through regulating drug-metabolizing enzymes), however, has been demonstrated to contribute a significant proportion of drug response variability. For example, genetic variants in CYP2C9 (cytochrome P450, family 2, subfamily C, polypeptide 9) and VKORC1 (vitamin K epoxide reductase complex, subunit 1) were found to contribute significantly to the variability of warfarin dose requirement, independent of other non-genetic factors². In oncology, using a cell-based model, the genetic contribution (e.g., through single nucleotide polymorphisms [SNPs] regulating drug response associated gene expression phenotypes, i.e., expression quantitative trait loci [eQTLs]) have also been clearly demonstrated for several chemotherapeutic agents^{3, 4}(e.g., etoposide, carboplatin, cisplatin, daunorubicin, cytarabine)⁵. Pharmacogentics and pharmacogenomics, therefore, hold the promise for the ultimate aim of personalized medicine by investigating the relationships between a patient’s genetic make-up and various therapeutic phenotypes.

The Genetic Basis for Pharmacogenomic/Pharmacogenetic Research

The completion of the first human genome sequences represented the beginning of the genomic era of biomedical discovery ⁶. Human genetic variation refers to the observed differences in DNA sequences (i.e., polymorphisms or genetic variants) both within and across populations. With the advances of sequencing and genotyping technologies, it is now feasible to obtain comprehensive genetic variation data of multiple genes, even at a genome-wide scale, in multiple samples efficiently (e.g., using whole-genome genotyping arrays and the next-generation sequencing technologies).

During the past decade, the general patterns of human genetic variation and the genetic architectures of a variety of complex traits and phenotypes such as gene expression, susceptibilities to common diseases, and response to therapeutic treatment have begun to be elucidated with the availability of comprehensive population-based genetic variation data for large genetic variation cataloging efforts including the International HapMap Project^{7, 8} and the 1000 Genomes Project⁹. For example, significant variation in quantitative gene expression, and gene regulation through genetic variants in the forms of SNPs and copy number variants (CNVs) have been investigated both within and across human populations using these models and genetic variation data^10-17. Similar to quantitative gene expression, drug response is likely to be a heritable phenotype, which may be at least partially attributed to genetic diversity. For example, a previous study estimated that genetic factors could determine more than 40% of the susceptibility to cytotoxicity induced by cisplatin¹⁸, a platinating agent used in treating various types of human cancers including lung cancer, lymphoma, thus indicating that sensitivity to the cytotoxic effects of chemotherapeutic agents could be under substantial genetic influence.

General Approaches in Pharmacogenomic/Pharmacogenetic Research

For investigating the relationships between genes/genetic variants and therapeutic phenotypes, in general, there are two categories of research approaches: i.e., the candidate gene approach and the whole-genome approach. Briefly, studies applying the candidate gene approach focus on a limited number of candidate genes based on a priori knowledge for a therapeutic phenotype of interest. Therefore, a successful candidate gene study relies on the selection of candidate genes and polymorphisms that may be potentially relevant to the therapeutic phenotype of interest. In general, candidate genes and polymorphisms can be selected based on the relevance of a known pathway or evidence from previous studies. For example, in a previous study using the candidate gene approach, genetic polymorphisms in TPMT (encoding thiopurine S-methyltransferase) were successfully identified to decrease TPMT enzyme activity and subsequently increase 6-mercaptopurine toxicity²³. In this case, the choice of candidate gene TPMT was based on the key role of thiopurine methyltransferase plays in converting the active drug to its inactive metabolite, 6-methylmercaptopurine ²⁴. As mentioned above, another successful example for the candidate gene approach is the identification of polymorphisms in CYP2C9 and VKORC1 associated with inter-individual variability in the dose response of warfarin²⁵. Warfarin is metabolized primarily via oxidation in the liver by CYP2C9, and exerts its anticoagulant effect by inhibiting the VKORC1 protein.

Though the candidate gene approach focuses on potential genes with functional relevance (e.g., drug-metabolizing enzymes for a pharmacokinetic phenotype), there still can be unknown genes or unidentified genetic variants that may help determine or modify the phenotypes of interest, e.g., through interactions with those known candidate genes. As a matter of fact, drug response phenotypes are often not explained sufficiently with known genes, this suggests that there could be novel genes, pathways and genetic variants unidentified in previous studies also responsible significantly for therapeutic phenotypes.

Furthermore, epigenetic markers including microRNAs ¹⁹ and DNA methylation^20-22 have emerged to be crucial regulators of quantitative gene expression phenotype, thus providing the foundation for investigating the complex interactions among genetic, epigenetic factors and complex traits or phenotypes. Indeed, various genome-wide molecular profiling platforms are available for this purpose, including platforms for whole-genome gene expression, microRNA, as well as cytosine modification levels quantification. Therefore, the whole-genome approach could be used to expand and complement the candidate gene approach to interrogate the entire human genome, transcriptome, or epigenome for the relationships between genes/genomic loci and therapeutic phenotypes. For example, a SNP in NRG3 gene (a gene that has never been linked to cancer) was found to be associated with cellular sensitivity to platinum agents through its effect on genome-wide gene expression and this association was replicated in ovarian cancer patients who underwent platinum-based therapy ²⁶. More recently, the next-generation sequencing technologies²⁷ further opened up the possibility to identify all genetic variants and mRNA transcripts (i.e., through RNA-sequencing), thus potentially allowing direct mapping of the causal genetic variants or transcripts contributing to therapeutic phenotypes of interest without relying on linkage disequilibrium between genotyped variants and causal variants or sequences of known transcripts.

Interpretation of GWAS pharmacogenomic findings

GWAS in theory provided an unbiased way to demonstrate the relationship between genome-wide genetic variants and the phenotypic traits of interest. A Manhattan plot or QQ plot are often employed to illustrate these relationships. With the evolution of the genomic technology, GWAS gradually becomes a common method employed to study the relationship between human genome and drug response. Indeed, GWAS has been performed to evaluate genetic contribution to the treatment effect of various commonly used medications in the recent 5 years; for example, response to anti-TNF treatment in rheumatoid arthritis ²⁸, flucloxacillin-induced liver injury ²⁹, QT prolongation ³⁰ and response to ³¹ iloperidone in treatment of schizophrenia, metabolic side effects³² and parkinsonism severity ³³ induced by antipsychotic drugs, treatment response to antipsychotics ³⁴, antidepressant response with escitalopram and nortriptyline ³⁵, response to citalopram in major depressive disorder ³⁶, antihypertensive response to thiazide diuretic ³⁷, angiotensin II receptor blocker (candesartan) for antihypertensive response ³⁸, glycemic response to metformin in type 2 diabetes ³⁹, statin-mediated reduction in total and LDL-cholesterol ⁴⁰, smoking cessation ⁴¹, bisphosphonate-induced osteonecrosis of the jaw ⁴², methotrexate disposition in acute lymphoblastic leukemia (ALL) children⁴³, overall survival of advanced non-small cell lung cancer patients treated with carboplatin and paclitaxel ⁴⁴, glucocorticoid response in asthma patients ⁴⁵, and short-acting β(2)-agonist bronchodilator response ⁴⁶.

These examples demonstrated the improvement of genomic technology where genomic coverage increased significantly from the initial 100,000 SNPs microarray to those that covered 1 million SNPs. They also showed the diversity of phenotypic traits studied, where both therapeutic efficacy and toxicity were of interest. Changes in technology, phenotypic diversity as well as increased sophistication in analytical methods in GWAS highlight the needs to properly interpret the GWAS findings.

From a phenotypic point of view, individual sensitivity to a given drug can be assessed through wide arrays of methods, from PK end points (e.g., amount of active drug in the blood stream) to disease symptom control (e.g., heart rate, blood pressure, tumor size reduction) to survival outcome and to patient quality of life. Various technologies can be employed to quantify each of these measurements. Therefore, it is extremely important when interpret GWAS findings to be within the context of the exact phenotypic traits studied. For example, when significant associations were observed between a set of genetic variants and blood level of a drug, it is incorrect to extrapolate to a conclusion that the genetic variants are affecting drug effect in patients. In addition, when conducting pharmacogenomic GWAS in clinical patient populations, the phenotypic heterogeneity can become a major barrier in proper identification of genetic contributors for the trait of interest.

Early GWASs tend to focus on the top associated relationship(s) as the so called “low hanging fruit”, for which the strongest genotypic-phenotypic relationship was seen for a given trait. Furthermore, the terms of genetic variants and genes are often used interchangeably in these early GWASs. In another word, when a genetic variant was found to be significantly associated with a phenotypic trait, the gene where the variants reside in or near was often considered to be the cause of the observed genotype-phenotype relationship. The method used to identify these signals is straight forward (e.g., case control association test) and the interpretation of the results is often straight forward as well. This type of findings usually has larger effect size and therefore is clinically relevant and can be adapted into clinical practice easily (see the statin example). Note that these early pharmacogenomics GWAS successes where genetic variants concentrated in a few genes/genomic loci and show very significant association, tend to relate to more rare clinical phenotypes, for example flucloxacillin-induced liver injury happens in less than 1 out of 5000 individuals ²⁹ and statin-induced myopathy at regular statin dosage occurs at a rate of 0.01% ⁴⁷.

Despite the success of these GWASs, the issue of missing heritability dampened the enthusiasm of using GWAS to improve our understanding of genetic links to human complex traits ⁴⁸. While considerable efforts are ongoing looking for rare but functional important genetic variants; more people started to evaluate those moderate genetic associations that could act together to affect phenotypic traits. For example, the development of various polygenic models that combine/integrate individual genetic signals together to predict phenotypic traits. The methods used to integrate moderate genetic signals include but not limit to location based approach (e.g., SNP located in transcription factor or microRNA binding sites) and functional based approach e.g., eQTL studies ^{5, 19, 49}. Instead of focusing on only association p value, which could be moderate in these cases, these studies applied system biology tools which integrate various functional relevant endophenotypes (e.g., gene expression in the form of mRNA or protein, microRNA expression and DNA methylation) to depict potential biological mechanism behind the observed genotype-phenotype association. These new integrative approaches facilitate the filtering of true genetic signals from potential “noise” from a GWAS and expand our understanding of genotypic-phenotypic relationship from simple association to a biological functional network.

Implementation of GWAS pharmacogenomic findings

From the implementation point of view, the field of medicine has always strived to improve the therapeutic outcomes by providing the right drug to the right person based on the knowledge available. Prior to the genomic era, the medication choice is often based on individual physiological (e.g., allergic history) and pathological characteristics (e.g., specific disease); while dose adjustment is made based on patients’ age, organ function, and sometimes body weight, body surface area, or gender. These long validated parameters are and will continue to be used in Medicine to guide drug therapy. Some of these clinical parameters could be driven by genetic factors and therefore served as an endophenotype in guiding medication choices. In the post genomic era, multi-disciplinary teams that consist of laboratory, computational and clinical research communities are working together to further improve drug therapy by incorporating individual’s genetic information (both germline and somatic) into the existing prescribing models ⁵⁰.

Ideally, for a genetic variant to be considered important and get incorporated into clinical therapeutic decision, it needs to 1) robustly predict clinical treatment outcomes; 2) clinical consequence is relevant and significant; 3) added value to existing prescribing method; 4) cost-effective; and 5) actionable with existing alternative dosage or therapy. Furthermore, there are many practice hurdles, for example, the knowledge dissemination (e.g., prescriber education), the reimbursement, and the technical consistency ⁵¹. In the next section, we will provide examples of a few pharmacogenomic studies that employed the whole-genome approach and their current status of clinical implementation.

Examples

Statins

Statins (namely simvastatin, lovastatin, pravastatin, fluvastatin, atorvastatin, rosuvastatin and pitavastatin) are among the top 10 prescribed drugs (atorvastatin, simvastatin and rosuvastatin) in the US (http://www.pharmacytimes.com/publications/issue/2012/July2012/Top-200-Drugs-of-2011). They are commonly used to lower low-density lipoprotein cholesterol by inhibiting cholesterol synthesis pathway. Statins are relatively safe with the most common side effect being skeletal muscle toxicity. Depend on the severity, this muscle toxicity can be classified as incipient myopathy, myopathy, or rhabdomyolysis, which is life threatening but rare with a frequency of 1 in 100,000 ⁵². Statin induced myopathy is dosage dependent ⁴⁷. Therefore, in 2011 the FDA announced an update to the simvastatin product label recommending against initiation of the 80-mg dose and cautioning against continuation of an 80-mg dose unless the patient has already tolerated it without muscle problems for more than 1 year (http://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/ucm258338.htm).

The first GWAS for statin-induced myopathy was conducted using approximately 300,000 markers in 85 subjects with definite or incipient myopathy and 90 controls, all of whom were taking 80 mg of simvastatin daily as a part of a trial involving 12,000 participants. Genetic variants within SLCO1B1 [solute carrier organic anion transporter family, member 1B1], were found to be significantly associated with the risk of developing myopathy ⁴⁷. Resequencing indicated rs4149056 is the putative causative allele with OR of 4.5 per minor C allele ⁴⁷. This finding has since been replicated in several independent studies including the Heart Protection study ⁵³, the STRENGTH (Statin Response Examined by Genetic Haplotype Markers) study ⁵⁴ and a retrospective case-control study⁵⁵. Note that although the association between rs4149056 and simvastatin induced adverse event is strong in STRENGTH trial, there is rather modest relationship between this SNP and atorvastatin and no significant association observed for pravastatin ⁵⁴. This lack of association between rs4149056 and pravastatin was also confirmed elsewhere ⁵⁵.

Based on these strong scientific evidence, the Clinical Pharmacogenomics Implementation Consortium (CPIC) recently published a Guideline for SLCO1B1 and Simvastatin-Induced Myopathy⁵². This guideline focused on reducing muscle toxicity and optimize patient adherence with simvastatin treatment. To date, the pharmacogenetic testing of SLCO1B1 variants is not mandated by either the US FDA or European Medicine Agency (EMA) ⁵⁶.

Warfarin

Warfarin is the most commonly prescribed oral anticoagulant worldwide ⁵⁷. It is used for the treatment and prevention of thrombotic disorders by inhibiting the vitamin K-dependent clotting pathway. Despite its efficacy, warfarin has a narrow therapeutic index and individual response to warfarin is highly variable. Too much of the drug will result in hemorrhage, (e.g., cranial hemorrhage which could be deadly); while too little of the drug will result in undesired coagulation, (e.g., deep vain thrombosis (DVT)). Not surprisingly, adverse events related to inappropriate warfarin dosing is one of the most common reasons for emergency room visits ⁵⁸. Therefore, considerable efforts are invested to dose warfarin correctly based on patients’ age, size, other medications, diet, and more recently genetics. For a GWAS with warfarin, many phenotypic traits can be studied (e.g., incidence of cranial hemorrhage, or INR). Each has its pros and cons. For cranial hemorrhage, although it is extremely important, due to the low incidence of such event, it is very difficult to recruit sufficient number of patients for a GWAS. To date, GWASs conducted for warfarin have been mainly focusing on identification of genetic variants that could be used to predict stabilized warfarin dose, which has a direct translational importance.

Prior to GWAS era, candidate gene studies have been focusing on CYP2C9, an enzyme predominantly responsible for metabolizing S-warfarin (the more potent enantiomer as compared to R-warfarin) and VKORC1, the target of warfarin. CYP2C9 variant alleles were found to associate with low warfarin dose requirement ⁵⁹, with as high as 6.21 odds ratio ⁶⁰. VKORC haplotypes was also found to link to warfarin dose ⁶¹.

Initial GWAS was performed in 181 White warfarin treated patients, followed by independent replication in 374 patients for stabilized warfarin dose ⁶². Approximately 550,000 polymorphisms were evaluated and the most significant independent effect was found for VKORC1 polymorphism (p = 6.2 × 10⁻¹³) with moderate association observed for CYP2C9 variants (rs1057910 and rs4917639, p approximately 10⁻⁴) ⁶². Subsequent GWAS was performed on a larger scale with a sample size of 1,053 ⁶³. Univariate regression analysis found the strongest statistical signals (p<10⁻⁷⁸) at SNPs clustering near VKORC1 and the second lowest p-values (p<10⁻³¹) emanating from CYP2C9. Multiple regression adjusting for known influences on warfarin dose (VKORC1, CYP2C9, age, gender) identified a single SNP (rs2108622) with genome-wide significance (p = 8.3 × 10⁻¹⁰) that alters protein coding of the CYP4F2 gene⁶³. The role of variants in CYP4F2 gene was also confirmed by a separate GWAS conducted in Japanese ⁶⁴.

Based on these scientific findings, the US FDA modified the warfarin label, stating that CYP2C9 and VKORC1 genotypes may be useful in determining the optimal initial dose of warfarin in 2007 ⁶⁵. The label was further updated in 2010 to include a table describing recommendations for dosing ranges for patients with different combinations of CYP2C9 and VKORC1 genotype. To date, several dosing algorithms have been developed that incorporate genetics information along with nongenetic clinical algorithms ^{66, 67}.

Several small scale randomized prospective studies have been conducted to examine the role of pharmacogenetics in guiding initiation of warfarin ^{68, 69}. Although limited in sample size and study design, they tend to support the role of pharmacogenetic testing in warfarin dosing. Furthermore, there are several large scale ongoing randomized controlled trials to evaluate the role of genetic in warfarin dosing, for example, the Clarification of Optimal Anticoagulation through Genetics (COAG), the European Pharmacogenetics of AntiCoagulant Therapy (EU-PACT), the Genetic InFormatics Trial (GIFT), Pharmacogenetic dosing of warfarin: a controlled randomized trial by the Taiwan Warfarin Consortium, and Warfarin Adverse Events Reduction for Adults Receiving Genetic Testing at Therapy Initiation (WARFARIN) ⁵⁷.

Recently, the CPIC developed a peer-reviewed gene-drug guideline to assist in the interpretation and usage of CYP2C9 and VKORC1 genotype data for estimating therapeutic warfarin dose to achieve an INR of 2-3, should genotype results be available to the clinicians ⁵⁷. The guideline states “the greatest benefit of genetics in warfarin dosing is likely to be in patients requiring less than 21 mg/week or more than 49 mg/week (constituting >40% of all patients). Genetics-based algorithms also better predict warfarin dose than the FDA-approved warfarin label table. The best way to estimate the expected stable dose of warfarin is to use the algorithms available on http://www.warfarindosing.org and the dosing algorithm published by the International Warfarin Pharmacogenetics Consortium at http://www.pharmgkb.org/do/serve?objId=PA162372936&objCls=Dataset#tabview=tab2.

Conclusion

In summary, technical advancement in high throughput genotyping has allowed the application of GWAS in pharmacogenomic discovery. Compared to the traditional candidate gene approach in pharmacogenetics, GWAS can provide an unbiased, comprehensive scan of the human genome for potential genomic loci responsible for therapeutic outcome phenotypes. Given that pharmacological phenotypes are likely to be complex and influenced by a variety of genetic and non-genetic factors, future genome-wide approaches integrating molecular data (e.g., epigenetic markers) and genotypes could enhance our understanding of the complex networks, including gene-gene and gene-environment interactions, underlying pharmacological phenotypes. Furthermore, the next-generation sequencing technologies may advance the next wave of pharmacogenomic discovery to the unprecedented depth beyond the current GWAS efforts.