Personalized medicine: Genetic risk prediction of drug response

Abstract

Pharmacogenomics (PGx), a substantial component of “personalized medicine”, seeks to understand each individual's genetic composition to optimize drug therapy reactions (ADRs). Drug responses are highly variable because innumerable factors contribute to ultimate phenotypic outcomes. Recent genome-wide PGx studies have provided some insight into genetic basis of variability in drug response. These can be grouped into three categories. [a] Monogenic (Mendelian) traits include early examples mostly of inherited disorders, and some severe (idiosyncratic) ADRs typically influenced by single rare coding variants. [b] Predominantly oligogenic traits represent variation largely influenced by a small number of major pharmacokinetic or pharmacodynamic genes. [c] Complex PGx traits resemble most multifactorial quantitative traits –– influenced by numerous small-effect variants, together with epigenetic effects and environmental factors. Prediction of monogenic drug responses is relatively simple, involving detection of underlying mutations; due to rarity of these events and incomplete penetrance, however, prospective tests based on genotype will have high false-positive rates, plus pharmacoeconomics will require justification. Prediction of predominantly oligogenic traits is slowly improving. Although a substantial fraction of variation can be explained by limited numbers of large-effect genetic variants, uncertainty in successful predictions and overall cost-benefit ratios will make such tests elusive for everyday clinical use. Prediction of complex PGx traits is almost impossible in the foreseeable future. Genome-wide association studies of large cohorts will continue to discover relevant genetic variants; however, these small-effect variants, combined, explain only a small fraction of phenotypic variance –– thus having limited predictive power and clinical utility.

1. Introduction

How successful will we be in predicting drug responses caused by inherited genetic variation? This is the primary question addressed herein. Inter-individual variability in response is defined as an “effect of varying intensity occurring in different individuals receiving a specified drug dose”, or a “requirement of a range of doses (concentrations) in order to produce an effect of specified intensity in all patients.” Patients are well known to vary widely in their responses to drugs (Brunton, Chabner, & Knollman, 2011).

A substantial subset of personalized medicine comprises personalized medication. Over the past six decades, this field of study began by being called pharmacogenetics; however, since the Human Genome Project started in 1990, the term pharmacogenomics has become more popular. Although the terms are often used interchangeably –– for many of us in the field, there are subtle differences, i.e. effect of individual genes (pharmacogenetics) vs total genomic expression (pharmacogenomics), in response to a drug. Pharmacogenomics aims to develop rational methods to optimize drug therapy with respect to the patient's genotype, as well as to ensure maximum efficacy with minimal adverse effects in each individual. Any drug capable of producing a desired therapeutic effect –– can also produce unwanted, or adverse, side effects (Edwards & Aronson, 2000).

All the above-mentioned drug responses (therapeutic effect, adverse effect and toxic effect) are regarded in the field of genetics as phenotypes, or traits, which herein we collectively call “pharmacogenetic or pharmacogenomic traits” (PGx traits). Most drug responses are expressed as multifactorial traits –– similar in many ways to human complex diseases (e.g. type-2 diabetes, schizophrenia, cancer), as well as quantitative traits such as height, blood pressure or serum lipid levels. One major difference between drug-response and complex-disease is obvious: any individual not challenged with a particular drug will never know his or her phenotype for that drug.

The degree of success in predicting outcome of a drug before treating the patient will depend on the “genetic basis of the PGx trait” (genotype), i.e. number of genetic variants contributing to that phenotype, allele frequency and effect-size of each contributing genetic variant, and interactions between them and with other environmental factors (Park et al., 2011). This review examines the genetic architecture (genetic basis) of various drug responses (PGx traits) and discusses statistical feasibility of genetic prediction.

2. Brief history of genetics

2.1. Gregor Mendel

Principles of “dominant-versus-recessive” classical genetics began in the 1860s by Mendel, using garden peas as his experimental model system (e.g. red flower color dominant, white recessive). Whereas the F₁ cross yields all red flowers, the F₂ generation yields three red (two are R/r heterozygotes, and one R/R homozygote) and one white flower color, r/r; this F₂ population represents the “Mendelian pattern of inheritance”, or distribution (Fig. 1A).

Fig. 1.

Phenotype distribution of different traits. A: a recessive Mendelian trait with two discrete phenotypes; B: a distinct codominant Mendelian trait with three discrete phenotypes; C: a quantitative trait –– controlled predominantly by a large-effect gene, and undoubtedly additional modifiers showing continuous distribution, with three distinct modes; D: a quantitative trait influenced by numerous genetic and environmental factors, i.e. a polygenic trait that follows a normal distribution.

Recent advances have greatly expanded Mendel's studies, e.g. showing in the garden pea (Pisum sativum) at least two genes involved in red flower color: the A gene codes for a bHLH transcription factor that regulates anthocyanin pigmentation. The white-flowered mutant allele most likely used by Mendel is a simple G-to-A transition in a splice-donor site –– resulting in a mis-spliced mRNA and premature stop codon; the A2 gene encodes a WD40 protein, which is part of an evolutionarily highly conserved regulatory complex (Hellens et al., 2010).

2.2. Garrod's “inborn-errors-of-metabolism”

In the first decade of the 1900s, Sir Archibald Garrod described “in-born-errors-of-metabolism”: albinism, alkaptonuria, cystinuria and pentosuria. Each of these distinct clinical autosomal recessive traits show a pattern of inheritance similar to that of white flower color of Mendel's garden pea. Garrod is credited with ushering in the era of human genetics; the predominant underlying tenet was “one gene, one disease”, or “one wild-type (healthy) allele, and one disease allele”. For each pregnancy, two healthy parents, “carriers” heterozygous for a disease allele, bring a 25% chance of producing a child having both disease alleles and, hence, inheriting the unwanted disorder.

2.3. Single-coding mutations with severe effect

Garrod's four disorders (described above) are among the first single-coding mutations causing severe clinical disorders; these usually skip a generation. During the next several decades, many additional examples of autosomal recessive clinical diseases were described (e.g. maple syrup urine disease, phenylketonuria, cystic fibrosis, Gaucher disease, congenital adrenal hyperplasia); similarly, these diseases are autosomal recessive.

Autosomal dominant traits began to be discovered (e.g. Huntington disease, achondroplastic dwarfism, Marfan syndrome, neurofibromatosis, hereditary spherocytosis) in which the trait typically appears in each generation because heterozygotes exhibit the disease. In addition, random germline mutations (de novo mutations) –– not present in somatic cells of either parent –– can also occur in gametes, thereby giving the child a disease that had never before occurred in that genetic pedigree (“family tree”).

Traits transmitted as X-linked recessives were also described (e.g. red-green colorblindness, hemophilia –– due to mutations on the X chromosome). Female carriers have a 50% chance of passing the defective allele to offspring, thereby causing the disorder in males but not females.

Further, X-linked dominant traits were identified (e.g. incontinentia pigmenti, Coffin-Lowry syndrome –– also caused by mutations on the X chromosome). Affected females have a 50% chance of passing the mutant allele to all offspring, whereas affected males have a 100% chance of transmitting the defective allele to daughters. A single copy of the defective allele is sufficient to cause the disorder.

All the above-mentioned phenotypes follow bimodal distributions of Mendelian inheritance (Fig. 1A). Large-effect alleles can also result in a trimodal distribution (Fig. 1B), in which additive traits from both parents result in an intermediate phenotype. An additional caveat to virtually all Mendelian diseases includes “modifier genes”, i.e. any number of additional genes that can affect age of onset and/or degree of severity of the disease (Fig. 1C).

2.4. Biometric view of inheritance and reconciliation with Mendelian inheritance

In addition to these simple Mendelian traits having distinct patterns of inheritance, many traits such as body height and cognitive ability exhibit continuous variation in the population as well as strong resemblance within each family; however, the inheritance pattern of these continuous traits could not readily be explained by Mendelian laws. In the late 19th century and early 20th century, a group of statisticians (biometricians) –– including Francis Galdon, Karl Pearson and Raphael Weldon –– argued against Mendelian inheritance, insisting that most phenotypic variation was continuous, and that a “blending model” seemed more appropriate to explain the inheritance pattern of continuous traits, because an offspring's phenotype is roughly the average of that of the two parents.

The debate between supporters of Mendelian inheritance (“Mende-lians”) and the biometric school of inheritance lasted for two decades, until the disagreement was finally resolved by R.A. Fisher in 1918 in his publication of “The Correlation between Relatives on the Supposition of Mendelian Inheritance” (Fisher, 1918). In this monumental paper, Fisher showed that the “continuous variation” described by biometricians could represent the cumulative result of many discrete genetic loci (Fig. 1D), and that the family resemblance of continuous traits could be explained by Mendelian inheritance. Hence, it became appreciated that most human complex traits and diseases (e.g. height, body mass index, serum lipid levels, blood pressure, psoriasis, dementia, numerous types of cancer) represent multifactorial traits –– i.e. the result of contributions from dozens or hundreds, perhaps thousands, of genes (polygenic), combined with additional modifying effects such as contribution of epigenetic effects and other environmental factors.

2.5. Beginning of the genomics era

Subsequent to advances in molecular biology, recombinant-DNA cloning and DNA sequencing that began in the 1970s –– the field of genomics quickly helped us appreciate that “monogenic diseases” represent many more than one “disease allele.” Among the earliest breakthroughs was phenylketonuria (PKU), described as an autosomal recessive disorder caused by phenylalanine hydroxylase deficiency. After cloning the PAH gene from one chromosome of a “carrier” parent of a PKU child (Woo, Lidsky, Guttler, Chandra, & Robson, 1983), the critical PAH mutation represented a GT > AT transition at the 5′ splice donor-site of intron 12 (Marvit et al., 1987). At first, this discovery was hailed as “the disease allele” for PKU; however, within months, a second mutation (this one changing an amino acid) was reported. Three years later (Cotton, 1990), 18 distinct mutations had been identified, and the concept of allelic heterogeneity was widely accepted as the norm for most genes in which single mutations cause serious diseases.

Currently, 955 mutations (in and near the PAH gene), causing variable symptoms of PKU, have now been reported, worldwide [http://www.biopku.org/home/pah.asp], with ethnic differences in allelic frequencies. Similar stories can be told about many other Mendelian diseases.

2.6. Single-nucleotide polymorphisms (SNPs)

Following launch of the Human Genome Project in 1990, the field of genomics exploded with new knowledge. Yeast, worm and fly geneticists had been using the term “nucleotide substitutions” since the early 1980s, but in the mid-1990s several human genetics laboratories coined the term “single-nucleotide polymorphisms” (SNPs); “SNP fever” then began (“SNiPping through the DNA,” etc.). In retrospect, a better name for SNPs might have been “single-nucleotide variants”, but SNVs (“SNiVs”) doesn't sound as “catchy” as SNPs (“SNiPs”).

In the mid-1990s, many publications began to appear in top-tier journals, demonstrating “statistically significant” associations(P < 0.05) between one or a few SNPs and a complex disease –– such as hypertension, Alzheimer disease, asthma, or schizophrenia. Just as quickly, however, experiments to corroborate those data failed to do so, and it was soon realized that the genetics of complex diseases is not nearly as simple as that of Mendelian diseases.

Methods such as linkage studies that had been successful in finding major genes were found to have limited power for detecting genes of modest effect or lower penetrance; thus, a different method of association studies –– one that tested all candidate genes — would have greater power, even if this meant testing every gene in the genome. The landmark 2-page paper in 1996 by Risch and Merikangas (1996) proved to be visionary in this respect.

2.7. Recent rapid advances in our understanding of human genetics and genomics

Our appreciation of the concepts of human genetics and genomics has rapidly exploded during the past 25 years. We therefore believe that a brief review of past and present thinking in these fields, with regard to genome-wide association studies (GWAS), missing heritability and rare vs common variants, is necessary in order to acquaint the reader with the very complicated properties of multifactorial traits. As will be described, clearly the vast majority of drug responses (efficacy, adverse effect and toxic effect) involves –– not Mendelian or predominantly oligogenic –– but rather polygenic, multifactorial phenotypes.

2.7.1. Genome-wide association studies (GWAS)

The earliest GWAS was probably published in 2002, in which an association between the LTA gene and myocardial infarction was determined by typing almost 93,000 gene-based SNP markers (Ozaki et al., 2002). Another early GWAS –– typing > 116,000 SNPs –– reported an association between the CFH gene and age-related macular degeneration (Klein et al., 2005). More recently, easy-to-use DNA chips containing 1 million to 5 million SNPs have become readily available. The GWAS field has expanded exponentially; today, >24,000 SNP-trait associations have been reported in >2500 studies [https://www.ebi.ac.uk/gwas/]. These robust GWAS –– having P-values ranging from <10⁻⁸ to <10⁻⁴⁰⁰ –– underscore the value of using more stringent statistical significance levels when one is studying ≥ 1 million SNPs in large cohorts containing thousands, or even hundreds of thousands, of samples.

GWAS quickly became far more reliable for genotype-phenotype association tests –– when compared with studies involving one or several SNPs in small cohorts of several dozen or even several hundred individuals. These latter publications of (type-I and type-II error) artifacts have variously been called “the incidentalome” (Kohane, Masys, & Altman, 2006) and “the P < 0.05 false-positive studies” (Nebert, Zhang, & Vesell, 2008).

It should be recognized that multiple parameters (effect-size, allelic frequency, significance level, sample size) will all affect statistical power for any genotype-phenotype association study. Statistical power obviously improves with larger numbers of cases and controls. As any minor allele frequency (MAF) increases, fewer subjects are usually needed in the study group, and the level of detectable contribution by a genetic variant to a phenotype will be lower. If the MAF is low, greater numbers per group will be required, and the level of detectable contribution by a variant to a phenotype will be higher.

However, one cannot always embrace the above statements. For example, when the contribution (effect-size) of the risk variant is measured by heritability, or variance explained, influence of the MAF on statistical power of an association study is not significant. Nevertheless, if effect-size is measured by odds ratio (OR) or genotypic relative risk (GRR), the influence of a MAF will be greater –– especially for MAFs <0.05. In addition to the MAF and prevalence of a disorder, the inheritance model (“additive,” “dominant,” or “recessive”) of a risk allele can also influence power of an association study.

Virtually never do GWAS have sufficient statistical power to detect epistasis (gene × gene interactions; G × G) (Bhattacharjee et al., 2010; Sackton & Hartl, 2016) or gene-environment (G × E) interactions (D. Thomas, 2010). Moreover, an additional underappreciated class of genetic interactions is intergenic compound heterozygosity, i.e. interactions between multiple rare variants contributing to a trait (Gibson, 2011). Such interactions encompass heterozygous combinations of multiple alleles. In a broad sense, if hundreds of mutations –– each having a frequency of ≤0.1% –– all contribute to the phenotype, then such an event could contribute substantially to individuals that are heterozygous at many of these loci.

2.7.2. “Missing heritability”: real or imagined?

Twenty years after launching the Human Genome Project, initial GWAS findings became frustrating to those who wanted clear-cut explanations into the etiology of a complex disease (Goldstein, 2009). For most complex diseases, even multiple GWAS variants considered together (e.g. using polygenic risk score) typically explain too little variability in disease occurrence to be of much predictive value (Manolio, 2013; Wray et al., 2013). However, more importantly, some GWAS data had identified potential novel therapeutic targets for treating a complex disease –– without knowing its precise etiology. Similarly, some GWAS data might uncover potential therapeutic targets for treating an environmental disease, by learning something about its mode-of-action without necessarily understanding its precise mecha-nism-of-action.

Many GWAS often explain only a small proportion of heritability (defined as additive genetic variance). The absent proportion became known as “missing heritability” (Lander, 2011; Manolio et al., 2009), leading to renewed awareness about the genetic architecture of human complex disease and traits (Gibson, 2011; Zhang, 2015) –– a topic that had been extensively debated in the early 20th century.

First, it was realized that heritability attributable to some common variants could be substantial. However, as GWAS cohort sizes continued to increase –– thereby identifying additional variants contributing smaller effects to the trait –– the “revealed heritability” continued to grow, albeit rarely reaching >20–25% for various diseases and traits (Lander, 2011).

In addition, current GWAS may overlook many variants of lower frequency (MAFs 1–5%), because existing SNP-typing arrays often lack a more useful marker. Many complex disease-related alleles are probably included in this frequency class. Also, new genotyping arrays and imputation methods, based on the 1000 Genomes Project (The 1000 Genomes Project Consortium, 2012, 2015) or The Haplotype Reference Consortium (McCarthy et al., 2016), are able to capture these less frequent variants. This topic is discussed in detail later.

GWAS also miss many common small-effect variants –– due to limited sample-size and/or stringent statistical thresholds imposed to ensure reproducibility. Efforts seeking to infer contributions of loci that fall just short of statistical significance were then addressed (Park et al., 2010; Yang et al., 2010). Contributions of loci that fall even further short of statistical significance, likewise, will result in even smaller effect-sizes on phenotype. Although their individual contributions may be too small ever to detect by investigators designing feasible sample cohort sizes, these very-small-effect variants collectively will probably also explain a significant fraction of heritability (Gibson, 2010; Lander, 2011).

Furthermore, rare variants of large-effect will sometimes contribute substantially to common diseases, although their roles are just recently being explored. Increases in our understanding of rare variants have now advanced –– via whole-exome sequencing (WES) (Bertier, Hetu, & Joly, 2016) and whole-genome sequencing (WGS); also called “next-generation” sequencing (NGS) (Cirulli & Goldstein, 2010; Pinto, Ariani, Bianciardi, Daga, & Renieri, 2016).

Whether rare variants will lead to discovery of a substantial number of new genes –– is a question that perhaps can be answered by systematic WES or WGS. Given the background rate of rare variants, many thousands of samples will be required to achieve statistical significance. Correspondingly, how to quantify total heritability due to rare variants remains unclear. Although the inferred effect-sizes are larger, overall contribution to heritability may be small because of their low frequencies (Gibson, 2011; Zhang, 2015).

Lastly, some “missing heritability” might be purely an illusion (Lander, 2011) because heritability is estimated from epidemiological data by applying principles for inferring additive genetic effects. These approximations may be overestimated –– due to methods that are not effective at excluding nonlinear contributions of G × G interactions, G × E interactions, or epistasis –– which are likely to be important.

2.7.3. Rare versus common variants

Numerous explanations for “missing heritability” have been postulated (Gibson, 2011), including: the ‘infinitesimal model’ (large number of variants across the entire allele frequency spectrum of small effects); the ‘rare variant model’ (multiple large-effect rare variants that are poorly tagged by genotyping arrays); and the ‘broad-sense heritability model’ (contributions from G × G, G × E, and/or epigenetic interactions).

The infinitesimal model is a classic model for polygenic traits –– first introduced by Fisher (1918). This model assumes a very large number of genetic loci, each with an infinitesimally small-effect. Those loci detected by GWAS are simply the largest effect-sizes (i.e. “low hanging fruits”). If a half dozen common variants explain 10% of the phenotype in a population, then the remaining 90% reflects innumerable variants in which each contributes very small effect on the phenotype (i.e. explaining ≪ 1% variance) (Gibson, 2011).

Ultimately, every very-small-effect variant will contribute to every trait –– but, with effect-sizes so small that it might take samples greater than the population size of the cohort, or even the species size on the planet, to detect! Examples include the massive meta-analyses of GWAS for height (Lango Allen et al., 2010; Wood et al., 2014) and body mass index (BMI) (Locke et al., 2015; Speliotes et al., 2010) in each of these cases, involving several hundred thousand people. In these studies, it has been shown that several hundred loci can contribute to some multifactorial traits; still, these loci often will not necessarily explain even one-fourth of all genetic variance. For example, using genome-wide data from > 253,000 individuals, 697 variants were identified at genome-wide significance, which, together, explain ~20% of the heritability for adult height (Wood et al., 2014). Therefore, the infinitesimal model suggests that heritability is not so much “missing”, but simply “hidden” beneath statistically-significance thresholds (Gibson, 2010). The largest-yet GWAS on human height has recently appeared (Marouli et al., 2017) in which almost 459,000 individuals, followed by an independent set of more than 252,000 participants, were studied using ExomeChip (Grove et al., 2013) to test low-frequency coding variants association. In addition to the previously reported ~700 common variants, 83 additional height-associated coding variants were found –– having lower MAFs in the range of 0.1% to 4.8%. These data confirm further that increasingly larger sample sizes will continue to uncover additional rare and low-frequency SNPs having moderate-to-large effects associated with polygenic human traits. However, these newly identified rare and low-frequency variants only explained an additional small fraction (1.7%) of the heritability of height.

The rare variant model emphasizes contributions of multiple rare (MAFs <0.5%) or low-frequency (MAFs <5%) –– as long as they are large-effect variants (Pritchard, 2001). In this model, expressivity can be modified by other loci or by the environment (Bodmer & Bonilla, 2008), but the idea is that the rare-susceptibility genotype is largely responsible for the trait. The rare-allele model would generally refer to dominant effects –– caused either by haploinsufficiency or gain-of-function alleles. This would elevate risk > 2-fold above background; under these conditions, penetrance need not be anywhere near 100%. In fact, the vast majority of unaffected individuals would be expected to carry one or more risk alleles. Thus, the notion is that a multifactorial trait is actually a combination of hundreds, even thousands, of similar but genetically heterogeneous conditions –– attributable to rare variants at individual loci. If each of these variants explains most of the phenotype in just a few people, their effects would be insufficient to explain variance in a total population, detectable by the usual GWAS. The total number of loci that may contribute to any multifactorial trait would be a function of the baseline incidence, number of rare variants per locus, and GRR (i.e. effect-size).

The broad-sense heritability model suggests that additive contributions, of common variants and large-effects of rare variants, are still insufficient to explain “missing heritability.” There are numerous studies detecting G × G interactions clinically and G × G interactions in model-organism quantitative genetic research. Moreover, an increasing number of studies have described epigenetic effects (Slatkin, 2009), parent-of-origin genetic contributions (Connolly & Heron, 2015), and inheritance of DNA-methylation patterns (Haque, Nilsson, Holder, & Skinner, 2016). The hypothesis of this third model is that, because GWAS measure only average effects of alleles across thousands of individuals, GWAS are incapable of capturing heterogeneity of effect-sizes at the family or the individual level –– which would be a guarantee of the proposed broader components of genetic architecture. It cannot be ruled out that broad-sense heritability is likely to contribute some indeterminate amount of heritability to multifactorial traits.

3. Genetic studies of PGx traits

“Genetic differences in drug response” have long been recognized. Some suggest that Pythagoras –– in southern Italy, about 510 BCE –– was the first to recognize “dangers of some individuals, but not others, who eat fava beans” [reviewed in (Nebert, 1999)]; the adverse reaction in this case is hemolytic anemia due to glucose-6-phosphate dehydrogenase (G6PD) deficiency. During the 1940s, genetic differences in G6PD protein stability and function were discovered to be the cause [reviewed in (Motulsky, 1964)]. Others propose that Snyder's original study in 1932 –– in which he described the “phenylthiourea nontaster” phenotype inherited as an autosomal recessive trait (Snyder, 1932) –– represents the dawn of modern pharmacogenetics.

3.1. Early-day PGx examples

3.1.1. CYP2D6 gene polymorphism

The first PGx trait –– elucidated by recombinant DNA methodology –– was the debrisoquine/sparteine polymorphism (Gonzalez et al., 1988); the defective-enzyme disorder had been discovered independently in England (Mahgoub, Idle, Dring, Lancaster, & Smith, 1977) and Germany (Eichelbaum, Spannbrucker, Steincke, & Dengler, 1979). Dominant, additive, and recessive phenotypes were named “extensive-metabolizer” (EM), “intermediate-metabolizer” (IM), and “poor-metabolizer” (PM), respectively. Cloning the CYP2D6 gene and identifying several variant alleles (Gonzalez et al., 1988) was the first time a polyallelic mechanism could be explained in molecular terms as a quantitative PGx trait. PM recessive alleles might encode an inactive CYP2D6 enzyme, unstable protein, incorrect splicing of the gene transcript, or even complete deletion of the gene; hence, clinical phenotypes resulted in lowered, or completely absent, CYP2D6 activity.

Subsequently, an “ultra-metabolizer” (UM) phenotype was discovered, caused by two to 13 CYP2D6 duplicated genes (Bertilsson, Dahl, Dalen, & Al-Shurbaji, 2002). Thus, distribution of N individuals as a function of CYP2D6 phenotype is not simply bimodal or trimodal; it begins to look like a gradient with perhaps four distinct peaks (Fig. 2). Moreover, striking ethnic differences have been found among CYP2D6 alleles responsible for PM, IM, EM and UM phenotypes (Teh & Bertilsson, 2012) –– suggesting a genetic bottleneck perhaps triggered by some sort of environmental selective pressure.

Fig. 2.

Distribution of measurable CYP2D6 phenotypes. Four phenotypic groups can be distinguished: ultra-rapid (UM), extensive, sometimes called “efficient” (EM), intermediate (IM), and poor (PM) metabolizers [modified from The Lancet 2000; 356: 1667–71].

Eliglustat is a potent selective inhibitor of glucosylceramide synthase, important for the biosynthesis of glucosylceramides that accumulate in Gaucher disease. Because CYP2D6 predominantly metabolizes this drug –– CYP3A to a lesser extent –– the patients' CYP2D6 metabolizer status, and use of concomitant drugs, must be carefully considered (Balwani et al., 2016; Belmatoug et al., 2016).

The “CYP2D6 panel” now comprises at least 25% of all commonly-prescribed drugs [(Lu, Wang, & Lin, 2003); http://medicine.iupui.edu/clinpharm/ddis/]. In the vast majority of cases, deficient CYP2D6 causes an adverse effect –– the parent drug reaching toxic levels due to insufficient metabolism in PM patients. The opposite occurs in the case of opiates such as codeine, however. The high-CYP2D6-activity phenotype patient is at greater risk of an adverse effect, because the inactive parent drug requires CYP2D6 metabolism to form the active moiety, morphine (Linares, Fudin, Schiesser, Daly Linares, & Boston, 2015).

3.1.2. N-acetylator phenotype

Originally termed “isoniazid acetylation polymorphism”, this unusual drug response was identified clinically in the late 1940s when subjects –– who had converted to a positive tuberculin test –– were routinely treated with isoniazid. The clinical presentation was an unexpectedly high incidence of peripheral neuropathy. Using a 6-h plasma-level timepoint after an isoniazid standard dose, a bimodal distribution was reported (Price Evans, Manley, & McKusick, 1960). This was the first example of clearance of a parent drug that is genetically determined –– i.e. the major enzyme that metabolizes isoniazid (N-acetyl-transferase) is defective in efficient removal of the parent drug. As the result of “slow” vs “rapid” plasma clearance, individuals were phenotyped as “slow acetylators” (r/r) or “rapid acetylators” (R/R, R/r); the autosomal recessive inheritance is consistent with the earlier findings of Mendel and Garrod. Three decades later (Blum, Grant, McBride, Heim, & Meyer, 1990), the NAT2 gene, and not NAT1, was shown to be responsible for Mendelian inheritance of isoniazid N-acetylation.

The NAT2 polymorphism is not only relevant to isoniazid, but also to dozens of other drug and environmental toxicant substrates of NAT2 (Grant, Goodfellow, Sugamori, & Durette, 2000; Walraven, Zang, Trent, & Hein, 2008). As is the case for virtually every drug-response gene due to the great human diasporas (Nebert & Dieter, 2000), striking inter-ethnic differences in frequency of NAT2 alleles were discovered, e.g. among Japanese the PM phenotype subset is <10%, whereas among Caucasians PMs comprise >70% [reviewed in (Nebert, 1999)].

3.1.3. CYP2C19 gene polymorphism

CYP2C19 participates in metabolism of at least four dozen commonly-prescribed drugs [http://medicine.iupui.edu/clinpharm/ddis/]. The story is similar to the two above-described examples: the EM phenotype appears to be dominant over the PM phenotype, and almost always the parent drug in PMs exhibits ADRs due to overdose. Ethnic differences are very striking for CYP2C19 alleles; e.g. the East Asian PM subset is ~33% and Oceanian PM subset > 50%, whereas the Caucasian PM subset represents <6% (Mizutani, 2003).

CYP2C19 catalyzes bioactivation of clopidogrel, an anti-platelet prodrug. Loss-of-function alleles (e.g. CYP2C19*2) impair formation of active metabolites, resulting in decreased platelet inhibition and increased risk for adverse cardiovascular events, especially in patients undergoing percutaneous coronary intervention (PCI) (Johnson et al., 2012). Therefore, alternative anti-platelet therapy (e.g. prasugrel, ticagrelor) is now recommended for CYP2C19 PM or IM patients –– if there is no contraindication (Scott et al., 2013).

3.1.4. Abacavir-induced hypersensitivity

In an early example of identification of a PGx disorder by a candidate-region study, 185 patients treated with abacavir –– an HIV-1 nu-cleoside-analog reverse-transcriptase inhibitor –– resulted in a life-threatening hypersensitivity syndrome in 18 patients (Mallal, Nolan, et al., 2008). SNP-typing of loci in the major histocompatibility complex (MHC) region revealed a very strong association with the HLA-B*57:01 allele (OR = 117) and also with the combination of HLA-DR7 and HLA-DQ3 loci (OR = 73), when compared with 167 abacavir-tolerant controls. A high association of abacavir-induced hypersensitivity with HLA-B*57:01 was later confirmed in an elegant double-blind prospective, randomized study involving almost 2000 patients from 19 countries (Mallal, Phillips, et al., 2008).

3.1.5. Coumarin-associated gene polymorphisms

Optimizing warfarin, coumarin and acenocoumarol dosage for anticoagulation is an oligogenic example of a substantial contribution of several genes –– resulting in rapid vs slow PK and/or PD of coumarins. CYP2C9 (J. A. Goldstein & de Morais, 1994), VKORC1 (D'Andrea et al., 2005), and CYP4F2 (Caldwell et al., 2008) polymorphisms were independently discovered in candidate-gene studies and then later confirmed in GWAS (Cooper et al., 2008; Takeuchi et al., 2009). The combination of mutations in the CYP2C9, CYP4F2 and VKORC1 genes explains as much as 35–50% of the patient's total variation in drug response (Daly, 2009). This means the remaining genetic impact on coumarin response is derived by contributions of genes other than these three major genes.

3.1.6. Further examples of PGx traits

The above-described NAT2, CYP2D6, CYP2C19, CYP2C9 and CYP4F2 polymorphisms represent large-effect PK genes encoding enzymes involved in metabolism.

On the other hand, HLA-B*57:01 represents an immune-response gene, whereas VKORC1 and IL28B are considered to be large-effect PD genes. HLA-B*57:01 encodes a particular cell-surface molecule responsible for presentation of endogenous peptides to immune system cells. VKORC1 codes for vitamin K-epoxide reductase complex subunit-1 that is targeted directly by coumarins. Considered as vitamin K antagonists (Jackson & Suttie, 1977), coumarins are potent inhibitors of this reductase complex, leading to depletion of reduced vitamin K –– essential for normal coagulation. This epoxide reductase is therefore regarded as a “drug target” for coumarins, and, except for the vitamin K substrate, does not appear to metabolize any other drug.

Three additional large-effect PK genes showing substantial quantitative clinical differences in drug response include the thiopurine S-meth-yltransferase gene (TPMT), UDP glucuronosyltransferase-1A1 gene (UGT1A1), and dihydropyrimidine dehydrogenase gene (DPYD).

TPMT encodes an enzyme of critical importance in thiopurine drugs and therefore treatment of childhood leukemia (Eichelbaum, Ingelman-Sundberg, & Evans, 2006; Weinshilboum & Sladek, 1980). Of interest, a recent GWAS of > 1000 children with leukemia found TPMT as the only gene achieving genome-wide significance –– top hit was rs1142345; 719A > G; P = 8.6 × 10⁻⁶¹ (Liu et al., 2016), thereby supporting the idea that differences in TPMT activity might reflect a Mendelian, or predominantly oligogenic, trait. These data further confirm the clinical usefulness of the red-cell TPMT enzyme assay, which is a phenotype (not genotype) test, because discovery of a completely new DNA variant that causes low TPMT enzyme activity is always possible.

UGT1A1 encodes the enzyme UGT1A1 that metabolizes, among many other drugs, irinotecan. Poor-metabolizer UGT1A1*28 homozygosity is strongly associated with irinotecan-induced neutropenia; it is recommended –– on the irinotecan product label for UGT1A1*28 homo-zygotes –– to decrease the starting dose (Kweekel, Guchelaar, & Gelderblom, 2008). Due to lack of sufficient prospective data, however, it remains uncertain whether dose-reduction leads to decreased toxicity or altered antitumor effect. Combined toxicity analysis indicates that most patients experiencing grade 3–4 diarrhea and/or neutropenia are not homozygous for UGT1A1*28.

Frequently prescribed anticancer drugs, the fluoropyrimidines, are inactivated by hepatic dihydropyrimidine dehydrogenase (DPYD). As much as 5% of cancer populations exhibit DPYD-deficiency, which is considered having practical value (and even cost-effective) gene product for its anticancer drug substrates; DPYD*2A is strongly associated with fluorouracil (5-FU) induced severe and life-threatening toxicity (Deenen et al., 2016). There is convincing evidence to implement prospective DPYD genotyping with an up-front dose adjustment in DPYD-deficient patients (Lunenburg et al., 2016). In each of these examples, if the PGx assay reveals diminished enzymatic activity or deficiency, lower doses or alternative drugs are usually recommended.

3.2. PGx genome-wide association studies (GWAS)

Advances in genomic technology –– such as microarray-based genotyping platforms and NGS technologies, along with development of statistical packages for large-scale data analysis –– have enabled researchers to perform genome-wide searches for novel and less obvious genes associated with drug responses. Obviously, it is more difficult to identify a large number of individuals being treated with a specific drug (Daly, 2010), compared with, for example, identifying a large number of individuals for height, blood pressure, or BMI. Since 2007, however, dozens of PGx GWAS with sufficient numbers in the cohorts have identified many genetic variants associated with various drug responses and adverse effects (Table 1). GWAS are hypothesis-free and thus do not require a priori assumptions about genomic location of functional variants (McCarthy et al., 2008; Stranger, Stahl, & Raj, 2011); therefore, GWAS provide an unbiased and powerful tool for systematic discovery of genetic variants associated with PGx traits (Motsinger-Reif et al., 2013).

Table 1.

List of selected PGx GWAS with genome-wide statistically significant findings (P < 5e–8).

Drug	Response	Study sample size		Associated genes and variants					Reference
		Discovery	Replication	Genes	Variant type	Freqb	Effect	P-value	PMID	Year
ADR (resembles Mendelian traits with incomplete penetrance)
Statin	Myopathy	85/90	20,000	SLCO1B1	Non-synonymous	0.15	OR = 4.5	4.00e–09	18650507	2008
Flucloxacillin	DILIa	51/282	23 cases	HLA-B*57:01	HLA haplotype	0.05	OR = 80.6	8.70e–33	19483685	2009
Lumiracoxib	DILI	41/176	98/405	HLA-DRB1*15:01	HLA haplotype	0.1	OR = 5.0	6.80e–25	20639878	2010
Carbamazepine	Hypersensitivity	22 + 43/3987	145	HLA-DRB*15:02	HLA haplotype	0.02–0.05	OR = 12.4	3.50e–08	21428769	2011
Heparin	Thrombo-cytopenia	67/884	94/178	TDAG8	Missense?	0.22	OR = 18.6d	3.18e–09	25503805	2015
Anthracycline	Cardiotoxicity	280	96 + 80	RARG	Missense	0.25	OR = 4.7	5.90e–08	26237429	2015
Lapatinib	DILI	1192 (34/810)		HLA-DRB1*07:01	HLA haplotype	0.2	OR=14d	7.8e–11	25987243	2016
Asparaginase	Hypersensitivity	3308(589)		NFATC2	Intronic	0.08	OR = 3.11	4.10e–08	25987655	2015
Cisplatin	Hearing loss	238	68	ACYP2	Intronic	0.07	HR = 4.5	3.90e–08	25665007	2015
Dosage/efficacy (having major gene effect)
Warfarin	Dosage	181	374	VKORC1 and CYP2C9	Multiplec		34%	6.20e–13	18535201	2008
Warfarin	Dosage	1053		VKORC1, CYP2C9 and CYP4F2	Multiple		43.1%	<1e–78	19300499	2009
Acenocoumarol	Dosage	1451	287	VKORC1, CYP2C9, CYP4F2 and CYP2C18	Multiple		48.80%	<5e–8	19578179	2009
PegIFN-alpha	Viral clearance	1137 + 1475		IL28B	Upstream	0.3	OR = 7.3	1.00e–38	19684573	2009
		293	555	IL28B		0.27	OR = 1.98	9.30e–09	19749758	2009
		142	172	IL28B		0.15	OR = 27.4	2.70e–32	19749757	2009
Citalopram and escitalopram	Concentration	435		CYP2C19 and CYP2D6	Multiple		NA	<5e–8	24528284	2014
Complex PGx traits (influenced by multiple genetic variants each with small effect)
Statin	LDL response	18,596	22,318	SORT1/CELSR2/PSRC1,SLCO1B1, APOE, LPA	Multiple		1.3%–5.2%e	<5e–8	25350695	2014
Metformin	HbA1c <7%	1024	1783 + 1113	ATM	Intronic	0.44	OR = 1.35	2.90e–09	21186350	2011
		10,577		SLC2A2	Intronic	0.29	0.17% greater	6.60e–14	27500523	2016
Antidepressant	Questionnaire, physician's opinion	865	1529	HPRTP4	Intergenic	0.4	OR = 1.36	5.03e–08	25897834	2015

^aDILI, drug-induced liver injury.

^bFreq, minor allele frequency (MAF) of the most significant or putative functional variant.

^cMultiple associated variants (probably due to allelic heterogeneity); variant types and frequencies are not shown.

^dOR of homozygotes.

^ePercentage of extra LDL-C lowering in carriers versus non-carriers of the SNP.

3.2.1. GWAS of ADRs

In the past few years, several GWAS have detected novel genetic loci associated with severe ADRs (Daly, 2012). Probably the first successful example is identification of a genetic risk variant for myopathy in patients treated with simvastatin (Link et al., 2008). Authors first screened ~300,000 markers in 85 cases and 90 controls selected from a clinical trial of ~12,000 participants, and then replicated the finding in ~ 20,000 subjects; they identified a nonsynonymous variant (rs4149056) in the SLCO1B1 gene that confers an OR of 16.9 in the homozygous state (Table 1). Collectively, >60% of myopathy cases, receiving a high simvastatin dose (80 mg), can be attributed to the risk allele. The SLCOB1 gene codes for an organic anion transporter, suggesting that it functions in cellular uptake of statins (König, Seithel, Gradhand, & Fromm, 2006); however, this hypothesis was notfirmly established, according to previous candidate-gene association studies (Morimoto et al., 2005). Subsequently, the correlation was reproduced in further clinical studies (Voora et al., 2009), and simvastatin dosing, based on the SLCO1B1 genotype, has been recommended by the Clinical Pharmacogenomics Implementation Consortium (CPIC) (Ramsey et al., 2014). Intriguingly, association of the SLCO1B1 rs4149056 variant with myopathy has been less compelling for other statins (Niemi, 2010).

In another PGx GWAS (Daly et al., 2009), the HLA-B*57:01 allele was identified as a major risk factor for flucloxacillin-induced liver injury –– a severe, but quite rare, ADR (~8.5 in every 100,000). In a small cohort (51 cases, 282 controls), authors identified a strong association at a SNP marker in complete linkage disequilibrium (LD) with HLA-B*57:01 having an OR =80.6 (Table 1). However, despite the strong association, only one in every 500 to1000 flucloxacillin-treated individuals with this allele will develop liver injury, thereby limiting clinical utility (i.e. the test has a very high false-positive rate). This is an example in which incomplete penetrance makes it difficult for the physician or pharmacologist, who must make clinical decisions on a patient-by-patient basis.

Following this study, more GWAS have demonstrated strong associations between several HLA haplotypes and various ADRs (Table 1) –– including lumiracoxib-induced liver injury (Singer et al., 2010), carbamazepine-induced hypersensitivity (McCormack et al., 2011), and lapatinib-induced liver injury (Parham et al., 2016). However, predictive values for typing each of these HLA alleles associated with these ADRs, unfortunately, are very low (Daly, 2014).

Other than the HLA region, PGx GWAS have also identified a number of associations with various ADRs (Fernandez et al., 2015; Karnes et al., 2015; Xu et al., 2015) –– which would not have been suspected, based on established knowledge about each drug's PK and/or PD (Table 1). Although these discoveries may not have immediate clinical utility, they point to new biological pathways underlying each drug's mode-of-action and/or toxicity.

Unfortunately, not all GWAS of ADRs have yielded robust associations. For example, many GWAS of drug-induced QT-interval prolongation did not produce consistent results [recently reviewed in (Niemeijer, van den Berg, Eijgelsheim, Rijnbeek, & Stricker, 2015)]. Osteonecrosis of the jaw (ONJ) is a severe ADR, associated with chronic bisphosphonate treatment; initial GWAS (Sarasquete et al., 2008) revealed a correlation between ONJ and CYP2C8 variants, and this P450 enzyme is a plausible candidate for having a significant role in metabolism of several antican-cer drugs; however, this finding has not been replicated in subsequent studies (English et al., 2010; Katz et al., 2011; Such et al., 2011). ADRs such as ONJ are well known to be extremely heterogeneous phenotypes, similar to many other complex diseases; therefore, unless one has a very large sample-size, loci that can be replicated will continue to be difficult to identify with certainty.

3.2.2. GWAS examples of drug efficacy and dosage

Warfarin is a widely used anticoagulant but its usage is complicated, due to wide inter-individual variability in dosing regimen and its narrow therapeutic “window”; this makes warfarin dosage a very active area of PGx studies. Candidate-gene association and molecular studies –– well before the genomic era –– had already revealed significant roles of two important genes (CYP2C9 and VKORC1; vide supra). Multiple GWAS have now confirmed these known associations (Cooper et al., 2008; Takeuchi et al., 2009) and provided more accurate estimates of their effects in determining the observed variance (VKORC1 = ~30% and CYP2C9 = ~12%). Furthermore, a GWAS with larger sample-size (Takeuchi et al., 2009) also discovered an additional locus with relatively small-effect (CYP4F2 = 1.1%); this CYP4F2 polymorphism, however, has not been included in most dosing algorithms (Johnson et al., 2011; Pirmohamed, Kamali, Daly, & Wadelius, 2015). GWAS have also identified similar genetic determinants involving other coumarin anticoagulants (Teichert et al., 2009), as well as in several different ethnic groups (Cha et al.,2010; Parra et al., 2015; Perera et al., 2013).

Another early successful example of GWAS of drug efficacy came from studies of variable response to interferon-α treatment of hepatitis C virus (HCV) infection. In 2009, three independent studies (Ge et al., 2009; Suppiah et al., 2009; Tanaka et al., 2009) reported genetic variants near the IL28B gene associated with interferon-α treatment responses; the effect is measured by sustained virological response (SVR; i.e. absence of detectable virus at end of follow-up). The same variant was also reported to be correlated with spontaneous clearance of HCV (Thomas et al., 2009), suggesting the variant might influence host immune response –– with or without drug intervention.

Also relevant, the advantageous allele (C allele at rs12979860) is more frequent in Caucasian and Asian populations than in Africans –– a difference that very likely explains the previous report (Conjeevaram et al., 2006) of relatively poor outcomes among HCV-in-fected African Americans in trials of interferon-α treatment. Despite the large estimated effect-size (OR = 7.3), however, the imperfect predictive power, combined with lack of alternative treatments, had limited immediate clinical utilization of determining the IL28B genotype in personalized treatment decisions (Iadonato & Katze, 2009).

The new direct antiviral agents (DAAs), in association with the pegylated interferon/ribavirin combination, is the new standard of HCV treatment making it necessary to re-evaluate the clinical utility of the IL28B genotype (Thompson & McHutchison, 2012). Although IL28B genotype testing continues to provide important clinical information regarding expected treatment response, IL28B genotyping will eventually lose its utility –– as HCV therapy moves away from interferon-based regimens (Jensen & Pol, 2012; Liapakis & Jesudian, 2012).

3.2.3. GWAS of more complex PGx traits

GWAS have also been used to study efficacy of additional widely used medications for common complex disorders –– such as type-2 diabetes (T2D), dyslipidemia, hypertension, and psychiatric disorders. The GoDarts Group (GoDarts et al., 2011) conducted GWAS for glycemic response to metformin in >1000 individuals with T2D and identified variants in the ataxia-telangiectasia-mutated gene (ATM) that were associated with treatment success; the association was replicated in an independent cohort with almost 1800 samples. The combined effect was 1.35 times higher for reaching the treatment goal (HbA1c <7%) for each minor allele (C) of rs11212617, at which the strongest association was observed.

Consequently, this association was replicated in multiple additional cohorts (van et al., 2012); however, this variant has not modified the metformin response in the Diabetes Prevention Program (DPP) (Florez et al., 2012). The rs11212617 variant thus appears to have no significant impact on the predictive value with regard to metformin action on gly-cemic control. Nevertheless, for reasons not clear, several observations appear to suggest a substantial role of ATM in glucose homeostasis. For example, ataxia telangiectasia patients can manifest a severe form of diabetes, and Atm(−/−) knockout mice exhibit insulin resistance and abnormal adipose distribution (Takagi et al., 2015). More recently, in a large GWAS consisting of ~13,000 participants, the Metformin Genetics (MetGen) Consortium reported an association between an SLC2A2 (facilitated glucose transporter) variant and metformin treatment response (Zhou et al., 2016).

Treatment response to statins has also been examined by multiple GWAS (Chasman et al., 2012; Deshmukh et al., 2012; Hopewell et al., 2013; Postmus et al., 2014) [reviewed in (Leusink, Onland-Moret, de Bakker, de, & AH, 2016)], with large sample-sizes ranging from ~2000 to >40,000 samples. These well-powered studies have identified common SNPs in five genes (LPA,APOE,SLCO1B1,SORT1 and ABCG2) ––albeit with modest effect-sizes. These variants are robustly associated with LDL-C response; however, none consistently modifies risk reduction for cardiovascular events, and therefore these findings have limited clinical utility.

3.2.4. Unsuccessful examples

Hypertension is the most prevalent modifiable risk factor for global disease burden, and many PGx studies have attempted to identify genetic variants responsible for the highly varying responses to one or another anti-hypertensive medication. Despite extreme efforts, these studies have yet to generate robust independently replicated findings (Lupoli, Salvi, Barcella, & Barlassina, 2015; Menni, 2015); such failures likely reflect the extreme heterogeneity and complexity of hypertensive disease. Elevated blood pressure has multiple etiologies and involves varying physiological changes throughout one's lifetime. Identified GWAS variants associated with blood pressure are far fewer than those associated with blood glucose or blood lipid levels, and identified blood pressure variants also explain less phenotype variance (<3%) (Munroe, Barnes, & Caulfield, 2013) –– which is probably further evidence consistent with a high degree of complexity for hypertensive disorders.

Moreover, the multiple classes of antihypertensive drugs (diuretics, ACE inhibitors, angiotensin II receptor blockers, calcium channel blockers, beta-blockers) –– alone, and in combinations –– further complicate “clean” PGx studies, not to mention variance due to within-person fluctuations of blood pressure (e.g. changes in diet, exercise, time-of-day, combinations of prescribed drugs, alcohol usage, etc.). Several PGx GWAS have reported variants in plausible genes associated with thiazide diuretics (Chittani et al., 2015; Turner et al., 2008; Turner et al., 2013), angiotensin II receptor blockers (Frau et al., 2014; Turner et al., 2012), and beta-blockers (Gong et al., 2015; Gong et al., 2016); yet, most of these studies involve relatively small discovery sample-sizes and are underpowered, as far as detecting associations having genome-wide significance (i.e. P < 5 × 10⁻⁸). Large international consortia have been established –– aimed at increasing opportunities for discovery and replication by assembling large samples (Cooper-DeHoff & Johnson, 2016). However, even though large consortium studies might be able to identify genetic-variant associations in the future, it is hard to imagine that, even in combination, any of such newly identified small-effect variants will have important clinical predictive power.

Similarly, GWAS of antidepressant treatment response (Biernacka et al., 2015; Garriock et al., 2010; Ising et al., 2009; Uher et al., 2010) have not yielded promising results. One particular challenge in PGx studies of psychotropic medications is the evaluation of disease conditions and quantitative measurements of treatment responses, which –– in contrast to blood pressure or serum cholesterol levels –– are all based on multidimensional psychiatric diagnostic criteria. This method has long been criticized for its lack of reliability (e.g. degree of consensus on a diagnosis or treatment response as viewed by different clinicians; the wording and how questionnaires are answered by patients) [reviewed in (Aboraya, Rankin, France, El-Missiry, & John, 2006)]. This complexity, comparing the “equivocal phenotype” vs the “unequivocal phenotype”, has previously been emphasized (Nebert et al., 2008).

The same is especially true for major depressive disorder (MDD) (Freedman et al., 2013). Clinically-based psychiatric criteria do not directly reflect biologically homogeneous conditions –– which makes phenotypic definition of psychiatric conditions highly heterogeneous and measurement of treatment responses very inaccurate. At least in part due to these difficulties, GWAS of MDD with reasonably large sample size (>9000 cases) have failed to identify any replicable association (Ripke et al., 2013). In contrast, a PGx study, evaluating parent drug (citalopram and escitalopram) and metabolite plasma concentrations in 435 MDD patients, successfully identified highly significant associations with CYP2C19 and the CYP2D6 variants (Ji et al., 2014). This important distinction –– between diffuse psychiatric clinical phenotypes and quantitative measurements of blood or urine drug- or metabolite-level phenotypes –– has previously been emphasized (Nebert et al., 2008). Therefore, it becomes problematic to discern how these pharmacometabolomic association studies [e.g. (Ji et al., 2014)] will aid in genetic risk prediction of many complex clinical responses.

4. Genetic architecture of PGx traits

Genotype-phenotype association studies, especially those recently described PGx GWAS (vide supra) have identified many genetic variants correlated with numerous drug-response phenotypes. These results have opened a window into the “genetic architecture of PGx traits.” Genetic architecture defines how genotypes map to phenotypes (cf. Glossary of Terms); specifically, it includes descriptions about the number of traits-associated variants, their allele frequencies and effect-size distribution, combined with patterns of pleiotropy and epistasis (Zhang, 2015). Knowledge about genetic architecture of PGx traits is not only important to our understanding the biology, but also is instrumental in PGx study-design and analysis. Genetic architecture primarily defines potential predictive power of PGx response using genomic data (Chhibber et al., 2014; Zhou & Pearson, 2013). In what follows, we have classified PGx traits into three broad categories (Mendelian, predominantly oligogenic, and complex) –– based on their genetic architecture –– and we discuss the implications in study design and clinical utility.

4.1. Monogenic (Mendelian) PGx traits

4.1.1. PGx response

Many (ordinarily rare) severe ADRs are usually expressed as binary traits (i.e. “present”, or “not present”). Occurrence of such rare/severe reactions depends almost exclusively on the presence of causative PGx variants (or mutations), whereas the ADR is less influenced by dosage in the therapeutic range. This class can be further categorized into three subgroups, based on molecular etiology:

1. Severe ADRs due to a major defect in drug metabolism or transporter gene (PK gene). Examples include statin-induced myopathy (SLCO1B1) and slow- vs rapid-acetylator phenotype (NAT2).

2. Dose-independent (idiosyncratic) ADRs due to inappropriate immune response –– including severe drug-induced liver damage and hypersensitivity reactions (usually affecting skin) (Daly, 2013).

3. Genetic mutations in genes involved in mechanism-of-action of a therapeutic drug. For example, hereditary predisposition for venous or arterial thrombosis, such as mutations in F5, PROC and PROS1, have been associated warfarin-induced skin necrosis (Yang & Algazy, 1999).

4.1.2. Genetic basis

The genetic basis of Mendelian PGx traits usually involves nonsynonymous mutations in PK genes (e.g. CYP2D6, SLCO1B1 or NAT2) or HLA genes (Daly, 2013). Because of the discrete nature of this type of binary PGx response and relatively large-effect of the causative genetic variant(s), consequence of genetic changes can be efficiently measured by OR or penetrance.

It should be emphasized that –– even this simplest form of PGx trait is not simple, for two reasons. First, multiple mutations in a gene often influence gene function to varying degrees, e.g. to date, there are >180 different alleles reported for the CYP2D6 gene [The Human Cytochrome P450 (CYP) Allele Nomenclature Database; http://www.cypalleles.ki.se/], which can cause a gradient (from mild to severe) of CYP2D6 enzyme deficiency. The same can be said for SLCO1B1 or NAT2. Second, although presenting with high OR, risk alleles may not always cause the ADR; as described earlier, this is an example of incomplete penetrance. For instance, despite the strong association (OR = 80) of flucloxacillin-in-duced toxicity with the HLA-B*57:01 allele, only one in every ~500 to ~1000 individuals with this genotype, when treated with flucloxacillin, will develop drug-induced liver injury, (Daly et al., 2009).

4.1.3. Genetic study

Given the high OR (or penetrance), case/control studies with even small numbers of cases can be an efficient approach to identification of genetic variants responsible for this type of PGx trait. Especially for those idiosyncratic cases in which we have no plausible a priori biological hypothesis, GWAS with a small number of case/control samples can be a powerful approach to identifying causative variants (Daly, 2012). However, given the low occurrence of such rare ADRs –– identification of even a small number of cases can be a daunting task –– likely requiring international collaborations and a worldwide reporting system (Wilke et al., 2007).

4.1.4. Clinical utilization

Because the severity of such rare ADRs can sometimes be life-threatening, therefore, screening for these PGx variants that might cause this type of rare/severe ADR can be used as an exclusion criterion (contraindication). However, because of this rarity, sometimes it can be inefficient for pharmacoeconomical reasons, to screen every patient for such rare variants (Daly, 2013). Although individually less frequent, each individual has a higher chance of carrying at least one large-effect PGx variant that, in combination, can cause a major ADR, which can be prevented by preemptive PGx testing [cf. Section 7].

4.2. Predominantly oligogenic PGx traits

4.2.1. PGx response

This group of PGx traits usually involves quantitative variations in concentration of drug metabolites, drug efficacy, dosage, and/or toxic effects. Variability is expressed in a continuous spectrum among individuals in any population. Because of the existence of major contributing factors having predominantly large effect-size, distribution can exhibit multiple distinctive modes (e.g. CYP2D6-metabolizer groups; Fig. 2). However, these effects are usually seen only in well-controlled experimental studies. In clinical settings, the distinction between different modes (subgroups) will usually be blurred –– due to other factors such as drug dosage, drug-drug interactions, and many environmental influences.

4.2.2. Genetic basis

The genetic basis of predominantly oligogenic PGx traits is usually influenced by common genetic variants in one or several major (PK or PD) genes having large-effect. In each of these genes, there may be multiple variants, coding or regulatory, that elicit a functional impact on drug response. Because of the quantitative nature of this group of PGx traits, effect-sizes of the associated genetic variants can be measured by percentage of variance explained (R²). For example, common variants in VKORC1 and CYP2C9, combined, can account for ~40% variance in warfarin dosing (Fung et al., 2012). Cigarette smoking-induced CYP1A2, as well as the heritable CYP1A2 and CYP2D6 polymorphisms, appear to explain ~20% of variance in serum fluvoxamine levels (Carrillo et al., 1996).

4.2.3. Genetic study

Given their relatively large effect-sizes, common variants having a major impact on predominantly oligogenic PGx traits can be identified by genetic association studies –– using a regression-based approach for studying continuous variation in drug responses, or a case/control design based on dichotomization of a continuous drug response gradient, or the extreme discordant phenotype design (Zhang, Nebert, Chakraborty, & Jin, 2006).

Sample-size requirement for this type of study is principally defined by effect-size of the genetic variant and significance level (which needs to be adjusted for multiple-testing, if multiple genetic markers are examined). For example (Fig. 3), a GWAS with ~1000 samples could have ~80% power to identify variants that explain 4% variance at a significance level of 5e–8 (adjusted for 1 million independent tests). After identification of the major genes or variants having large impact (say, R² > 5%), increasing sample-size will continue to reveal more variants having minor effects. However, these small-effect variants –– even in combination –– will not substantially increase variance explained in drug response; hence, these variants will not substantially improve the predictive value. That is to say, those genetic variants that cannot be identified within a reasonably large sample-size (e.g. N = 1000) will have minimal predictive value in each individual patient. Using warfarin as an example, although larger GWAS did identify additional variants robustly associated with the phenotype, they added limited predictive power to the two major genes (VKORC1 and CYP2C9).

Fig. 3.

Effect-size (measured by variance explained, R²) that can be identified by GWAS as a function of different sample sizes [power = 80%, significance level α = 5 × 10⁻⁶ (green dashed line) and 5 × 10⁻⁸ (red solid line)]. As shown by the figure, GWAS with ~1000 samples have appropriate power (>80%) to detect genetic variants that are likely to explain ~4% of phenotypic variance. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

4.2.4. Clinical utilization

Given the relatively large effect-size, it is possible to predict predominantly oligogenic PGx traits based on major contributing genetic variants. However, the true clinical utility might be complicated by several challenges:

1. In addition to major genetic determinants, this type of PGx trait is usually influenced by several or many environmental factors;

2. The predictive value can be different between ethnic populations. For example, VKORC1 and CYP2C9 variants account for as muchas 40% of variability in warfarin-dosing among Europeans, but <10% among African Americans (Limdi et al., 2008; Limdi et al., 2010; Momary et al., 2007);

3. Predominantly oligogenic PGx traits under evaluation (such as dosage, drug or metabolite concentrations) may not be directly related to the ultimate treatment outcome (Zineh, Pacanowski, & Woodcock, 2013).

In addition, there might be a better alternative to genetic testing –– a more accurate dosing or treatment option can often be achieved more rapidly by intensive monitoring. All these factors will require the true clinical utility and pharmacoeconomics of PGx tests to be established based on clinical trials (Ioannidis & Khoury, 2013; Rioux, 2000), with sample-sizes that are much larger than original PGx studies that identified the genetic association.

Using warfarin PGx dosing as an example, the initial association with CYP2C9 and VKORC1 was identified and validated by studies with < 100 samples (Wadelius & Pirmohamed, 2007); a dosing algorithm was later established by a large clinical study involving >5000 samples (International Warfarin Pharmacogenetics et al., 2009); however, the true clinical utility has recently been challenged by three independent clinical trials (Kimmel et al., 2013; Pirmohamed et al., 2013; Verhoef et al., 2013) –– each containing several hundreds to one thousand samples. Some fundamental differences in trial design were believed to cause the discordant findings (Pirmohamed et al., 2015; Zineh et al., 2013). This example underscores the need for more elegantly designed and methodologically rigorous clinical studies, to definitively clarify issues of public health relevance of pharmacogenetics.

4.3. Complex PGx traits – principally drug efficacy and ADRs

4.3.1. PGx response

The group of complex PGx traits usually involves variations in efficacy, side effect, toxicity, and common ADRs. Clinically, although one might simply dichotomize the trait as “efficacious” vs “inefficacious” or “causing” vs “not causing” an ADR –– this type of variability is decidedly continuous in nature and usually follows a unimodal bell-shaped distribution (Fig. 1D), indicating involvement of multiple factors each having small-effect. Examples of variable responses to commonly used antidiabetic, lipid-lowering, antihypertensive, and antidepressant medications (vide supra) probably fall into this category.

4.3.2. Genetic basis

Although complex PGx GWAS are still sparse (Table 1), as far as providing a complete picture of the genetic architecture (Zhou & Pearson, 2013), it is reasonable to presume that the genetic basis of this group of PGx traits might resemble many other complex human traits, i.e. influenced by at least hundreds if not thousands of variants. And, similar to other complex traits, the genetic architecture of complex PGx traits might follow more closely the infinitesimal model described earlier (Gibson, 2011) –– the variance being additively contributed by a very large number of causal variants, each having small-effect. Large-effect rare variants may exist, but will contribute little to variance explained in a population, due to their low allelic frequencies (Zhang, 2015). Genetic variants associated with complex PGx traits identified so far (Section 3.2.3 and Table 1) are mostly common variants (Zhou & Pearson, 2013).

4.3.3. Genetic study

Because of the complexity in genetic etiology and the uniformly small-effect sizes of underlying genetic variants, the only means to identify genetic variants associated with complex PGx traits is by carrying out well-powered (with at least thousands of samples) hypothesis-free GWAS. The resultant identified variants might help in revealing new biology behind drug action or even lead to identification of potential therapeutic targets; however, the results will have very limited predictive power or clinical utility, given their small effect-size. Large sample-sizes may enable identification of small-effect variants; however, it is unrealistic to imagine that ever-increasing sample sizes (to tens of thousands, or more) will reveal sufficient numbers of small-effect variants that, in combination, can explain a substantial fraction of variance and have predictive power.

4.3.4. Clinical utilization

Because precision of predicting PGx complex traits is primarily defined by the percentage of variance explained by predictors, it is difficult to imagine that there will be any chance to predict complex PGx traits by identifiable small-effect genetic variants, even in combination (Fig. 4).

Fig. 4.

Performance of diagnostic test for a dichotomized outcome (evaluated by AUC, area-under-curve) is principally determined by percentage of variance in phenotype that can be explained by predictors. The three curves illustrate the AUC as a function of phenotypic variance explained by predictive variable(s), providing different prevalences of 0.01 (red solid line), 0.1 (green dashed line), and 0.50 (blue dotted line), of the dichotomized outcome. Given the same level of variance explained, greater discriminatory power (i.e. larger AUC) can be obtained for less frequent outcomes. Generally, in order to achieve good discriminative power (AUC >80%; grey horizontal line), the predictors should explain at least 20%–40% phenotypic variance; however, for most complex PGx traits, the identified genetic variants explain far less phenotypic variance than this amount. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

It should be noted that the above categorization of PGx traits is not absolute: the genetic architecture of PGx traits should more or less follow a continuous spectrum, with overlapping among the three categories. For example, the slow-acetylator type –– as measured by serum isoniazid levels in experimental subjects –– might follow a Mendelian inheritance pattern (autosomal recessive); however, the concentration of isoniazid metabolites could be influenced by drug dosage, drug-drug interactions or other PK genes, such as CYP2E1 and GST1, thereby changing the pattern to a predominantly oligogenic distribution. Moreover, even additional genes might contribute to isoniazid-induced hepatotoxicity.

5. Current status of clinical implementation of PGx – review of FDA PGx labels

We surveyed the current status of clinical utility of PGx by examining the Table of Pharmacogenomic Biomarkers in Drug Labeling, published by the FDA (http://www.fda.gov/Drugs/ScienceResearch/ResearchAreas/Pharmacogenetics/ucm083378.htm). The 204 records in that table annotate PGx-labeling information between 165 unique drugs and 67 unique genes or chromosomal abnormalities that influence action of these drugs.

Of the 165 drugs with FDA-approved PGx labeling information, the largest group is the antineoplastic agents (N = 52), 42 of which display target-specific somatic changes. The second largest group is the psychotropic drugs (N = 23), followed by central nervous system agents (N = 19) and anti-infectives (N = 16) (Fig. 5).

Fig. 5.

Different classes of drugs having FDA PGx labels.

According to PharmGKB (https://www.pharmgkb.org/view/drug-labels.do) (Whirl-Carrillo et al., 2012), the level of action implied in each PGx label can be categorized into: [a] genetic testing required; [b] genetic testing recommended; [c] actionable PGx, and [d] informative PGx. Basically, the first two categories (genetic testing required or recommended) imply that some genetic testing should be conducted, or is recommended, before using the drug for its indication (or contraindication) of usage, dosage and administration, and/or use in a specific population. The actionable PGx category contains information about changes in drug efficacy or risk of ADRs due to genetic variations, with or without dosage recommendations. The actionable PGx information is usually established on relatively strong evidence. The last category (informative PGx) refers to possible involvement of a gene or protein in drug-action, but no clear evidence to date suggests a specific genetic variant associated with a variable drug response. Definitions of these PGx levels can be found at (https://www.pharmgkb.org/page/drugLabelLegend).

Following this categorization method, PharmGKB annotated FDA drug labels with the “PGx level” tag (https://www.pharmgkb.org/view/drug-labels.do). The majority (32 of 42) of antineoplastic drugs are known to target specific somatic changes that require genetic testing in order to guide usage; however, for drugs with PGx traits related to germline genetic variation, only 18 (14.6%) contain genetic-testing information. Among those 18, eight drugs target inborn-errors-of-me-tabolism (e.g. urea-cycle enzyme deficiency, N-acetylglutamate synthase deficiency), or genetic disorders (CFTR), or cancer (BRCA1 or BRCA2 mutations/olaparib). The remaining ten drugs require genetic testing to detect major deficiencies in PK genes (TPMT, CYP2D6 and CYP2C19) or HLA genes to avoid adverse reactions.

Of the 67 genes or chromosomal changes in the FDA PGx labels, 20 are somatic changes targeted by numerous (N = 42) antineoplastic drugs. For the remaining 47 genes with germline variants that have a PGx effect, 28 genes are actually those in which mutations cause rare metabolic disorders (such as CYB5R1–4: cytochrome b₅ reductase deficiency; AKR1D1, AMACR, CYP27A1, CYP7A1, HSD3B7: bile acid synthesis disorders; HPRT1: abnormal uric acid metabolism) or genetic diseases (CASR, F2, F5, PROC, PROS1, SERPINC1, CFTR, BRCA1, BRCA2). For understanding these diseases, cf. https://www.ncbi.nlm.nih.gov/omim.

These rare conditions involve either the treatment target or they can increase toxicity of drugs being administered for other purposes. After excluding these 28 “rare genetic disorder” genes, there are only 19 PGx genes in the conventional sense –– ten of these are PK genes, including five phase I (CYP2D6, CYP2C9, CYP2C19, CYP2B6, CYP3A5), four phase II (UGT1A1, TPMT, NAT2, NAT1), one transporter gene (SLCO1B1) and one inborn-error-of-metabolism gene (G6PD). The remaining genes in the list of FDA PGx labels include four PD genes (VKORC1, IFNL3, LDLR, RYR1) and four HLA genes (HLA-A, HLA-B, HLA-DQA1, HLA-DRB1).

We might also categorize PGx genes based on the “PGx level” of associated PGx information in the FDA labels (annotated by PharmGKB). Most of the genes/chromosomal abnormalities (16 of 20) that are involved in somatic-cell alterations require genetic testing. However, only five of the “conventional” 19 PGx genes have genetic-testing information with regard to at least one associated drug; these include: G6PD, CYP2D6, HLA-A, HLA-B (Genetic testing recommended) and TPMT [Genetic testing recommended, but the red-cell enzyme assay (phenotype) is preferred]. Those five genes having genetic testing information are either PK genes with extreme metabolic phenotypes (i.e. CYP2D6 PM or UM, TPMT deficiency), or inborn-error-of-metabolism genes (i.e. G6PD deficiency), or HLA genes responsible for severe ADRs. For the majority of PK or PD genes that show a more continuous variation-gradient in the population, although one can explain a portion of dissimilarity in drug response, there are no established genetic tests yet.

6. Genomic prediction of drug response—challenges and opportunities

6.1. Importance of the environment and epigenomics

Beyond the scope of this review is the role of the environment (including hepatic and renal function, diet, physical activity, mental stress, cigarette smoking, use of alcohol or recreational drugs, taking over-the-counter drugs concomitantly, occupational exposures to high doses of toxicants, etc.) which obviously can affect drug efficacy, intensity of ADRs, and drug toxicity; hence, the environment can interfere with genetic risk prediction of drug response.

Also, with each additional decade of life, we must keep in mind that environmental insults will cause alterations to our epigenomes. Epigenetic effects include DNA methylation, RNA-interference (RNAi), his-tone modifications, and chromatin remodeling –– all of which contribute overall to the genetic architecture of the individual. Moreover, among the ~70 trillion cells in the human body, there are −210 distinct human cell types. Unlike DNA sequence, which is identical in all cells of the body, humans probably have ~210 distinct epigenomes, all of which are theoretically capable of changing, as a function of time and environmental stimuli. Environmental perturbations also contribute to transgenerational heritable phenotypes, presumed to be manifestations of epigenetic effects, but the mechanisms are poorly understood [reviewed in (Nebert, Zhang, & Vesell, 2013) and (van Otterdijk & Michels, 2016)].

With recent advances in high-throughput technologies, including WGS of a single cell, it will soon be possible to analyze many features of the epigenome in depth –– allowing for the establishment of complete epigenomic profiles for basically every cell-type of interest (Martens, Stunnenberg, & Logie, 2011).

6.2. Challenges of clinical utility in risk prediction vs pharmacoeconomics

A recent initial assessment of the benefits of implementing PGx into the medical management of patients in a long-term-care facility (Saldivar et al., 2016) emphasized that healthcare costs associated with prescription drugs are enormous, particularly in patients with polypharmacy (taking more than five prescription medications). Applying PGx-guided prescription recommendations across this patient population –– resulted in elimination and/or replacement of one to three drugs for 50% of the polypharmacy patient population tested –– with an estimated annual savings of US$621 per patient. Initial assessment of this study shows that there is a clear opportunity for concrete healthcare savings, based solely on careful prescription drug management when incorporating PGx testing.

Interestingly, the positive predictive value of HLA-B*38:02:01 in anticipating carbimazole/methimazole-induced agranulocytosis is 0.07 and the negative predictive value is 0.999. Thus, it was extrapolated that ~211 cases need to be screened in order to prevent one case; screening for the risk allele will therefore be useful in preventing agranulocytosis –– but only in populations in which frequency of the risk allele is high (Cheung et al., 2016). This is an excellent example of a potent effect of a rare-variant allele with, unfortunately, low penetrance.

A recent review (Abul-Husn, Owusu, Sanderson, Gottesman, & Scott, 2014) summarized the current state of implementing genetic testing for personalized medicine, with emphasis on clinical pharmacogenetic testing. This included a table listing company-based PGx tests approved by the US FDA for “in vitro diagnostic” use.

6.3. The field of “precision oncology”

Precision oncology (Relling & Evans, 2015), as discussed in Section 5, is not the same as PGx prediction of risk in therapeutic treatment of the whole patient. All cancers arise as a result of DNA mutations, which confer growth advantage of tumor cells. These specific mutations may be targets for treatment and might easily influence therapeutic responses significantly. Specific malignant cells –– today, in fact, even a single cell –– can be genotyped and the most appropriate drug can then be selected, because it is quite certain to be tailored to inhibit growth successfully of that patient's tumor.

6.4. Consideration of cost-benefit ratios

ADRs are an important public health concern, accounting for ~5% of all hospital admissions and two-thirds of all complications occurring shortly after hospital discharge. There are often long delays between when a drug is approved, and when serious ADRs are identified. Recent and ongoing advances in drug safety surveillance include: [a] establishment of government-sponsored networks of population databases, [b] use of data-mining approaches, and [c] formal integration of diverse sources of drug safety information. These advances promise to decrease the delay-time in identifying drug-related risks, or in providing reassurance about the absence of such risks (Hennessy & Strom, 2015).

7. Conclusions and future perspectives

As we have indicated throughout this review, the promise of personalized medicine is far from being realized –– especially in genetic prediction of responses to commonly used drugs, and in the treatment of extremely complex phenotypes such as T2D, hypertension and major depression disorder. Lack of evidence has been widely accepted as one of the most significant challenges of PGx implementation into clinical practice (Khoury et al., 2008; Swen et al., 2007). To overcome this barrier, a continuum of “translational research” agenda has been proposed (Khoury et al., 2007) to accelerate translation of genomic discoveries into healthcare benefits. However, as we have emphasized throughout this review, many valid but small-effect PGx associations, even in combinations, are not likely to improve genetic prediction of drug response in the individual. Among other major challenges of PGx implementation into clinical practice –– is the lack of processes required to interpret genotype information, and to translate genetic information into clinically doing something (Relling & Klein, 2011).

Genomic studies with large sample-sizes, using GWAS or NGS technologies, will continue to identify genetic variants associated with countless drug responses, and this information might increase our understanding of the biology of a drug's mode-of-action or mechanism-of-action. However, the multifactorial nature of complex PGx traits may ultimately limit the possibilities for accurate prediction (Janssens & van Duijn, 2008; Nebert & Zhang, 2012). In view of this realization, it might be important to accept that there are “translational boundaries” in PGx prediction of drug response –– usually restricted to simple (Men-delian or predominantly oligogenic) PGx traits influenced by genetic variants having large-effect.

As WGS technologies become more readily available and cheaper in clinical settings, the cost-benefit balance will gradually become more favorable to clinical implementation of genomic tests. Coupled with the development of more accurately annotated and clinically relevant resources –– such as PharmGKB (Whirl-Carrillo et al., 2012), CPIC (The Clinical Pharmacogenetics Implementation Consortium) (Relling & Klein, 2011), as well as integration of genomic data into healthcare record systems –– computer-aided interpretations of PGx test results will become more feasible and should be helpful in clinical decision-making (Relling & Evans, 2015). In the context of clinical implementation of PGx, a preemptive approach (Dunnenberger et al., 2015) is particularly attractive, because the genotype does not change over time, i.e. a one-time genome-wide genotyping result can almost always be used to guide any prescribing decision throughout one's entire life. One exception to this generalization would be the patient's epigenomes of various cell-types, which are constantly changing. The contribution of epigenetics to genetic risk prediction is expected to become a very active area of investigation in the near future.

Although the overall health benefit has yet to be established, several studies (Ji et al., 2016; Van Driest et al., 2014) have already shown that the vast majority of patients carry at least one actionable PGx variant that could potentially be used to improve prescribed drug treatment in the clinic. Despite these promising aspects of PGx, however, we must recognize that –– as we have attempted to convey throughout this review –– genetic risk prediction of drug response works only for a few drugs, and it is usually based on the genotype of a few genetic variants having large-effect contributions to the phenotype (efficacy effect, adverse effect, and toxic effect). Hence, genomic prediction of complex PGx responses, based on the individual patient's whole-genome readout, is most likely unattainable in the foreseeable future.

Abbreviations

ADR

adverse drug reaction

GRR

genotypic relative risk

GWAS

genome-wide association study or studies

LD

linkage disequilibrium

MAF

minor allele frequency

OR

odds ratio

PD genes

those encoding proteins involved in pharmacodynamics

PGx

pharmacogenetics or pharmacogenomics

PK genes

those encoding proteins involved in pharmacokinetics

SNP

single-nucleotide polymorphism or variant

WES

whole-exome sequencing

WGS

whole-genome sequencing