For reprint orders, please contact:

Review For reprint orders, please contact: reprints@futuremedicine.com Prediction of warfarin dose: why, when and how? Prediction models are the key to individualized drug therapy. Warfarin is a typical example of where pharmacogenetics could help the individual patient by modeling the dose, based on clinical factors and genetic variation in CYP2C9 and VKORC1. Clinical studies aiming to show whether pharmacogenetic warfarin dose predictions are superior to conventional initiation of warfarin are now underway. This review provides a broad view over the field of warfarin pharmacogenetics from basic knowledge about the drug, how it is monitored, factors affecting dose requirement, prediction models in general and different types of prediction models for warfarin dosing. KEYWORDS: dosing algorithm human CYP2C9 protein human VKORC1 protein pharmacogenetics models statistical regression ana lysis warfarin There is overwhelming evidence that the individual maintenance dose of warfarin can be predicted by pharmacogenetic dose models [1 4]. In 2007 the US updated the label of warfarin to encourage, but not require, pharmacogenetic testing of patients initiating warfarin therapy [5]. In 2010 dosing recommendations according to genotypes of CYP2C9 and VKORC1 were added to the label (Table 1). Starting a patient s treatment with an individually predicted warfarin dose would, in theory, lower the risk of early bleeding, which is highly related to the intensity of anticoagulation [6]. In spite of this, the translation of pharmaco genetic dose models into clinical practice has been slow. Clinical studies aiming to show whether pharmacogenetic warfarin dose predictions are superior to conventional initiation of warfarin are now underway. This review provides an overview of the field of warfarin pharmacogenetics including basic information about the drug and how the anticoagulation effect is monitored, common problems encountered during treatment and the impact of genetic factors. Special focus is put on warfarin dose prediction models and how they are developed and evaluated. Warfarin Since the introduction of warfarin in the early 1950s, its use has steadily increased. It is now the most widely used anticoagulant in the world with annual prescriptions equaling 0.5 1.5% of the population [7]. Warfarin is a vitamin K-antagonist that targets the vitamin K cycle, thereby inhibiting the vitamin K-dependent coagulation factors II, VII, IX and X. It is administered as a racemic mixture of (R) and (S)-enantiomers with the (S)-isomer being three- to five-times more potent then the (R)-form [8]. Warfarin is mainly used for primary and secondary prevention of venous thrombo embolism, and for prevention of systemic embolism in patients with prosthetic heart valves or atrial fibrillation [6]. Studies on atrial fibrillation patients show that the overall risk reduction of stroke is approximately 70% compared with placebo [6]. Monitoring of anticoagulation One of the main problems with warfarin therapy is the large interindividual variation in the dose needed to reach adequate levels of anticoagulation. Dose requirements vary more than tenfold, ranging from <10 to >100 mg per week. The anticoagulant effect therefore needs to be monitored, especially during the initiation of therapy. This is done by measuring the prothrombin time (PT) International Normalized Ratio (INR), which is a measure of three of the four vitamin K dependent coagulation factors: II, VII and X. The INR is determined by dividing the PT of a patient with the geometric mean of PT for at least 20 healthy subjects of both sexes with the same test system [9]. An INR of 1.0 is considered to be normal coagulation and an INR of 2.0 means that the clotting time has been doubled. INR target ranges vary between countries and indications, but the most common target range is 2.0 3.0. For patients with a high risk of thrombosis, for example, patients with Niclas Eriksson* & Mia Wadelius 1 1 Department of Medical Sciences, Clinical Pharmacology, Uppsala University, Uppsala University Hospital, entrance 61, SE-751 85 Uppsala, Sweden *Author for correspondence: Uppsala Clinical Research Center, Uppsala University, Dag Hammarskjölds väg 50A, SE-751 85 Uppsala, Sweden Tel.: +46 18 611 9516 Fax: +46 18 506 638 niclas.eriksson@ucr.uu.se part of 10.2217/PGS.11.184 2012 Niclas Eriksson Pharmacogenomics (2012) 13(4), 429 440 ISSN 1462-2416 429

Review Eriksson & Wadelius Table 1. Ranges of recommended warfarin doses (mg/day) from the US drug label. VKORC1 CYP2C9 (mg/day) *1/*1 *1/*2 *1/*3 *2/*2 *2/*3 *3/*3 GG 5 7 5 7 3 4 3 4 3 4 0.5 2 AG 5 7 3 4 3 4 3 4 0.5 2 0.5 2 AA 3 4 3 4 0.5 2 0.5 2 0.5 2 0.5 2 Reproduced from the updated warfarin (Coumadin) product label. mechanical heart valves, the target range can be increased to 2.5 3.5 or even higher [10]. Patients receiving warfarin have frequent INR checks during the induction phase of therapy and once the target range is achieved the INR is measured once or twice a month. A measure of how well a patient is anticoagulated during a specific time interval is the time in treatment range (TTR). The standard way to calculate TTR is by the Rosendaal method, which uses linear interpolation to calculate the percentage of time a patient is within treatment range (usually INR 2.0 3.0) [11]. TTR is associated with the efficacy of warfarin treatment, and a 10% increase in time spent outside therapeutic range relates to an augmented risk of mortality, ischemic stroke and other thromboembolic events [12,13]. The induction phase There is no gold standard for how to initiate warfarin. The half-life of warfarin varies greatly between individuals, and is on average 36 42 h [6]. This means that it takes more than a week to reach a pharmacokinetic steady state when starting on a maintenance dose. Some apply a defensive strategy in patients not requiring rapid anticoagulation by starting with low doses and raising them over time with the guidance of frequent INR tests. This method is safe with respect to over-anticoagulation, which is related to bleeding, but patients requiring higher doses than average will be under-anticoagulated for a large part of the induction phase [10]. Loading doses are therefore often used to reach steady state more quickly. Commonly used initiation methods include giving 5 10 mg on the first 2 days or giving age-stratified initiation doses before switching to maintenance doses [14,15]. Adverse effects The most common adverse effect of warfarin is bleeding and the risk is highly related to the intensity of anticoagulation [13]. The incidence of major bleeding during warfarin treatment is, according to the recent RE-LY trial, on average 3.36% per patient year, but varies between studies depending on inclusion criteria and quality of medical care in terms of INR control [16]. Studies have demonstrated that increasing the INR target range from 2.0 3.0 to 3.0 4.5 also increases the risk of clinically significant bleeding [6]. An interesting observation from these studies is that the mean increase in dose per day is 1 mg when changing from the low to the high target range. If a patient is severely over-anticoagulated with high INR levels or bleeding, treatment is normally stopped and vitamin K is administered to reverse the effects of warfarin. In case of severe bleeding there is also an option to infuse fresh plasma or prothrombin concentrate [6]. Dietary interactions Vitamin K is the natural antidote to warfarin. Vitamin K is found in food, and diet therefore imposes variability in warfarin response. Normal intake of vitamin K is in the range of 60 200 µg/day. Most dark green vegetables such as broccoli, Brussels sprouts and spinach contain high levels of vitamin K (>100 µg/2 dl), but also other common foodstuffs contain a fairly large amount of vitamin K [17]. It is estimated that an increase in vitamin K intake of 100 µg per day for 4 consecutive days lowers the INR by 0.2 and vitamin K intake has therefore been incorporated into some warfarin dose prediction models [18,19]. Studies on combining warfarin treatment with a vitamin K supplement have been performed with the aim to reduce the variability in drug response caused by a low or sporadic dietary intake [20]. These studies show varying results, but the overall conclusion is that vitamin K supplements do decrease the variation in drug response caused by dietary intake. The current recommendation for patients on warfarin treatment is to keep a constant intake of vitamin K through foodstuffs to minimize variation in warfarin response. Drug interactions The two warfarin isomers are metabolized by different pathways. The main enzyme involved in the metabolic elimination of (S)-warfarin is CYP2C9, while (R)-warfarin is eliminated by CYP1A1/CYP1A2/CYP3A4 [5]. Patients starting or stopping drugs that are known inducers or inhibitors of these enzymes should have extra INR tests, and their dose of warfarin adjusted 430 Pharmacogenomics (2012) 13(4)

Prediction of warfarin dose: why, when & how? Review accordingly. Examples of drugs that interact with warfarin are amiodarone that inhibits the clearance of warfarin thereby potentiating the anticoagulant effect, and carbamazepine, phenytoin and rifampicin, which induce the metabolism of warfarin [1]. Drugs that inhibit clotting and increase the risk of bleeding, like aspirin, diclofenac and ibuprofen, are also regarded as interacting drugs even though they have no inhibiting or inducing effect on the elimination of warfarin [6]. Clinical & demographic effects on dose Although warfarin is almost entirely eliminated by metabolism, recent studies show that patients with moderate and severe renal impairment (estimated glomerular filtration rate: 30 59 ml/min/1.73 m 2 and <30 ml/min/1.73 m 2 ) require less warfarin than those with no or mild kidney impairment [21,22]. Among other factors that have effect on warfarin dose are age, height, weight, ethnicity, smoking and target INR value [2,4,23]. Genetic effects A genetic explanation to the large inter - individual variation in warfarin dose requirement was first sought with the candidate gene approach and later with genome-wide association studies. In the 1990s candidate gene studies focused on the enzyme CYP2C9, which metabolizes the potent (S)-isomer of warfarin, found an effect of CYP2C9 genetic variants on warfarin dose [24]. It is now well-known that patients with normal metabolism, that is, no genetic variants of CYP2C9 (*1/*1), have an estimated half-life of 30 37 h, while it is prolonged to 92 203 h in patients with severely impaired metabolism (*3/*3) [25,26]. Some years later the gene VKORC1 that codes for the enzyme vitamin K epoxide reductase [27], which is the target of warfarin, was discovered [28 30]. Initial studies showed that common variants of VKORC1 make patients more sensitive to warfarin [31 33]. Recent research has also shown that there are rare mutations in VKORC1 that cause warfarin resistance [34]. The genes CYP2C9 and VKORC1 have been confirmed as the major genetic determinants of warfarin dose through genome-wide studies [35,36]. In addition, a variant of the gene CYP4F2 has been shown to have a small but significant effect on warfarin dose [35,37,38]. The remaining genetic variance is likely to be caused by multiple genetic variants with a very small effect. Developing & critically reviewing prediction models A prediction model is any algorithm relating to a certain measure or measures of an outcome. In the field of medicine, it is most commonly the result of a regression model with estimated co efficients as weights of each variable. Although the process of prediction modeling can differ substantially between different projects/outcomes, there are still some general points to think of. These points are highlighted under Box 1, and are described in more detail here. Knowledge of certain important steps in prediction modeling can both help in future projects and when reading literature including prediction models. First of all, it is of high importance that the data a prediction model is built on reflect the population that it will be used on. For instance, if we have a prediction model that includes age, then the data used to build the model should include the age interval of the future populations where predictions will be made. If a model has been built on data from a cohort aged 20 40 years and is used to predict in patients aged 60 80 years, there is a chance that the co efficient for age is wrong; the ageoutcome relationship might for instance be of second order. This problem could arise if the cohort used to build the model is from a clinical study with narrow inclusion criteria for some key variables, while prediction in a more general population is desired. Think of different validation techniques before any calculations are done. If there is no external data set available and the data at hand is fairly large, validation can be done by splitting the data into a training set and a validation set. This, however, requires that no calculations/assumptions/modeling have been done on the whole data. Data splitting could still give a positive bias, that is, the performance will appear Box 1. Key aspects when prediction modeling. Is data representative for the population the model will be used on? Validation of model; internal or external? Missing data? Imputation? Is the effect of continuous variables linear or should transformation be applied? Read the literature; gather as much information as possible from previous publications about important variables Before discarding any variables from the model, estimate univariate estimates (one estimate of variable effect against outcome at a time) and multiple estimates (all variables included in the model) If variable selection is necessary, decide which variables to drop by investigating estimates of effects with confidence intervals, not only p-values. Do not automatically discard variables with p > 0.05 Collaboration between statistical and medical expertise is essential www.futuremedicine.com 431

Review Eriksson & Wadelius better than in real life, because the validation set is not entirely independent of the derivation set in terms of study wise definitions and design. Unbiased measures of performance are estimated by external validation on data totally i ndependent from the modeling process. Depending on study type, missing values could be a problem. A clinical study is most likely to have a low number of missing values, while data from registries often have more missing values. Missing values cause problems in the ana lysis, since patients with any missing value are discarded if nothing else is done. Most serious is the case of informative missing, that is, when the missing pattern has something to do with the outcome being analyzed. Various imputation techniques can be use to fill in missing values to reduce the amount of patients removed from the ana lysis. If continuous predictors are used, ensure that the relationship to the outcome is linear by plotting Y against the predictor. If there is evidence of a nonlinear effect, then apply some kind of transformation (e.g., log or square root) to attain linearity. In the worst case it is possible to categorize the variable, but think twice before categorizing a continuous variable since this means loss of information. Utilizing previous knowledge about predictors for the outcome will firstly lower the risk for spurious results by making it possible to drop some variables from the ana lysis before they are analyzed against the outcome. Second, it will help select variables for the modeling process. As a first step in the modeling, estimating univariate effects is a good way to gain knowledge about the data. The next step would be to fit a full multiple model using all predictors. This model could be the final prediction model if there is no need for variable selection. However, the model is often improved, in terms of simplicity and clinical utility, by dropping some variables. For instance, do not keep a variable that, although significant, shows signs of being nonimportant (estimate close to zero) if it also costs money to produce in the clinic. Another reason for dropping variables is that a simple model is easier to interpret and use in a clinical setting than a more complex model. In the process of variable selection, do not always obey the p 0.05 significance rule that is inherited from clinical trials; point estimates and CIs for each coefficient include more information. Why is a variable with p = 0.045 more interesting than a variable with p = 0.055 or even p = 0.15? The effect of all three variables could actually be the same, but the precision of the estimate could be different due to one variable being infrequent in the cohort. As an example, if a known interacting drug is used seldom in your cohort and therefore has a low frequency, the precision will be low. This gives large CIs for the estimate, but the estimate itself may be of the same magnitude as in previous studies. In this case the p-value will probably be >0.05, but given the previous knowledge the variable should be kept in the model. The most common performance measure of linear regression models is the coefficient of determination, R 2. It measures the percentage (or proportion) of variability of the outcome that is explained by the model. While the R 2 value gives a point estimate for the predictive ability of the model, examining scatter plots of observed values versus predicted values gives more information about performance across the whole range of outcome values. If external validation of the final model is used, and it includes or only consists of another ethnicity, the estimates of performance might be lower than expected. This could be due to the ethnicities having different allele frequencies of the genetic markers included in the model. This is truly the case for warfarin (see section on Stable warfarin maintenance dose ). For further in depth details about prediction modeling, the authors recommend Regression Modeling Strategies by Frank E Harrell Jr [39]. Why predict warfarin dose? The ultimate goal of pharmacogenetics is to individualize therapy and warfarin is a good example of where pharmacogenetics could help tailor the dose from the beginning of treatment. The narrow therapeutic index of warfarin and high interindividual variation in dose needed is a challenge in clinical care. Complications from inappropriate warfarin dosing remain as one of the most common reasons for emergency room visits in the USA [7,40]. Unintentional overdosing of warfarin is the most common cause of hospitalization due to adverse reactions in the USA, and the second most common cause in the UK [41,42]. There is evidence of a pronounced risk of bleeding during the initiation phase (1 3 months), which a genotype-guided dose has the potential to reduce [6,43]. For instance, a study by Wadelius et al. showed that the risk of over anti coagulation (INR >4) is highly related to CYP2C9 and VKORC1 genotypes [4]. A small prospective cohort study by Gong et al. indicated that the genotypic risk could be removed 432 Pharmacogenomics (2012) 13(4)

Prediction of warfarin dose: why, when & how? Review by using a pharmacogenetics-guided initiation dosing protocol [44]. When to predict warfarin dose In the following sections three different types of prediction model are discussed. The first is the prediction of stable warfarin maintenance dose, which is the most studied outcome in warfarin pharmacogenetics. However, maintenance dose prediction models do not tell us how to initiate treatment in the best way, which leads us to the second type of model: prediction of starting doses. The third type of prediction model is the dose refinement model that also includes INR response after the initial doses. For the different types of models examples from the largest international studies are shown and discussed. Stable warfarin maintenance dose The first prediction models for stable warfarin dose were estimated on relatively small populations mainly consisting of Caucasian patients included from one area or clinic [23]. While these models might predict the dose within the same catchment area, they are probably not portable to a more broad population from other regions or countries [45]. To address the problem of generalizability, the International Warfarin Pharmacogenetics Consortium (IWPC) was established with the primary aim to develop an algorithm for warfarin dose prediction. The IWPC includes 21 research groups from nine countries that have contributed clinical and genetic data for 5700 patients who were treated with warfarin [1]. The resulting pharmaco genetic model was shown to predict patients requiring low doses ( 21 mg/week) or high doses ( 49 mg/week) significantly better than a clinical model. In the validation data the pharmacogenetic algorithm explained 43% of the variance in warfarin dose (R 2 = 43%), while the clinical algorithm had an R 2 of 26%. The IWPC model includes the variables VKORC1 (rs9923231), CYP2C9 (*2 and *3), age, height, weight, race and the following interacting drugs: enzyme inducers and amiodarone (Table 2) [1]. Several studies have externally validated the IWPC model together with other models, and the overall conclusion is that the IWPC model performs well; however, the sample size has been small in these studies [46 48]. In 2010 the US updated their Coumadin label with a table giving stable dosing recommendations per genotype of CYP2C9 and VKORC1 (Table 1). A dose nomogram for the IWPC model illustrating the effect of VKORC1, CYP2C9 and age on week dose (mg) for a Caucasian patient with no interacting drugs, height 175 cm and weight 75 kg is given in Figure 1. The s recommended dosing recommendations from the Coumadin label are shown for comparison (shaded areas). A pharmacogenetic dosing algorithm has been shown to be superior to the table on dosing recommendations [49]. The SNPs commonly used in warfarin maintenance dose prediction models are presented in Table 3 together with the estimates of minor allele frequencies (MAF) from Scott et al. and Limdi et al. [1,2,4,50,51,52]. Notable is the difference in MAF between ethnicities, with more genetic variation among individuals of European descent than in individuals of African and Asian descent. For the VKORC1 SNP rs9923231 there is also a shift in minor allele between Europeans, Asians and Africans with the minor allele being A for Europeans and Africans and G for Asians. The differences in MAF affects how well a prediction model containing this variable, in terms of R 2, performs in the population [53]. The second study by IWPC showed that the reason for the lower performance of the IWPC algorithm in African Americans and Asians (R 2 24%) than in Caucasians (R 2 = 40%) was mainly due to less variation in the VKORC1 SNP rs9923231 in African Americans and Asians. By simulations the IWPC showed that given the population MAF, the R 2 for the VKORC1 SNP rs9923231 was in the range of what would be expected if the genotype effect was the same across ethnicities [52]. The conclusion was that non-caucasian patients with variant alleles have similar benefit of genetic-based dosing with the IWPC model, even if the performance of the algorithm is lower on a population level. In other words genetic based dosing for a single patient will work as well for African Americans and Asians as it does for Caucasians but on a population level Caucasians, with most variation in VKORC1, will benefit most from genetic dosing. Starting doses The time to reach steady-state concentrations of a drug depends on how quickly it is cleared from the system. If you initiate a drug with the maintenance dose, it takes approximately five-times the half-life of the drug to reach steady state. Compared with a person without CYP2C9 variant alleles, the half-life of the more potent enantiomer (S)-warfarin is doubled in a *2/*2 person and increased three- to six-fold www.futuremedicine.com 433

Review Eriksson & Wadelius Table 2. Three different prediction models for warfarin dosing. Study Klein et al., IWPC Avery et al. Lenzini et al., IWDRC Type of model Maintenance dose Initial doses Dose revision on days four or five Algorithm Dose (mg/week) = 5.6044 0.2614 age (in decades) + 0.0087 height (cm) + 0.0128 weight (kg) - 0.8677 VKORC1 A/G - 1.6974 VKORC1 A/A - 0.4854 VKORC1 genotype unknown - 0.5211 CYP2C9 *1/*2-0.9357 CYP2C9 *1/*3-1.0616 CYP2C9 *2/*2-1.9206 CYP2C9 *2/*3-2.3312 CYP2C9 *3/*3-0.2188 CYP2C9 genotype unknown - 0.1092 Asian race - 0.2760 Black or African American - 0.1032 missing or mixed race + 1.1816 enzyme-inducer status - 0.5503 amiodarone status The result is dose (mg/week) and should be squared to calculate the weekly warfarin dose in mg Age in decades should be entered as 1 for 10 19 years, 2 for 20 29 years, 3 for 30 39 years Enzyme-inducer status = 1 if patient taking carbamazepine, phenytoin, rifampin, or rifampicin, otherwise 0 For VKORC1, CYP2C9, race and amiodarone status enter 1 if present, otherwise 0. For example, if the patient is VKORC1 A/G then the coefficient for VKORC1 is -0.8677 x 1 R 1 k* k*2 - (1+ - x e + - x V S ) k* (1 - - x W e ) x = MD * S W S -k* x 1 2 * e k*2 + + - x W S e 3 3 W T X Where MD is the calculated maintenance dose per day in mg. Could be calculated by the IWPC model above (note that the IWPC model gives weekly dose, divide by seven to get the daily dose) k is the elimination rate constant for the CYP2C9 genotypes: *1/*1 = 0.0189 h -1, *1/*2 = 0.0158 h -1, *1/*3 = 0.0132 h -1, *2/*2 = 0.0130 h -1, *2/*3 = 0.009 h -1 and *3/*3 = 0.0075 h -1 t is the warfarin-dosing interval, use 24 (24 h) The initial doses on days one to three are then calculated as: day one = MD + x, day two = MD + 2x/3 and day three = MD + x/3 Dose (mg/week) = EXP (3.10894-0.00767 age - 0.51611 ln[inr] - 0.23032 VKORC1-1639 G>A - 0.14745 CYP2C9*2-0.3077 CYP2C9*3 + 0.24597 BSA + 0.26729 target INR - 0.09644 African origin - 0.2059 stroke - 0.11216 diabetes - 0.1035 amiodarone use - 0.19275 fluvastatin use + 0.0169 dose - 2 + 0.02018 dose - 3 + 0.01065 dose - 4 ) Where INR is the INR on days four or five VKORC1-1639 G>A should be entered as 0 for homozygous G/G, 1 for heterozygous and 2 for homozygous A/A CYP2C9*2 and CYP2C9*3 should be entered as 0 if absent, 1 if heterozygous and 2 if homozygous Dose - 2, - 3 and - 4 refers to dose given 2, 3 and 4 days before INR is measured BSA is calculated according to the formula by Dubois and Dubois where BSA = (weight [kg] 0.425 height [cm] 0.725 )/139.2 For African origin, stroke, diabetes, amiodarone use and fluvastatin use, enter 1 if present, otherwise 0 BSA: Body surface area; INR: International Normalized Ratio; IWDRC: International Warfarin Dose Refinement Collaboration; IWPC: International Warfarin Pharmacogenetics Consortium; MD: Maintainance dose. Ref. [1] [25] [56] in a *3/*3 person. Hence, if treatment is initiated with the predicted maintenance dose, the time to reach steady state is prolonged two- to six-fold in patients with only CYP2C9 variant alleles. Recent work by Hamberg et al. has described these pharmacokinetic differences in more detail [54,55]. The time to reach steady state can be optimized by combining a stable dose prediction algorithm with information on the estimated clearance of CYP2C9 genotypes [25]. This individualized pharmacogenetics-based warfarin initiation dose regimen is especially beneficial in patients with CYP2C9 variant alleles. A table of pharmaco genetic starting doses calculated using the formulas by Avery et al. (see formula in Table 2) is given in Table 4 for an example patient with six different combinations of CYP2C9 and VKORC1 [25]. The predicted maintenance doses are taken from the dose nomogram in Figure 1. These predicted loading doses show that for a Caucasian patient at the age of 60 years, height 175 cm, weight 75 kg, with no interacting drugs and the most common combination of genotypes for Caucasians (CYP2C9 *1/*1 and VKORC1 rs9923231 A/G) the predicted maintenance dose would be 4.71 mg per day and the loading doses would be 7.6 mg on day one, 6.61 mg on day two and 5.66 mg on day three. If a patient with the same clinical and demographic variables had impaired metabolism (CYP2C9 *3/*3) and required a low dose according to VKORC1 (rs9923231 A/A) the predicted maintenance dose would be 1.00 mg and the loading 434 Pharmacogenomics (2012) 13(4)

Prediction of warfarin dose: why, when & how? Review doses would be 3.23 mg on day one, 2.49 mg on day two and 1.74 mg on day three. This shows that the loading doses are less than half and the maintenance dose less than a quarter for a patient with CYP2C9 *3/*3 and VKORC1 A/A compared with a CYP2C9 *1/*1 and VKORC1 A/G patient. Dose refinement A dose refinement model is used after the first doses of therapy and incorporates knowledge about given doses and the INR response. The International Warfarin Dose Refinement Collaboration investigated if genetic information adds to dose prediction after initiating treatment by deriving both pharmacogenetic and clinical dose refinement models [56]. These models that can be used on days four to five were derived on 969 patients and internally validated in 20% of the data that were left out of the modeling. In a final step the model parameters were refitted on the pooled derivation and internal validation cohorts (n = 1213) and evaluated in an external validation cohort (see prediction model in Table 2). Results of internal validation were R 2 60% for the pharmacogenetic model on days four and five and for the clinical model R 2 44% on days four and five. External validation gave somewhat lower accuracy of prediction with R 2 42% for the pharmacogenetic model on days four and five and for the clinical model R 2 28% on days four and five. The overall conclusion was that the pharmacogenetic dose refinement model significantly improved the R 2 by 12 17% compared with a clinical dose refinement model. This indicates that knowledge of CYP2C9 and VKORC1 genotypes still adds information 4 5 days into warfarin therapy. How the prediction models could be used All three types of prediction models can be used in all patients. If possible, use a model derived in a population with similar characteristics to the patients it will be used on, for example, age, ethnicity and diet. Imagine a patient that is to be treated with warfarin due to atrial fibrillation and additional risk factors for stroke with a target INR of 2.5. The patient has the following characteristics; Caucasian, 60 years of age, height 175 cm, weight 75 kg, no interacting drugs and the genotypes CYP2C9 *3/*3 and VKORC1 A/G. To predict the initial doses for this patient first use a pharmacogenetic maintenance dose model, such as the IWPC model 60 60 60 VKORC1 rs9923231:a/a VKORC1 rs9923231:a/g *1/*1 VKORC1 rs9923231:g/g 50 50 50 *1/*1 *1/*2 *1/*1 40 *1/*1 40 *1/*1 *1/*2 *2/*2 30 *3/*3 20 Predicted mean dose/week (mg) 40 *1/*1 *1/*3 *2/*2 *2/*3 30 *3/*3 30 *1/*1 *1/*2 20 Predicted mean dose/week (mg) 20 *3/*3 Predicted mean dose/week (mg) *2/*3 *3/*3 *3/*3 10 10 *1/*3 *2/*2 *2/*3 *3/*3 10 0 0 0 20 40 60 80 100 Age (years) 20 40 60 80 100 Age (years) 20 40 60 80 100 Age (years) CYP2C9 *1/*1 *1/*2 *1/*3 *2/*2 *2/*3 *3/*3 Figure 1. A nomogram that shows the predicted doses for varying age and VKORC1/CYP2C9 genotypes using the International Warfarin Pharmacogenetics Consortium model. Factors kept constant are ethnicity Caucasian, height 175 cm, weight 75 kg and no interacting drugs. Although the original International Warfarin Pharmacogenetics Consortium model has age input as decades (1 for ages 10 19 years, 2 for ages 20 29 years) this figure was created by using age as a continuous variable by dividing the age by 10 years. The shaded areas in the background correspond to dosing recommendations per genotype from the US drug label (Table 1). Prediction model taken from Klein et al. [1]. www.futuremedicine.com 435

Review Eriksson & Wadelius Table 3. SNPs included in current prediction models. Gene SNP Descent/ethnicity n used to estimate MAF MAF (%) Alleles Effect CYP2C9 rs1799853 (*2) White 3062 13.1 C>T The T allele is associated with warfarin Asian 1063 0.0 clearance; patients require lower warfarin doses Black 645 2.9 CYP2C9 rs1057910 (*3) White 3062 7.1 A>C The C allele is associated with warfarin Asian 1063 4.0 clearance; patients require lower warfarin doses Black 645 1.8 VKORC1 rs9923231 White 2426 38.0 G>A The A allele is associated with Asian 883 91.2 decreased warfarin dose Black 643 10.8 CYP4F2 rs2108622 Caucasian 600 34.2 C>T The C allele is associated with Asian 600 30.5 decreased warfarin dose African American 600 11.7 The descent/ethnicity is kept with the original wording from the cited papers. MAF is for the minor allele positioned to the right of the > sign under the Alleles column. Estimates according to Limdi et al. [52]. Estimates according to Scott et al. [51]. MAF: Minor allele frequency. then use the formulas by Avery et al. to calculate the initial doses (Tables 2 & 4) [1,25]. This patient would be predicted a maintenance dose of 11 mg/week (1.6 mg/day) and initial doses of 5.1, 3.9 and 2.7 mg on days one to three, respectively. On day four the patient comes back for an INR test and has an INR of 2.0. At this point a dose revision model could be used, such as the model by Lenzini et al. (Table 2) which would predict a dose of 15.6 mg/week (2.2 mg/day) [56]. Preferably prediction models should be used with the VKORC1 SNP they were derived for, but due to the haplotype structure of VKORC1 it is possible to use other SNPs in linkage disequilibrium (LD) as surrogate markers [32]. Four SNPs are in high LD (r 2 > 0.9) in Caucasians and Asians (rs9923231, rs2359612, rs9934438 and rs8050894) whereas the LD is lower in people of African American ancestry. This LD information was used by the IWPC group when rs9923231 was imputed, more information on the imputation of rs9923231 is in the section S4 of the supplementary information to the IWPC paper [1,101]. All of the discussed prediction models have been derived on datasets of adult patients. The first bullet point under Box 1, Is the data representative for the population the model will be used on, highlights that these models probably would not work in children aged 0 18 years. At the present time many research groups are evaluating genetic effects on warfarin dose in children and a recent paper by Biss et al. showed the current IWPC model constantly over estimates the warfarin dose in children by an average of 1.5 mg/day [57]. They however showed that a similar proportion of the variation in dose was explained by VKORC1 and CYP2C9 in c hildren as in adult patients. Randomized trials The clinical benefit of pharmacogenetic dosing is still to be shown. At least five clinical trials of pharmacogenetic dosing are underway [102]. Three of the largest are the COAG and the GIFT trials in the USA, and the European EU-PACT trial [58 60]. The COAG study is a two-armed, doubleblinded, randomized controlled trial that compares genotype-based dosing with clinical guided dosing. Both treatment arms use a baseline dose-initiation model [2] and a dose refinement model after four or five days of warfarin therapy [56]. The aim is to include over 1200 patients that are expected to be on warfarin therapy for at least 3 months. The primary outcome of the study is TTR within the first 4 weeks of therapy. The GIFT trial is a 2 2 factorial-design, randomized control trial. It has two aims where the first aim is to compare pharmaco genetic dosing with clinical dosing and the second aim is to compare a target range of INR 2.0 3.0 with a 436 Pharmacogenomics (2012) 13(4)

Prediction of warfarin dose: why, when & how? Review lower target range of INR 1.5 2.1. Dosing will be guided by the website WarfarinDosing.org for a minimum of the first 11 days of treatment [103]. The primary study outcome for aim one is the composite of venous thromboembolism, major hemorrhage, INR 4 or death and for aim two the primary outcome is the composite of nonfatal venous thromboembolism or death. The plan is to include 1600 patients undergoing elective total hip or knee replacement surgery. The EU-PACT study is a two-armed, singleblinded, randomized controlled trial aiming to recruit almost 3000 patients commencing anticoagulation therapy with warfarin, acenocoumarol or phenprocoumon. Patients in the intervention arm are dosed according to a drug-specific dosing algorithm, which is based on genetic information and clinical data. For warfarin, three to four starting doses are calculated by combining the IWPC pharmacogenetic prediction model with an estimated accumulation index that is based on the clearance of warfarin in different CYP2C9 genotypes [1,25,55]. Thereafter a pharmacogenetic dose revision algorithm is used [56]. Patients in the control arm are dosed without information about genotype. The follow-up period is 3 months. The primary aim of the study is to show improved anti coagulation therapy by increased TTR during the first 3 months of treatment. New alternatives to warfarin New alternatives to warfarin therapy have recently been investigated. They include the direct thrombin inhibitor dabigatran (RE-LY study) [16] and the factor Xa inhibitors rivoxaban (ROCKET AF study) [61] and apixaban (ARISTOTLE study) [62]. These new oral anticoagulants are proven noninferior or even superior to warfarin; however, the TTR for the warfarin arms of these three studies was low with 64% in the RE-LY study, 55% in the ROCKET study and 62% in the ARISTOTLE study. The cost effectiveness of dabigatran in atrial fibrillation compared with warfarin was studied in a recent paper by Shah et al. [63]. The results were that for atrial fibrillation patients dabigatran is more cost effective in patients with poor INR control (TTR <57.1%) whereas warfarin is more cost effective in patients with excellent INR control (TTR >72.6). Another aspect of these new drugs is that there is no natural antidote, and agents to reverse the effect are still under development [64]. Future perspective The dose of warfarin that is required for therapeutic INR levels is strongly related to SNPs in CYP2C9 and VKORC1. Together with clinical factors, these genetic markers explain a large amount of the variability in dose. The variability in dose requirements between ethnicities is largely caused by different allele frequencies of the SNPs in VKORC1. Compared with dosing according to a clinical algorithm, the IWPC pharmacogenetic model performs better, especially in the high- and low-dose groups. Drawbacks of current pharmacogenetic dosing algorithms, such as the IWPC model, are that they underestimate doses in patients with large dose requirements. The cause of warfarin resistance, resulting in high warfarin doses, is not yet fully understood; it could, for instance, be another polymorphism, diet or nonadherence to taking the medicine as prescribed. To date, the reported prospective clinical studies using prediction models for warfarin Table 4. Pharmacogenetic loading doses according to Avery et al. CYP2C9 VKORC1 IWPC (dose/week in mg) IWPC (dose/day in mg) Day one (mg) Loading doses Day two (mg) *1/*1 A/A 24 3.43 5.51 4.82 4.12 *3/*3 A/A 7 1.00 3.23 2.49 1.74 *1/*1 A/G 33 4.71 7.56 6.61 5.66 *3/*3 A/G 11 1.57 5.07 3.9 2.74 *1/*1 G/G 44 6.29 10.1 8.83 7.56 *3/*3 G/G 18 2.57 8.3 6.39 4.48 Day three (mg) Approximate predictions are from the example nomogram in Figure 1 for a Caucasian patient with no interacting drugs, age 60 years, height 175 cm and weight 75 kg [1]. Prediction of loading doses were made using k (elimination rate constant) of 0.0189 for *1/*1 and 0.0075 for *3/*3 as in Avery et al. [25]. IPWC: International Warfarin Pharmacogenetics Consortium. www.futuremedicine.com 437

Review Eriksson & Wadelius dosing have been small. There is as yet no conclusive evidence for the recommendation to genotype CYP2C9 and VKORC1 when initiating a patient on warfarin [44,50]. New, well designed, large studies are ongoing to provide clinical utility and validity for personalized warfarin therapy. These ongoing studies have also been a catalyst for the development of point of care systems for rapid genotyping. Using blood, without DNA extraction, these assays can successfully genotype one sample with high accuracy in 2 h [65]. The use of pharmacogenetic dosing of warfarin is dependent on the clinical utility of the point of care testing, the availability and cost of genotyping and the cost effectiveness compared with new oral anticoagulants [63,66,67]. Regardless of this, the future of pharmacogenetic warfarin dosing relies on the results of the ongoing clinical trials. Financial & competing interests disclosure M Wadelius is supported by the Swedish Research Council (Medicine 523-2008-5568), the Swedish Heart and Lung Foundation, EU FP7 (HEALTH-F2-2009-223062), and the Clinical Research Support (ALF) at Uppsala University. The funding organizations played no role in the writing of this review. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed. No writing assistance was utilized in the production of this manuscript. Executive summary Warfarin Warfarin is monitored by the International Normalized Ratio (INR), which is a measure of the clotting ability of blood. A level of 1 is associated with normal coagulation and 2 indicates doubled coagulation time compared with normal. The optimal effect for most patients is in the target range of INR 2.0 3.0 and the required dose to reach such levels varies between <10 mg per week to >100 mg per week. Adverse effects The most common adverse effect of warfarin is bleeding and the risk is highly related to the intensity of anticoagulation. Genetic effects CYP2C9 is associated with the clearance of warfarin. Patients with normal metabolism (CYP2C9 *1/*1), have an estimated half-life of 30 37 h whereas those with severely impaired metabolism (CYP2C9 *3/*3) have an estimated half-life of 92 203 h. VKORC1 codes for the enzyme vitamin K epoxide reductase, the target of warfarin. Why predict warfarin dose? Unintentional overdosing of warfarin remains as one of the most common causes of hospitalization due to adverse reactions in the USA and the UK. Genotype-guided dose has the potential to reduce the risk of bleeding during the initiation phase (1 3 months) of warfarin. When to predict warfarin dose Three different kinds of pharmacogenetic prediction models exist: - Prediction of stable maintenance dose. - Prediction of loading doses to be used during the initiation of warfarin. - Dose revision models including the INR response on day 4 and later. Randomized trials Randomized trials incorporating pharmacogenetic dosing of warfarin have to date been too small to draw solid conclusion about the value of genotyping before drug initiation. At least three large trials are ongoing; the COAG and the GIFT trials in the USA, and the European EU-PACT trial. References Papers of special note have been highlighted as: of interest of considerable interest 1 Klein TE, Altman RB, Eriksson N et al. Estimation of the warfarin dose with clinical and pharmacogenetic data. N. Engl. J. Med. 360(8), 753 764 (2009). Largest study to date used to derive a prediction model. Used in the EU-PACT trial. 2 Gage BF, Eby C, Johnson JA et al. Use of pharmacogenetic and clinical factors to predict the therapeutic dose of warfarin. Clin. Pharmacol. Ther. 84(3), 326 331 (2008). Large study used to derive a prediction model. Used in the COAG and GIFT trials. 3 Wadelius M, Chen LY, Eriksson N et al. Association of warfarin dose with genes involved in its action and metabolism. Hum. Genet. 121(1), 23 34 (2007). Extensive study investigating the effect on warfarin dose of genes in the warfarin pathway. 4 Wadelius M, Chen LY, Lindh JD et al. The largest prospective warfarin-treated cohort supports genetic forecasting. Blood 113(4), 784 792 (2009). Large pharmacogenetic study on genetic effects on warfarin dose and other outcomes. 438 Pharmacogenomics (2012) 13(4)

Prediction of warfarin dose: why, when & how? Review 5 Gage BF, Lesko LJ. Pharmacogenetics of warfarin: regulatory, scientific, and clinical issues. J. Thromb. Thrombolysis 25(1), 45 51 (2008). 6 Hirsh J, Fuster V, Ansell J, Halperin JL. American Heart Association/American College of Cardiology Foundation guide to warfarin therapy. J. Am. Coll. Cardiol. 41(9), 1633 1652 (2003). 7 Johnson JA, Gong L, Whirl-Carrillo M et al. Clinical Pharmacogenetics Implementation Consortium Guidelines for CYP2C9 and VKORC1 genotypes and warfarin dosing. Clin. Pharmacol. Ther. 90(4), 625 629 (2011). 8 Wittkowsky AK. Warfarin and other coumarin derivatives: pharmacokinetics, pharmacodynamics, and drug interactions. Semin. Vasc. Med. 3(3), 221 230 (2003). 9 Poller L. International Normalized Ratios (INR): the first 20 years. J. Thromb. Haemost. 2(6), 849 860 (2004). 10 Keeling D, Baglin T, Tait C et al. Guidelines on oral anticoagulation with warfarin fourth edition. Br. J. Haematol. 154(3), 311 324 (2011). 11 Rosendaal FR, Cannegieter SC, van der Meer FJ, Briet E. A method to determine the optimal intensity of oral anticoagulant therapy. Thromb. Haemost. 69(3), 236 239 (1993). 12 Jones M, Mcewan P, Morgan CL, Peters JR, Goodfellow J, Currie CJ. Evaluation of the pattern of treatment, level of anticoagulation control, and outcome of treatment with warfarin in patients with non-valvar atrial fibrillation: a record linkage study in a large British population. Heart 91(4), 472 477 (2005). 13 Connolly SJ, Pogue J, Eikelboom J et al. Benefit of oral anticoagulant over antiplatelet therapy in atrial fibrillation depends on the quality of International Normalized Ratio control achieved by centers and countries as measured by time in therapeutic range. Circulation 118(20), 2029 2037 (2008). 14 Heneghan C, Tyndel S, Bankhead C et al. Optimal loading dose for the initiation of warfarin: a systematic review. BMC Cardiovasc. Disord. 10, 18 (2010). 15 Kovacs MJ, Rodger M, Anderson DR et al. Comparison of 10 mg and 5 mg warfarin initiation nomograms together with low-molecular-weight heparin for outpatient treatment of acute venous thromboembolism. A randomized, double-blind, controlled trial. Ann. Intern. Med. 138(9), 714 719 (2003). 16 Connolly SJ, Ezekowitz MD, Yusuf S et al. Dabigatran versus warfarin in patients with atrial fibrillation. N. Engl. J. Med. 361(12), 1139 1151 (2009). 17 Booth SL, Centurelli MA. Vitamin K: a practical guide to the dietary management of patients on warfarin. Nutr. Rev. 57(9 Pt 1), 288 296 (1999). 18 Khan T, Wynne H, Wood P et al. Dietary vitamin K influences intra-individual variability in anticoagulant response to warfarin. Br. J. Haematol. 124(3), 348 354 (2004). 19 Pavani A, Naushad SM, Rupasree Y et al. Optimization of warfarin dose by populationspecific pharmacogenomic algorithm. Pharmacogenomics J. doi:10.1038/tpj.2011.4 (2011) (Epub ahead of print). 20 Ford SK, Moll S. Vitamin K supplementation to decrease variability of International Normalized Ratio in patients on vitamin K antagonists: a literature review. Curr. Opin. Hematol. 15(5), 504 508 (2008). 21 Limdi NA, Beasley TM, Baird MF et al. Kidney function influences warfarin responsiveness and hemorrhagic complications. J. Am. Soc. Nephrol. 20(4), 912 921 (2009). 22 Limdi NA, Limdi MA, Cavallari L et al. Warfarin dosing in patients with impaired kidney function. Am. J. Kidney Dis. 56(5), 823 831 (2010). 23 Sconce EA, Khan TI, Wynne HA et al. The impact of CYP2C9 and VKORC1 genetic polymorphism and patient characteristics upon warfarin dose requirements: proposal for a new dosing regimen. Blood 106(7), 2329 2333 (2005). 24 Furuya H, Fernandez-Salguero P, Gregory W et al. Genetic polymorphism of CYP2C9 and its effect on warfarin maintenance dose requirement in patients undergoing anticoagulation therapy. Pharmacogenetics 5(6), 389 392 (1995). 25 Avery PJ, Jorgensen A, Hamberg AK, Wadelius M, Pirmohamed M, Kamali F. A proposal for an individualized pharmacogenetics-based warfarin initiation dose regimen for patients commencing anticoagulation therapy. Clin. Pharmacol. Ther. 90(5), 701 706 (2011). Pharmacogenetic prediction of loading doses. Used in the EU-PACT trial. 26 Linder MW, Bon Homme M, Reynolds KK et al. Interactive modeling for ongoing utility of pharmacogenetic diagnostic testing: application for warfarin therapy. Clin. Chem. 55(10), 1861 1868 (2009). 27 Bell RG, Sadowski JA, Matschiner JT. Mechanism of action of warfarin. Warfarin and metabolism of vitamin K 1. Biochemistry 11(10), 1959 1961 (1972). 28 Fregin A, Rost S, Wolz W, Krebsova A, Muller CR, Oldenburg J. Homozygosity mapping of a second gene locus for hereditary combined deficiency of vitamin K-dependent clotting factors to the centromeric region of chromosome 16. Blood 100(9), 3229 3232 (2002). 29 Li T, Chang CY, Jin DY, Lin PJ, Khvorova A, Stafford DW. Identification of the gene for vitamin K epoxide reductase. Nature 427(6974), 541 544 (2004). 30 Rost S, Fregin A, Ivaskevicius V et al. Mutations in VKORC1 cause warfarin resistance and multiple coagulation factor deficiency type 2. Nature 427(6974), 537 541 (2004). 31 D Andrea G, D Ambrosio RL, Di Perna P et al. A polymorphism in the VKORC1 gene is associated with an interindividual variability in the dose-anticoagulant effect of warfarin. Blood 105(2), 645 649 (2005). 32 Rieder MJ, Reiner AP, Gage BF et al. Effect of VKORC1 haplotypes on transcriptional regulation and warfarin dose. N. Engl. J. Med. 352(22), 2285 2293 (2005). 33 Wadelius M, Chen LY, Downes K et al. Common VKORC1 and GGCX polymorphisms associated with warfarin dose. Pharmacogenomics J. 5(4), 262 270 (2005). 34 Watzka M, Geisen C, Bevans CG et al. Thirteen novel VKORC1 mutations associated with oral anticoagulant resistance: insights into improved patient diagnosis and treatment. J. Thromb. Haemost. 9(1), 109 118 (2011). 35 Takeuchi F, Mcginnis R, Bourgeois S et al. A genome-wide association study confirms VKORC1, CYP2C9, and CYP4F2 as principal genetic determinants of warfarin dose. PLoS Genet. 5(3), e1000433 (2009). 36 Cooper GM, Johnson JA, Langaee TY et al. A genome-wide scan for common genetic variants with a large influence on warfarin maintenance dose. Blood 112(4), 1022 1027 (2008). 37 Caldwell MD, Awad T, Johnson JA et al. CYP4F2 genetic variant alters required warfarin dose. Blood 111(8), 4106 4112 (2008). 38 McDonald MG, Rieder MJ, Nakano M, Hsia CK, Rettie AE. CYP4F2 is a vitamin K1 oxidase: an explanation for altered warfarin dose in carriers of the V433M variant. Mol. Pharmacol. 75(6), 1337 1346 (2009). 39 Harrell FE. Regression Modeling Strategies : With Applications To Linear Models, Logistic Regression, And Survival Ana lysis. Springer, NY, USA (2001). 40 Shehab N, Sperling LS, Kegler SR, Budnitz DS. National estimates of emergency department visits for hemorrhage-related adverse events from clopidogrel plus aspirin and from warfarin. Arch. Intern. Med. 170(21), 1926 1933 (2010). www.futuremedicine.com 439