Predicting Clopidogrel Response Using DNA Samples Linked to an Electronic Health Record

Abstract

Variants in ABCB1 and CYP2C19 have been identified as predictors of cardiac events during clopidogrel therapy initiated after myocardial infarction (MI) or percutaneous coronary intervention (PCI). In addition, PON1 has recently been associated with stent thrombosis. The reported effects of these variants have not yet been replicated in a real-world setting. We used BioVU, the Vanderbilt DNA repository linked to de-identified electronic health records (EHRs), to find data on patients who were on clopidogrel treatment after an MI and/or a PCI; among these, we identified those who had experienced one or more recurrent cardiac events while on treatment (cases, n = 225) and those who had not experienced any cardiac event while on treatment (controls, n = 468). We found that CYP2C19*2 (hazard ratio (HR) 1.54, 95% confidence interval (CI) 1.16–2.06, P = 0.003) and ABCB1 (hr 1.28, 95% CI 1.04–1.57, P = 0.018), but not PON1 (HR 0.91, 95% CI 0.73–1.12, P = 0.370), were associated with recurrent events. In this population, genetic signals for clopidogrel resistance in ABCB1 and CYP2C19 were replicated, supporting the use of EHRs for pharmacogenomic studies. Our data do not show an association between PON1 and recurrent cardiovascular events.

Current guidelines recommend dual antiplatelet therapy to be initiated after a recent myocardial infarction (MI) or intracoronary stent placement to prevent future adverse cardiac events. Aspirin in conjunction with clopidogrel is the most frequently prescribed combination.^1–6 Clopidogrel is a prodrug, bioactivated predominately by CYP2C19.^7,8 A loss-of-function polymorphism in CYP2C19, CYP2C19*2 (rs4244285), results in the clopidogrel-poor-metabolizer trait and has been associated with reduced on-drug platelet inhibition and an increased risk of adverse cardiac events, particularly stent thrombosis.^9–13 These results, from studies conducted in registries and large randomized clinical trials, have prompted the US Food and Drug Administration to include in the clopidogrel label a black-box warning that poor metabolizers merit consideration for alternative therapies. This risk appears to be present for patients who are either heterozygous or homozygous for the CYP2C19*2 loss-of-function allele.^14,15

Other genetic polymorphisms have been implicated in decreased efficacy of clopidogrel therapy. CYP2C19*17 (rs12248560) encodes a protein with increased function as compared to CYP2C19*1 and has been associated with both improved efficacy of clopidogrel and increased risk of bleeding.^16,17ABCB1 encodes the P-glycoprotein efflux transporter that is expressed at the apical membrane of intestinal mucosa (among other locations), and the C3435T (rs1045642) polymorphism has been associated with reduced bioavailability.¹⁸ Clinically, this variant has been associated with an increase in the rate of cardiac events during clopidogrel therapy.^11,19,20 The common Paraoxinase-1 (PON1) polymorphism, Q192R (rs662), which converts 2-oxo-clopidogrel to the active thiol metabolite, has also recently been associated with stent thrombosis; however, this finding remains controversial.^21,22 PON1 is believed to result in a lower rate of bioactivation of the compound and therefore to translate into an increased risk of stent thrombosis after percutaneous coronary intervention (PCI).²¹

Genetic associations such as those described above have been generally studied in clinical trial or registry settings. The extent to which findings in these research settings can be translated to real-world clinical care settings is uncertain. Biobanks linking DNA to electronic health records (EHRs) offer an approach to addressing this issue.^23–26 To date, such EHR-based studies have focused on disease susceptibility or physiologic traits, but their utility for pharmacogenomics has not been evaluated. The Vanderbilt University Medical Center EHR includes comprehensive information on ∼1.7 million patients; as of June 2011, BioVU, the Vanderbilt DNA biobank²⁷ included de-identified DNA samples from more than 120,000 individuals. From the BioVU database we identified individuals who were started on clopidogrel therapy after an MI and/or PCI with stent placement and then assigned “case” or “control” status depending on the occurrence of a recurrent cardiac event during a 1- to 2-year follow-up. We report here that the associations of adverse cardiac events with variants in CYP2C19 and ABCB1 were readily replicated, whereas the association with PON1 was not.

Results

Through the implementation of an electronic algorithm followed by manual review, we identified 807 individuals who had been started on clopidogrel therapy after an MI and/or intracoronary stent placement procedure (Figure 1). The primary outcome was a recurrent cardiovascular event, defined as a composite of all-cause mortality, MI, revascularization, and stroke. This primary outcome occurred in 260 individuals, and, after biobank opt-out and inadequate DNA samples were taken into account, 225 underwent genotyping. Among the genotyped cases, there were 14 deaths, 40 MIs (9 ST segment elevation MI, 28 non–ST segment elevation MI, and 3 MIs not categorized), 206 revascularization procedures, and 1 stroke. In addition, stent thrombosis occurred in 12 individuals. We identified 547 individuals with no second composite cardiac event, of whom 468 underwent genotyping. Of the controls, 76.7% had 2 years of follow-up, and the remaining 23.3% had at least 1 year of follow-up.

Figure 1.

Phenotype algorithm for selection of cases and controls for genotyping. CPT, Current Procedural Terminology; ICD, International Classification of Disease; MI, myocardial infarction; NLP, natural language processing.

Cases and controls were similar in age, gender, body mass index, and cardiac risk factors, including hypertension, diabetes mellitus, hyperlipidemia, and smoking status (Table 1). The overall rates of hypertension, hyperlipidemia, and positive smoking status were higher than in previously reported studies. Cases and controls were also similar in their indication for study enrollment.

Table 1.

study population characteristics

	Controls	Cases
	N = 468	N = 225
Demographics
Age (years) (± SD)	69 ± 12	67 ± 11
Gender (% male)	62.8	64.9
Race (%)
European American	88	88.4
African American	9.4	10.7
Other	2.6	0.9
Body mass index (kg/m2) (± SD)	30.2 ± 6.6	30.5 ± 6.0
Past medical history (%)
Hypertension	84.6	72.9
Diabetes	34.8	34.8
Hyperlipidemia	95.5	87.1
Smoking
Current	16.4	15.6
Ever	47.1	38.7
Never	36.4	45.8
Heart failure	13.7	19.1
Primary event (%)
MI	4.1	6.7
Stent	67.9	70.2
Both	28	23.1
Number of stents (%)	N = 449	N = 210
1 stent	65.2	56.6
>1 stent	32.7	43.3
Unknown	1.8	0.0
Stent type
Bare metal	33.6	21.4
Drug-eluting	60.8	67.1
Mixed	2.9	3.3
Unknown	2.7	8.1
Stent locationsa
LMCA	1.1	3.3
LAD	45.7	33.8
LCX	22.0	25.7
RCA	39.6	37.6
Graft	2.4	15.7

LAD, left anterior descending coronary artery; LCX, left circumflex coronary artery; LMCA, left main coronary artery; MI, myocardial infarction; RCA, right coronary artery.

^aPercentages total more than 100% because patients may receive more than one intervention.

We performed genotyping for six single-nucleotide polymorphisms (SNPs): rs1045642, ABCB1; rs4244285, CYP2C19*2; rs4986893, CYP2C19*3; rs28399504, CYP2C19*4; rs12248560, CYP2C19*17; and rs662, PON1. Two SNPs—rs28399504 (CYP2C19*4) and rs4986893 (CYP2C19*3)—were excluded from further analysis because of the very low frequency of the minor alleles. The genotype frequencies by case status for four SNPs—rs1045642 (ABCB1), rs4244285 (CYP2C19*2), rs12248560 (CYP2C19*17), and rs662 (PON1)—are available in Supplementary Table S1 online.

Kaplan–Meier curves (Figure 2) of the primary end point for CYP2C19 among cases and controls showed that carriers of at least one CYP2C19 loss-of-function allele (*1/*2 or *2/*2) had higher event rates than those without this allele (P = 0.005). For ABCB1, the recurrent cardiac event rate was greater among A/A genotypes as compared to G/G genotypes (P = 0.065). Neither the CYP2C19*17 nor the PON1 genotype affected the time to second cardiac event in dominant, recessive, or additive genetic models.

Figure 2.

Kaplan–Meier curves for time to recurrent cardiovascular event. (a)CYP2C19*2, (b)ABCB1, (c)CYP2C19*17, (d)PON1.

Unadjusted and adjusted results for the primary analysis are shown in Table 2 for all individuals. According to the adjusted estimates, CYP2C19*2 and ABCB1 were significantly associated with the primary end point, with hazard ratios (HR) equal to 1.54 (95% CI 1.16–2.06, P = 0.003) and 1.28 (1.04–1.57, P = 0.018), respectively. Individuals who were CYP2C19*2 heterozygotes were also at increased risk as compared to homozygote normal metabolizers, the HR being 1.61 (1.20–2.17, P = 0.001). There were 15 individuals who were CYP2C19*2 homozygotes, and no increased risk compared to normal metabolizers was detected in this small group. There was no evidence to suggest an effect of CYP2C19*17 or PON1 on the primary outcome measure: HRs were 0.91 (0.71–1.18, P = 0.48) for CYP2C19*17 and 0.91 (0.73– 1.12, P = 0.37) for PON1. For both CYP2C19*2 and ABCB1, the primary end point was driven largely by repeat revascularization, the HR values being 1.54 (1.14–2.07, P = 0.004) and 1.21 (0.99–1.47, P = 0.067), respectively (Table 3). For the secondary end points (death, MI, and stent thrombosis considered individually), none of the SNPs was found to have a significant association. However, ABCB1 approached significance for stent thrombosis, the HR being 2.64 (0.84–8.29, P = 0.096), and CYP2C19*2 approached significance for MI, the HR being 2.02 (0.99–4.12, P = 0.054).

Table 2.

Unadjusted and adjusted tests of association for ABCB1, CYP2C19*2, CYP2C19*17, and PON1

					Unadjusted			Adjusted
Gene	SNP	Chr	Loc	Coded allele	HR	95% CI	P	HR	95% CI	P
ABCB1	rs1045642	7	86976581	A	1.20	0.99–1.46	0.061	1.28	1.04–1.57	0.018
CYP2C19*2	rs4244285	10	96531606	A	1.50	1.13–2.00	0.0049	1.54	1.16–2.06	0.003
CYP2C19*17	rs12248560	10	96511647	T	0.88	0.68–1.13	0.308	0.91	0.71–1.18	0.481
PON1	rs662	7	94775382	G	0.93	0.76–1.13	0.466	0.91	0.73–1.12	0.370

Analysis performed using Cox proportional hazards model. The adjusted model was stratified by gender and race, and adjusted for age, smoking status, and body mass index. Chr, chromosome; CI, confidence interval; HR, hazard ratio; Loc, location.

Table 3.

Test of association for secondary end points for ABCB1, CYP2C19*2, CYP2C19*17, and PON1

	HR	95% CI	P
Stent thrombosis (n = 12)
ABCB1	2.64	0.84–8.29	0.096
CYP2C19*2	2.79	0.70–11.16	0.147
CYP2C19*17	0.27	0.04–2.09	0.210
PON1	0.77	0.26–2.22	0.622
Revascularization (n = 206)
ABCB1	1.21	0.99–1.14	0.067
CYP2C19*2	1.54	1.14–2.07	0.004
CYP2C19*17	0.86	0.66–1.11	0.247
PON1	0.93	0.75–1.14	0.470
Myocardial infarction (n = 40)
ABCB1	1.33	0.80–2.22	0.270
CYP2C19*2	2.02	0.99–4.12	0.054
CYP2C19*17	0.63	0.31–1.29	0.207
PON1	1.03	0.63–1.69	0.906
Death (n = 14)
ABCB1	1.70	0.78–3.70	0.178
CYP2C19*2	1.76	0.57–5.37	0.323
CYP2C19*17	1.62	0.69–3.81	0.267
PON1	0.57	0.23–1.44	0.231

All results report unadjusted Cox proportional hazard models. CI, confidence interval; HR, hazard ratio.

Supplementary Table S2 online shows the unadjusted and adjusted analyses individually by race. In European Americans, CYP2C19*2 and ABCB1 were significantly associated with the primary end point in both adjusted and unadjusted analyses, with adjusted HR equal to 1.44 (95% CI 1.05–1.97, P = 0.03) and 1.27 (1.03–1.57, P = 0.029), respectively. In African Americans, CYP2C19*2 was significantly associated with the primary outcome in the unadjusted analysis only, with an HR of 2.40 (1.04– 5.55, P = 0.04).

Discussion

This study validates the association of CYP2C19*2 (rs4244285) and ABCB1 3435 G>A (rs1045642) polymorphisms with recurrent cardiovascular events in patients on clopidogrel therapy. Because these associations were observed in a de-identified cohort compiled from an EHR populated with real-world phenotypic and treatment information, the clinical relevance of these associations, previously observed in controlled trials, is bolstered. This study, along with ongoing work with warfarin,²⁸ is among the first to demonstrate the use of EHR data to replicate and discover pharmacogenomic associations.

Multiple studies have examined the associations of CYP2C19*2 and ABCB1 G>A with clopidogrel resistance. These studies have included registry trials of patients who have had an MI or undergone PCI, cohort studies with those receiving clopidogrel for stent thrombosis or PCI, post hoc analysis of randomized control trials, and a meta-analysis.^{9,11–15,20,29,30} For the primary end point (a composite of all-cause mortality, MI, revascularization, and stroke), our results are within the ranges previously reported for both CYP2C19*2 and ABCB1. Through the use of an electronic database, both study recruitment and follow-up can be acquired instantaneously, thereby decreasing the time, personnel, and cost required to perform a study. Both the cases and the controls were identified from the same population, reducing a selection bias between the groups. However, the use of the EHR for such studies does limit the outcomes assessed to those recorded within the EHR and to hard end points such as revascularization status from cardiac catheterization records. We were unable to replicate the association of CYP2C19*2 with stent thrombosis that has been reported in previous studies,^14,29 possibly because of insufficient power due to the low numbers of cases. Because the indication for cardiac catheterization could not always be reliably assessed from EHR data, our data may underestimate the number of stent thromboses. In addition, given the nature of our study, the angiograms could not be reviewed in our de-identified cases to ascertain the validity of the diagnoses of stent thrombosis in line with the Academic Research Consortium criteria.³¹ However, the rate of stent thrombosis in our population was 1.7% over 2 years, which was within the range of 0.5–1% per year reported in previous studies.³² Future accrual of larger numbers of patients may yield adequately powered studies on stent thrombosis.

Bouman et al. recently reported a significantly increased risk of stent thrombosis associated with PON1 in a small cohort of 40 patients showing a gene–dose effect dependent on the number of major alleles.²¹ However, when a larger cohort of patients with stent thrombosis was examined, this relationship was not replicated.²² In addition, a previous genome-wide association study of clopidogrel's effect on adenosine diphosphate–induced platelet aggregation readily detected the CYP2C19*2 signal but failed to identify an association with PON1.³⁰

Although our study was not powered to detect an association with stent thrombosis, we would have expected a trend toward association between PON1 and either the composite outcome or stent thrombosis if a strong association with stent thrombosis does indeed exist. Moreover, these PON1 variants have also been associated with adverse cardiac events independent of clopidogrel therapy, further weakening the association of PON1 and adverse cardiac outcomes in the context of clopidogrel efficacy.^33,34

There are limitations to the results of the work we report here. Compliance with prescribed medication regimens is not readily assessable in the EHR. In our study, we used evidence of medication refills and visits with providers as evidence of ongoing adherence to prescribed medications. The strength of the signal in an EHR-based study may be reduced relative to randomized controlled trials that are characterized by strict medication adherence measurements such as pill counts. The fact that our findings are consistent with clinical trial data despite this limitation lends further support to the importance of clopidogrel pharmacogenomics as applied to clinical practice. Finally, because evidence of death status was obtained primarily through linkage to the Social Security Death Index, we were unable to assess cause of death and therefore could not ascertain the incidence of cardiovascular-related death.

In 2007, more than 25 million patients were prescribed clopidogrel.³⁵ The demonstration of the relationship between recurrent cardiac events during clopidogrel therapy and common variants in CYP2C19*2 and ABCB1 as obtained from EHR data improves the generalizability of clinical trial results, highlighting the possibility that a large proportion of “real-world” clopidogrel users are at risk for treatment failure.³⁶

The argument has been made that randomized clinical trials should be conducted to establish the optimal approach to therapy in patients with these variants; however, we believe that the totality of the available data to date make such a trial difficult to mount because of potential lack of clinical equipoise. In the absence of data from randomized clinical trials, we advocate personalized approaches to therapy. These may include platelet-function testing, early genotyping for patients likely to require clopidogrel, or point-of-care genotyping at the time of initiation of clopidogrel treatment.³⁷

Methods

BioVU database and subject selection

Cases and controls were identified from the BioVU DNA biobank as described previously.^23,24,27 In brief, BioVU links DNA extracted from blood samples (obtained during routine clinical care and about to be discarded) to a de-identified image of the Vanderbilt EHR, termed the Synthetic Derivative. Each DNA sample and associated de-identified health record is linked by a unique identifier generated by a one-way hash function.²⁷ The resource has been and continues to be reviewed by the institutional review board; it has been considered as containing data for nonhuman subjects in accordance with the provisions of Title 45 of the Code of Federal Regulations part 46, as have the individual research studies utilizing the resource. Exclusion criteria include insufficient or poor-quality DNA, patient opt-out, and absence of signed consent to treat. In the Vanderbilt EHR, the race parameter is administratively assigned; we have previously shown that this is highly concordant with race defined by ancestry-informative markers.³⁸

The selection of records for inclusion was a two-part process: an electronic algorithm was employed to identify possible cases and controls, followed by a manual review of all “possible” records by three physicians (Figure 2). Individuals who presented with an MI and/or intracoronary stent placement and who were subsequently placed on clopidogrel before discharge were identified through a combination of International Classification of Disease, 9th edition (ICD-9), and Current Procedural Terminology codes, laboratory values, medication orders, and natural-language processing in physician notes.

The primary outcome was defined as a composite of MI, revascularization, stroke, and/or all-cause mortality within a 2-year period starting 2 days after the initial event, while the patient was on clopidogrel treatment. Revascularization was defined as either intracoronary angioplasty or coronary bypass surgery. For the purpose of ascertaining all-cause mortality, all records in the Synthetic Derivative have been linked to the Social Security Death Index.²⁶ In addition, in-hospital deaths are documented in the EHR and transferred into the Synthetic Derivative. Controls were defined as those with an absence of the primary outcome while on clopidogrel treatment for up to 2 years of observation after the primary event. In addition, we required all controls to have outpatient visits, with clinical documentation being available for at least 1 year after the primary event. The electronic algorithm was then used to assign the individual's status as possible case or possible control, and a manual review was employed to determine the validity of the status assignment. Covariates for age, gender, and body mass index were obtained through a combination of ICD-9 and Current Procedural Terminology codes, natural-language processing, and manual review.

Genotyping and data analysis

Prior to genotyping, the data for 114 subjects were removed from the study because of inadequate DNA concentration or opt-out from the DNA biobank. The Vanderbilt DNA Resources Core performed genotyping on 693 samples, using TaqMan (Applied Biosystems, Foster City, CA). Genotyping was performed for six SNPs: rs1045642, ABCB1; rs4244285, CYP2C19*2; rs4986893, CYP2C19*3; rs28399504, CYP2C19*4; rs12248560, CYP2C19*17; and rs662, PON1. Quality-control analysis was assessed with the PLINK, version 1.07, http://pngu.mgh.harvard.edu/∼purcell/plink/ and Plato software packages, version 1.1 https://chgr.mc.vanderbilt.edu/plato.^39,40 SNP call rates were >95% for all six SNPs, with a concordance of >98% between duplicates and HapMap samples. One individual's sample was excluded because of poor genotyping efficiency. For all samples, the allele frequencies and tests of Hardy–Weinberg equilibrium were calculated. All SNPs were in Hardy–Weinberg equilibrium for European Americans, and all SNPs except rs662 were in Hardy–Weinberg equilibrium in African Americans. Four SNPs passed the quality-control criteria and were taken up for statistical analysis. The two SNPs excluded from further analysis were rs28399504 (CYP2C19*4), the minor allele of which was observed in only five Asian subjects, and rs4986893 (CYP2C19*3), which was monomorphic.

statistical analyses

Descriptive statistics were calculated for patient characteristics. Continuous variables were summarized as mean values and standard deviations, and categorical variables were expressed in percentages. The cumulative incidence of the primary event (stent thrombosis, revascularization, MI, or death) was estimated using the Kaplan–Meier product limit estimator, and unadjusted comparisons between target SNP–defined groups were compared using the log-rank test. Cox proportional hazards models, stratified by gender and race, were used to examine the effect of genotype on the primary event. Stratification was required because of violations to the proportional hazards assumption. We assumed an additive genetic model for rs1045642 (ABCB1), rs12248560 (CYP2C19*17), and rs662 (PON1). On the basis of information from the literature, we assumed a dominant genetic model for rs4244285 (CYP2C19*2) for the primary analysis.^12,13 Prespecified covariates included age, gender, race (European American, African American, and other/unknown), BMI, and smoking status (current, ever, or never). BMI was imputed for 34 subjects in whom the values were missing, using race, height, and/or weight where available. Secondary end points were all-cause mortality, MI, revascularization, and stent thrombosis (definite and probable, as defined by ARC criteria).³¹ Statistical analyses were performed using the R programming language.