Nonvalidation of Reported Genetic Risk Factors for Acute Coronary Syndrome in a Large-Scale Replication Study FREE

Compelling evidence from twin and epidemiological studies suggests a genetic basis for atherosclerotic heart disease and acute coronary syndromes (ACS), including unstable angina, non–ST-elevation myocardial infarction (NSTEMI), and ST-elevation myocardial infarction (STEMI).^1- 2 To date, numerous candidate genes have been implicated, mainly by case-control studies, as potential cardiovascular risk factors, but few, if any, have been established definitively.^3- 5 Factors undermining the validity of previous reports include inappropriately small sample sizes, multiple subgroup comparisons, and publication bias.⁴

Before use in clinical care, potential genetic risk factors would ideally be replicated en masse in large, well-characterized patient populations.⁶ To date, no such comprehensive validation of genetic variants potentially associated with ACS or atherosclerosis has been reported.

Accordingly, we first sought to identify genetic associations with ACS by systematically searching the medical literature for variants reported in association with MI, unstable angina, or atherosclerosis. We then attempted to validate these putative genetic risks in a large case-control study.

We searched PubMed and bibliographies of original and review articles for manuscripts published before March 10, 2005, that reported statistically significant associations between specific genotypes and coronary atherosclerosis or ACS (A list of the articles is available on request from the authors). MEDLINE search terms included: gene, genetic, polymorphism, myocardial infarction, atherosclerosis, coronary heart disease, and coronary artery disease. Reports were included if they contained a claim of a significant positive association, with an investigator-reported P value <.05. A total of 96 polymorphic genetic variants in 75 genes were identified and included (Table 1 and Table 2). Eleven of those were excluded because they had failed the multiplex genotyping assay.

Eight hundred eleven white patients of European ancestry with ACS were identified from a consecutive series of patients presenting at 2 Kansas City, Mo, hospitals (Mid-America Heart Institute and Truman Medical Center), from March 2001 through June 2003. Standard definitions were used to diagnose ACS patients with either MI or unstable angina.^92- 93 Myocardial infarction was defined by a positive troponin blood test in the setting of symptoms and electrocardiogram changes (both ST-segment elevation and non–ST-segment elevation changes) consistent with MI. Unstable angina diagnoses were confirmed, by concurrence of 3 physician chart reviewers, if patients had negative troponin blood tests and any one of the following: new onset angina (<2 months) of at least Canadian Cardiovascular Society Classification class III, prolonged (>20 minutes) rest angina, recent (<2 months) worsening of angina, or angina that occurred within 2 weeks of an MI.⁹³ Of the troponin-negative unstable angina patients, 203 (92.7%) had a cardiac catheterization, a nuclear stress test, or a stress echocardiogram to corroborate their diagnoses.

Each participating inpatient with ACS was interviewed to determine variables, such as smoking, alcohol use, family history (≥1 first-degree relatives with MI or coronary artery disease), and to obtain consent for a blood sample for genetic analysis. In addition, detailed chart abstractions were performed to collect relevant laboratory and clinical data.

A total of 1045 ACS patients (of which 811 white patients were included in the current study) agreed to participate and to provide a blood sample for genetic analysis. Patients self-reported their race/ethnicity by selecting one of the following descriptors that were provided by the investigators: white, white Hispanic, African American, and African American non-Hispanic. Age- and sex-matched controls were recruited from the ambulatory outpatient clinical laboratory of 1 of the centers, Saint Luke's Hospital of Kansas City. These patients were undergoing routine laboratory testing and were asked to complete a medical questionnaire defining cardiac risk factors and medical co-morbidities. Those controls reporting a previous ACS, prior coronary artery bypass graft surgery or prior percutaneous coronary intervention were excluded. To minimize the potential impact of genetic admixture, 650 white controls of mixed European ancestry who reported no history of coronary artery disease were selected from among the 1054 potential controls. Risk factor data were missing for 9 sex-, age-, and race-matched unaffected controls, and 56 additional matched controls were used for ALOX5AP haplotyping.

The research protocol was approved by the institutional review boards of both institutions; all study participants provided written informed consent for clinical and genetic studies.

Genomic DNA was isolated (Gentra PUREGENE, Minneapolis, Minn) from blood samples and subjected to whole genome amplification by multiple-strand displacement (Molecular Staging Inc, New Haven, Conn), using random priming and Phi-29 polymerase.^94- 95 Genotyping was performed using the Sequenom MALDI-TOF (Matrix Assisted Laser Desorption-Ionization Time-of-Flight) system, using Spectrodesign software for assay design (Sequenom, San Diego, Calif), and assay methods that have previously been described.^96- 97 Gene variants were excluded from analysis if they could not be genotyped using the Sequenom system due to persistent assay failure, defined as less than 95% scorable genotypes after 4 multiplex reaction design cycles. Eleven assays were ultimately excluded.^{13,20,42,45,72,88,98- 100} For the rare MEF2A 21–base pair (bp) deletion, cases and controls were genotyped by polymerase chain reaction to generate amplicons of 152-bp nondeletion or 131-bp deletion followed by electrophoresis on 3% agarose gels. Identified deletions were confirmed by direct DNA sequencing. Due to its rarity, MEF2A was analyzed separately, and thus only the other 84 genes were subjected to the full set of statistical analyses. PHASE Version 2.1 was used to estimate haplotype frequencies for ALOX5AP.^101- 102

Genotype distributions in cases and controls were examined for significant deviation (P<.05) from Hardy-Weinberg equilibrium. The number of departures was assessed by Monte Carlo simulation and compared with the number expected by chance alone (Resampling Stats Inc, College Park, Md).

In the primary analysis, each genetic variant was prespecified based on published reports, and the frequencies of risk-associated variants were compared in cases and controls by using a 100 000 iteration Monte Carlo extension of the χ² test (SPSS 13.0 Exact Tests, SPSS Inc, Chicago, Ill). The term statistically significant was reserved for a P value below the Bonferroni-corrected study-wide significance threshold (0.05/84 = 0.0006). Because the Bonferroni correction is conservative when applied to a replication study, the total number of all positive associations at the P<.05 level was also compared with the expected number by chance in 100 000 simulations. A surplus of positive associations over random expectations would imply that some are truly associated with ACS.

Secondarily, we also compared the overall genotype distributions at each locus in cases and controls by Monte Carlo χ² testing. Power to confirm individual genetic associations was determined using a log-likelihood-based method (Quanto 1.0).^103- 104

Finally, as a measure to increase power, the observed proportion of prespecified risk variants found to be even marginally more frequent in cases than in controls was assessed by the Sign test. Under the null hypothesis, each of the risk variants is equally likely to be more frequent in cases, or in controls. To estimate the Sign test's power to detect an excess of even weakly positive genetic associations (50 of 84 positive associations confers P = .05 in the Sign test), we simulated the resampling of 650 control and 811 case genotypes across 84 genetic comparisons, finding the minimum detectable odds ratio ensuring a critical probability level of a 63.3% win rate for each 84 risk variants that provides 80% confidence of having at least 50 wins.

The clinical characteristics of the 811 cases and 650 controls are described in Table 3 and the distributions of their genotypes are shown in Table 2. The population of ACS cases included 308 (38%) STEMI, 284 (35%) NSTEMI, and 219 (27%) unstable angina patients. Cases and controls had similar age, sex, and body mass index distributions. A family history of coronary artery disease or MI among first-degree relatives was 2.7-fold higher in male cases than in male controls and 2.0-fold higher in female cases than in female controls. Male and female cases were significantly more likely to be current smokers and to have type 2 diabetes mellitus but less likely to consume at least 1 alcoholic drink per month. Frequencies of hypercholesterolemia and hypertension were higher in female cases than in controls; no significant differences were observed in males. Previous revascularization had been performed in 35.6% of incident ACS cases and in none of the controls.

A total of 85 variants in 70 genes were genotyped in cases and controls.The overall genotype call rate for these variants was 98.5% (range, 95.0%-99.8%). Two percent of all samples were genotyped in duplicate for each marker in a blinded fashion as a measure of genotype reproducibility. Among the 2511 repeated genotypes, 5 were discordant, demonstrating a reproducibility of 99.8%.

Tests of Hardy-Weinberg equilibrium revealed that 1 variant violated it in both cases and controls, at the P<.05 level; 7 violated it in cases only; and 4 violated it in controls only (Table 1 and Table 2). This finding is not more than expected by chance (4 violations expected by chance in each group; see the Methods section) and therefore none was excluded from further analysis at this stage.

With respect to power parameters, the mean effective frequency (or 1-frequency, if q >0.5) in controls of the putative risk variants studied was 0.20, and 58 (68.2%) were common, (≥0.1), 25 (29.4%) were uncommon (<0.1; >0.01), and 2 (2.4%) were rare (≤0.01). Our sample had 80% power to confirm, by the Monte Carlo χ² test, a genotype-specific relative risk of 2.3 for a rare variant (q = 0.01), 1.4 for a relatively uncommon variant (q = 0.1), and 1.25 for a common allele (q = 0.5).

We tested whether each putative risk variant showed a significant difference in frequency between cases and controls (Table 1). An odds ratio greater than 1 indicates that the risk genotype was in higher frequency among cases, and if so, the genotype frequency difference was reported as a positive decimal number. Only 1 genetic variant was significant at the P<.05 level, which is the number most likely by chance alone. The −455 variant, which lies upstream of the transcription initiation site in the β-fibrinogen gene, replicated the originally reported association, with the GG genotype being more frequent in cases than controls (frequency, 66% in cases vs 61% in controls; odds ratio, 1.27; P = .03). In addition, we found the MEF2A 21-bp deletion in 1 case and 1 control, confirming that this is a rare variant in the population.¹⁰⁵

Several supplementary analyses were performed. When the genotypes of cases and controls were analyzed by extension of 2 × 3 χ² tests to 100 000 simulations, 4 loci, RECQL2, THBS2, LIPC, and p22-PHOX, were marginally significant (Table 2). In each case, the specific genetic risk model providing significance was different from that reported in the literature; hence, these cannot be considered formal replications and the total number of positive associations is not in excess of random expectations.

Finally, we found that only 41 of 84 predefined risk variants were even marginally more frequent in cases than in controls (excluding 1 tie, the rare MEF2A deletion), representing a 48.8% win rate (95% confidence interval, 38.1%-59.5%) for the collective-risk genotypes. This observed proportion of wins is not different from the expected proportion (50%) under the null hypothesis (P = .91). Table 1 shows that the absolute differences in risk genotype frequencies between cases and controls (negative signs meaning that the putative risk genotype was more frequent in controls than in cases) were small, with a median difference of −0.0003, and maximum of 0.056 (β fibrinogen).

We were unable to confirm as risk factors for ACS 85 genetic variants because none was unequivocally validated in this large case-control study of 1461 participants. In the primary analysis, only the −455 promoter variant in β-fibrinogen) was nominally statistically significant (P = .03). Among the 4 variants in the secondary analysis that met nominal statistical thresholds, there was an excess of a different variant than was previously reported among cases in the original study, which does not support replication. We therefore conclude that our findings, in this large sample of well-characterized ACS patients and controls, cannot support that this panel of gene variants contains bona fide ACS risk factors.

Our findings come at a critical juncture in complex disease genetics. Some cardiovascular gene variants (eg, ACE, AGT, AGTR1, ITGB3, F2, F5, MTHFR) included in our study can already be ordered clinically, for indications that explicitly include possible ACS risk. However, our findings suggest that such clinical genetic testing is premature and underscore the importance of robust replication studies of reported associations prior to their application to clinical care.

These nonreplications include variants in several high-profile studies. For example, haplotypes A and B of 5-lipoxygenase activating protein (ALOX5AP) were reported in 1 study to be associated with MI in the general populations of Iceland, and the United Kingdom, respectively.¹⁷ We found neither haplotype was associated with ACS, in spite of our observed haplotype frequencies in cases and controls closely approximating those found in the total United Kingdom data set (cases and controls) previously (haplotype A, 0.165 vs 0.160, respectively; haplotype B, 0.062 vs 0.058).

Although our study raises significant doubts about the collective panel of putative genetic risk factors, it does not invalidate any particular previous study. Possible explanations of our negative results could include: (1) false-negative results in our study; (2) false-positive associations in previous studies; and (3) varied effects of risk variants in different genetic backgrounds.

False-negative results as a general explanation for our study's null findings are unlikely given that our sample size is substantially larger than all but a few reported prior studies and was powered to detect modest relative risks. Based on a random sample (n = 30) of articles included in this study (1 per gene variant), we estimated that the mean odds ratio reported in positive studies was 2.3 (range, 1.25-5.0), indicating that we had well in excess of 80% power to replicate most reports. However, isolated positive reports may overestimate genetic risks.^5- 6 Recently, a meta-analysis of 14 genes included in our study reported odds ratios ranging from 1.10 to 1.73 for risk of MI.³ It is possible that minute odds ratios are to be expected in complex disease genetics and that neither our study nor most previous studies were sufficiently powered. Accordingly, we augmented our power, by use of the Sign test, to detect a surplus of as few as 16 weakly positive genetic risk factors among the entire set that we genotyped (84 −16 = 50, the number required for a significant Sign test), corresponding to a mean odds ratio of 1.05 or higher given our sample size and the average risk genotype frequency.

Absence of genetic effect only in our cohort is also unlikely. Cases showed a 2-fold higher family history of ACS, consistent with a genetic effect contributing to phenotypes in this cohort. In addition, homozygosity coding for an arginine residue at position 158 of apolipoprotein E (E4 variant), considered 1 of the least controversial of the putative ACS susceptibility factors despite some inconsistency in certain cohorts,¹⁰⁶ was significantly associated (P = .04) among cases with hyperlipidemia (4.1%) vs controls without hyperlipidemia (1.6%).

False-positive results in previous studies are another potential explanation for the discrepancy between our findings and those of others. This issue has previously been recognized as a serious problem with association studies, particularly when sample sizes are underpowered.¹⁰⁷ It is difficult to identify true vs false positives by analysis of the literature alone.¹⁰⁸ Unrecognized stratification between cases and controls can create spurious associations,¹⁰⁹ and the absence of negative genomic controls in nearly all prior studies to exclude this possibility leaves this an open question. Also difficult to assess is the extent to which publication bias and multiple hypothesis testing have had an effect.

It could be argued that our research participants are distinct from those reported previously and that our results may not bear on the validity of positive associations reported in different populations and clinical subgroups (eg, analyses substratified by age, sex, or a clinical variable, such as hypertension, hyperlipidemia, or smoking status). Given that the vast majority of common variants in the human genome date to our shared ancestry in Africa,¹¹⁰ it is not likely that there are different common functional variants in linkage disequilibrium with risk variants in our population vs others. Less common mutations of more recent ancestral origin could conceivably be correlated with certain genetic variants in one population but not another. The extent to which linkage disequilibrium patterns might explain our findings is unknown, but our study population is quite typical of the mixed European background that is prevalent in the United States.

Another possibility is that the effect of risk variants is different in different genetic backgrounds; if true, the lack of generalizability of results will severely limit their application to the clinical arena. The fact that we failed to replicate positive associations in a consecutive series of study participants that are broadly representative of the disease encountered in clinical practice places limitations on the potential applicability of prior findings and supports our premise that it is premature to extrapolate these earlier findings to routine clinical care.

The failure of the candidate gene approach to identify variants conferring susceptibility to ACS risk prompts consideration of other approaches. One promising approach is to screen the entire genome in an unbiased way in a large sample for variants that are significantly associated with disease risk. Coupled with the understanding of underlying patterns of linkage disequilibrium in the human genome⁷ and the ability to inexpensively obtain genotypes across the genome, the field is moving rapidly toward a comprehensive genome-wide approach. Challenges of this approach include the unknown number of variants that impart effect, the magnitude of the effect imparted by each, and the extent to which common variants as opposed to rare independent mutations account for disease risk.

Regardless of the approach taken, it is clear that multiple large, well-matched cohorts of cases and controls will be required to achieve valid progress in the genetic analysis of ACS and other complex human diseases. Our null findings indicate the need for caution in the interpretation of genetic associations in different clinical populations and the need for extensive validation of genetic risk factors.