Association Between 9p21 Genomic Markers and Heart Disease:  A Meta-analysis FREE

Preventing and managing cardiovascular disease (CVD) presents a challenge for health care and public health.^1- 2Quiz Ref IDNonmodifiable risk factors include increasing age, male sex, and heredity. Modifiable risk factors include smoking, hypertension, dyslipidemia, obesity, physical inactivity, and diabetes.^3- 5 Among men, the annual rate of initial CVD events increases from 3 to 74 per 1000 from ages 35 to 44 years to ages 85 to 94 years, respectively. Similar increases occur among women a decade later.⁶ Biomarkers (eg, C-reactive protein) have been combined with traditional risk factors to predict CVD events,⁷ and molecular markers hold further promise. In 2007, genome-wide association studies on CVD identified a series of associated single-nucleotide polymorphisms (SNPs) in an intergenic region of chromosome 9p21, near the CDKN2A (NM_000077) and CDKN2B (NM_004936) genes.^8- 9

Currently, no comprehensive compilation of the 9p21 literature uses formal methods to estimate the strength of the association with heart disease (eg, effect size, heterogeneity, publication bias, credibility of cumulative evidence) and examine clinical utility. This analysis is part of a targeted systematic review on existing cardiogenomic panels that included 28 genes in addition to the 9p21 SNPs. That review was commissioned by the Evaluation of Genomic Applications in Practice and Prevention (EGAPP) initiative and overseen by the EGAPP Working Group (EWG).¹⁰

The review team included an experienced consultant to the EWG (G.E.P.) contracted by the Office of Public Health Genomics (OPHG) at the Centers for Disease Control and Prevention, Atlanta, Georgia, an OPHG researcher (S.M.), and a clinical geneticist (L.A.B.). Standard methods for evaluating clinical validity included systematic literature searches, preset inclusion/exclusion criteria, data abstraction, meta-analysis, and grading of studies and cumulative evidence.^11- 13 Expert guidance on literature searches and analytic methods was provided by the Technical Expert Panel. Electronic searches used the HuGE Navigator (version 1.3)^14- 15 with search term (9p21[Text+Mesh]). Previous validations had cross-checked HuGE Navigator and PubMed¹⁶ search results for selected gene-disease associations and found the HuGE Navigator searches equally sensitive but more specific (data available from the author on request). Limiting key questions and truncating search strategies are 2 common methods in rapid reviews¹⁷ that we chose to use. The Web site of the laboratory offering 9p21 testing was reviewed for references or gray data. Reference lists of included articles were hand searched. Two team members (G.E.P., S.M.) reviewed included articles on clinical validity and extracted the raw data and demographic information into spreadsheets; discrepancies were resolved through discussion.

Included articles were published in English; contained primary outcomes of coronary heart disease (CHD), myocardial infarction (MI), or coronary artery disease (CAD); tested for 9p21 SNPs; reported race and numbers of affected and unaffected participants; and provided the odds ratio (OR) with confidence intervals (CIs) or data sufficient to compute it. If more than 1 outcome was reported, the best-described phenotype was chosen (eg, MI rather than CAD). The authors' definition of phenotype was not an exclusion criterion. Excluded articles reported only on stroke, intermediate outcomes (eg, atherosclerosis), or subgroups (eg, patients with diabetes). Several included articles reported consortium results with multiple independent populations. These populations were listed as separate data sets.

Not all researchers use the same 9p21 SNPs, and most articles reported results for multiple SNPs (uniquely identified by their rs number). We extracted data for all SNPs used by at least 2 of the 22 included articles, but we report herein 3 common SNPs (rs1333049, rs10757274, and rs2383207) that were included in all but 3 articles.^18- 20 These SNPs are in high linkage disequilibrium (D′ = 1.00; r²>0.85).^8,21- 25 The remaining 3 articles used 2 additional SNPs, rs2891168²⁰ and rs10757278.^18- 19 These were also in high linkage disequilibrium.¹⁹ One additional SNP (rs2383206) was reported in at least 2 included articles. When possible, we restricted results in each data set to a single race. Information about age at diagnosis and race/ethnicity was also collected.

As previously reported,^26- 27 individuals with 1 at-risk SNP allele (heterozygotes) were designated as the reference group. Heterozygotes comprise about 50% of the white population. For each data set, the observed genotype frequencies in controls were compared with expected frequencies based on Hardy-Weinberg equilibrium (χ² test with 2 degrees of freedom). All P values are 2-sided at the P = .05 level.

Summary ORs and corresponding 95% CIs were derived (by reanalysis when possible) and summarized using random-effects modeling weighted by each data set's total variance (Comprehensive Meta-Analysis, version 2, Biostat Inc, Englewood, New Jersey).²⁸ Subgroup differences were compared using the Q test for heterogeneity for each covariate separately. A fixed-effects meta-regression was performed for the OR vs age at diagnosis.²⁸ Studies that did not report a value for any covariate were excluded. Publication bias was examined by performing a cumulative effects analysis.²⁸ Wider ranges of these summary ORs indicate potential for publication bias.

Quality of individual studies (levels 1-4) and overall quality of evidence for clinical validity (convincing, adequate, or inadequate) were evaluated with EWG methods,¹¹ consistent with other grading systems.^29- 30 Assessment of cumulative evidence used a consensus evaluation guideline (Venice criteria) specific to gene-disease association studies¹² and focused on amount of evidence, replication of evidence, and protection from bias. Cumulative evidence was strong for 3 “A” grades, moderate for “B” grades but no “C” grades, and weak if it received at least 1 “C” grade.

The net reclassification index (NRI)³¹ was used as an intermediate measure of potential clinical utility. To compute the NRI, the cases (events) and controls (nonevents) in a population were cross-classified by risk assessment using traditional factors with and without 9p21 SNP results. The proportions of all cases reclassified into higher (correct) or lower (incorrect) risk categories were then computed and added together. A similar computation was made for controls. The sum of the 2 proportions (expressed as a percentage) is the NRI, with positive results indicating improved risk prediction.

Twenty-five citations were identified through the original literature search through January 2009; none were meta-analyses. Ten citations were excluded: 2 related to other diseases,^32- 33 3 addressed subgroups of the population,^34- 36 3 addressed intermediate outcomes,^37- 39 and 2 addressed stroke.^40- 41 The remaining 15 articles underwent a full review along with 4 additional citations identified through hand-searching reference lists.^{20,25,42- 43} Three of these 19 articles were excluded: 2 due to inadequate or inconsistent data^44- 45 and 1 had availability of only an electronic abstract at that time.⁴⁶ The 16 remaining articles were included.^{8- 9,20- 25,42- 43,47- 52} In December 2009, the search was repeated using the same methods, and 6 additional articles satisfied the inclusion criteria.^{18- 19,46,53- 55} The analysis is based on data sets from all 22. In 1 article,⁴⁶ 9 data sets were reported; only 5 were included. The remaining 4 data sets were excluded because at least some data from these sets were in published studies already included in our analysis (written communication, Nilesh J. Samani, MD, January 15, 2010). Figure 1 provides a summary of the literature review.

Forty-seven distinct data sets were analyzed; most were case-control studies. Observed genotype frequencies were available from 33 data sets (70%), and all but one²³ satisfied Hardy-Weinberg criteria (eTable 1). Consensus genotype frequencies in controls were 27% (range, 22%-36%), 50% (range, 47%-57%), and 23% (range, 17%-30%) for 0, 1, and 2 at-risk alleles, respectively. To demonstrate that the choice of SNP is relatively unimportant, we compared the ORs using 4 SNPs that the first 2 articles^24,47 listed in the Table had in common (eFigure 1). Although the average ORs are different between the 2 articles (1.78 and 1.26), they are very similar within (eg, 4 SNPs⁴⁷ provide nearly identical ORs of 1.78, 1.67, 1.75, and 1.72). This is expected, given the high linkage disequilibrium, and justifies reporting on a single SNP per data set.

Using a random-effects model, the summary OR was 1.25 (95% CI, 1.21-1.29, P < .001). Heterogeneity was low (Q = 51; I² = 10%; P = .29). Figure 2 presents these data. The ORs were stratified by race, 9p21 SNP tested, heart disease outcome, and age cutoff at diagnosis for cases. No differences in ORs were found between Asians (8 data sets) and whites (39 data sets) (ORs, 1.32 and 1.24, respectively; P = .17); the SNPs used (18 data sets with rs1333049, 17 with rs10757274, 6 with rs2383207, 4 with rs2891168, and 2 with rs10757278) (ORs, 1.23, 1.24, 1.28, 1.29, and 1.27, respectively; P = .75); or the outcomes of CAD (19 data sets), MI (17 data sets), or CHD (9 data sets) (ORs, 1.27, 1.23, and 1.21, respectively; P = .45).

Finally, we evaluated age at CAD diagnosis. Some articles reported only on early onset disease^47,51 or included data sets restricted to early onset disease.^{9,23,42,46- 47} Others enrolled cases in wider age ranges.^25,48 Some provided an upper age cutoff (eg, MI occurring by age 50 years) while others provided mean age (and standard deviation). Missing age cutoffs were estimated using information from 11 data sets that provided both (Table).^{8- 9,25,42,50- 51} Fourteen data sets not reporting this covariate were excluded.^{8- 9,18- 19,43,46,49,53- 54} A meta-regression performed on the remaining 33 ORs vs the upper age cutoff level showed a significant association between higher ORs at earlier ages of disease onset (P < .001; slope and intercept of −0.00558 and 0.60881) (Figure 3). The regressed OR was 1.35 (95% CI, 1.30-1.40) for an upper age cutoff of 55 years and 1.21 (95% CI, 1.16-1.25) for 75 years; remaining heterogeneity was reduced (Q = 29; I² = 0%; P = .58). Consistent results were found using the mean age at onset (eFigure 2). One way to estimate the OR for all adults is to use the 14 data sets with upper age cutoff levels of greater than 70 years. In this group, the summary OR was 1.19 (95% CI, 1.13-1.25; P < .001), with low heterogeneity (Q = 13; I² = 1%; P = .44).

The summary OR was 0.80 (95% CI, 0.77-0.82; P < .001), with moderate heterogeneity (Q = 65; I² = 29%; P = .04) (Figure 4). No differences in ORs were found between Asians and whites (ORs, 0.81 and 0.79, respectively; P = .79); the SNP used (ORs, 0.82, 0.77, 0.84, 0.78, and 0.70; P = .22); or the outcomes of CAD, CHD, and MI (ORs, 0.79, 0.79, and 0.81; P = .87). Meta-regression again showed a significant slope (P = .001), with lower ORs associated with earlier ages of onset (slope and intercept of 0.00500 and −0.57854) (Figure 5). The regressed OR was 0.74 (95% CI, 0.71-0.77) f or an upper age cutoff of 55 years and 0.82 (95% CI, 0.79-0.85) for 75 years; heterogeneity was lower after fitting the model (Q = 40; I² = 22%; P = .39). Results were similar when mean age was used (eFigure 2). When restricted to the 14 data sets with an upper age cutoff levels of 70 years or higher, the summary OR was 0.83 (95% CI, 0.78-0.89; P < .001). However, heterogeneity was moderate (Q = 21; I² = 37%; P = .08). With the most discrepant finding removed (OR, 1.16),⁴⁸ this summary OR became 0.82 (95% CI, 0.78-0.87) and heterogeneity was reduced (Q = 13; I² = 7%; P = .37).

Several included articles reported the effect size as an allele-specific OR,^{8- 9,21- 22,46,49- 51,53} where an equivalent “dose” of risk was conferred per allele. That would imply that the reciprocal of our low OR should be similar to the high OR (eg, 1/0.80 = 1.25; the high OR is 1.25).

Three cohort studies (Table) were deemed level 1 designs and the remaining case-control studies were level 2.¹¹ Based on this and internal validity assessments, the overall quality of evidence for clinical validity was convincing to adequate. After accounting for age at onset, the Venice grading¹² was A for amount of evidence (>1000 cases/controls with the least common genotype) and A for replication (low heterogeneity after meta-regression and for the 12 data sets with later ages of onset). Bias required examining phenotype definition, genotyping, population stratification, and selective reporting using the predefined “typical” effect size category (ORs, 1.15-1.80).¹² The use of widely agreed-on definitions for CAD, CHD, and MI (eg, CHD defined as requiring coronary artery bypass graft surgery) in the data sets made the likelihood of phenotype bias low or none. Nearly all data sets were tested using commercial or well-described genotyping platforms, making the likelihood of genotyping bias low. Most case-control studies matched on race/ethnicity, and, where possible, we restricted results to a single race. We assigned the likelihood of stratification bias to be low, resulting in an overall protection from bias grade of A.

Data sets restricted to early onset heart disease are likely to be smaller in size (since these events are uncommon) and have a larger effect size (Figure 3). This might appear to be positive publication bias. To avoid confounding, each of the high ORs (Table) was divided by the expected OR given that data set's age at onset (Figure 3). These age-adjusted ORs (median, 1.000) could then be evaluated for publication bias. The cumulative effect size analysis indicated a trivial 4% change from the most precise to the least precise cumulative estimates (0.96 for the 6 largest data sets to 1.00 for all 33 data sets).

The OR associated with an individual's 9p21 test result can be used to modify heart disease risk based on traditional factors. For example, a 65-year-old man with no other traditional risk factors would have a 10-year heart disease risk of 11%,⁵⁶ while a 40-year-old woman with no other risk factors would have a 10-year risk of 2%. If both were to have 2 at-risk 9p21 alleles, risk estimates would increase to about 13.2% (11 × 1.2) and 2.4% (2 × 1.2), respectively, compared with an individual with 1 at-risk allele. Having no at-risk alleles would result in reduced risk estimates of 9.2% and 1.7%, respectively, compared with an individual with 1 at-risk allele. Comparing no at-risk alleles with 2 at-risk alleles could modify risk estimates by 1.44-fold.

Three studies that included 1721 cases and 34 797 controls also provided information to compute the NRI.^25,43,57 Four NRI estimates were made because 1 study⁴³ reported 2 risk algorithms. For the 4 data sets, the proportions of cases reclassified were 0.5%, 0.7%, 2.5%, and −0.1% (P = .65, .79, .03, and .87), respectively; proportions of controls reclassified were 0.3%, 4.2%, −0.1%, and 0% (P = .36, <.001, .52, and .87), respectively. The corresponding NRIs were 0.8%, 4.9%, 2.5%, and −0.2% (P = .51, .09, .03, and .96), respectively. Details are given in eTable 2.

Based on EWG and Venice criteria, evidence for the association between heart disease and the 9p21 SNP markers has strong credibility. The 9p21 markers have been identified through genome-wide association studies that were not hypothesis-driven^8- 9,51 and appear independent of traditional risk factors or family history.^24,57 The pathophysiological mechanism is not yet understood.^24- 25,57 Risk alleles are more strongly associated with heart disease events in younger persons (OR, 1.35) than with heart disease in general (OR, 1.21). One company offering 9p21 SNP testing and interpretation uses an OR of 1.3 in its clinical reports for “MI at any age” and 1.6 for “MI at a younger age.”²⁷

Most data sets are limited to white populations, usually of European descent, so the results for other racial/ethnic groups might differ. In 8 studies among Asians, however, the association appears similar. We did not include data for black populations, but limited evidence suggests a smaller effect.^9,23 Rather than adjusting ORs for age/sex, we based our OR on raw data when possible, because adjusted ORs appear to be close to raw estimates. For example, 1 article⁴³ reported per-allele hazard ratios that were 1.14 when unadjusted, 1.15 after adjustment for age, and 1.12 after adjustment for age, blood pressure, lipid measurements, smoking status, diabetes, and antihypertensive use. These differences were not statistically significant. Finally, we removed 2 articles from the meta-analysis for which we could not generate reliable ORs. Both were small (202 cases of CAD/MI⁴⁴ and 118 individuals with MI/coronary insufficiency)⁴⁵ and inclusion would be unlikely to influence the results reported using larger sample sets.

A test has clinical utility when its results lead to a measurable improvement in health outcomes. Intervention trials to establish the clinical utility of adding 9p21 testing were lacking.^58- 60 There is an expectation that genetic susceptibility information could increase motivation for long-term lifestyle changes (eg, improvement in risk-reducing behaviors, treatment adherence).^61- 62 However, measuring behavioral change is challenging. Quiz Ref IDCommunicating genetic information to patients has resulted in encouraging reports of short-term positive effects^63- 64 and has not shown reduced adherence or fatalistic thinking.^64- 65

Achievement of more accurate risk by adding 9p21 to traditional risk factors could be considered an intermediate measure of clinical utility. The computed NRIs ranged from a 0.2% decrease to a 4.9% improvement. The study showing the largest NRI²⁵ achieved most of the risk reclassification because of reduced risk in individuals without events and, therefore, would have little chance of improving outcomes. The study reporting 2 traditional risk factor models⁴³ showed that adding 9p21 testing to the Adult Treatment Panel III model showed significant improvement (NRI, 2.5%), while adding testing to the Reynolds Risk Score (RRS) showed no improvement (NRI, −0.2%). One main difference between the 2 models was inclusion of a family history of myocardial infarction in the RRS. None of the 3 studies modified treatment protocols based on addition of 9p21 or examined long-term health or behavioral outcomes. Quiz Ref IDHow best to analyze the effect of adding genomic markers to traditional risk factors (eg, area under the curve, NRI, integrated discrimination improvement) remains an active research area.^31,66- 68

The incidence of MI can be reduced by drugs that lower cholesterol/low-density lipoproteins and blood pressure in individuals with risk factors.⁶⁹ Improved risk assessment might influence decision making about effective interventions and behavioral change. However, only 37% of US physicians reported regular use of a heart disease risk score.⁷⁰Quiz Ref IDA systematic review found preliminary evidence that CHD risk scores may translate into modest benefits (eg, increased drug treatment, short-term blood pressure reduction) without clinical harms.⁷¹ However, the need for higher-quality evidence on long-term outcomes, and for replication of the results in different clinical settings, was emphasized.⁷¹ Other complicating factors may be patient adherence to lipid-lowering medications (about half reach target lipid levels and one-quarter continue long-term drug treatment)⁷² and access to medical care and medications. Therefore, the clinical utility of adding 9p21 markers to traditional risk factors cannot be assumed, even if risk assessment is improved. One proposal suggests that heart disease may be more effectively prevented by implementing an inexpensive, standardized, multidrug intervention (ie, the polypill) in all persons aged 55 years or older, regardless of individual risk levels.⁶⁹

In summary, showing that a genetic test has clinical validity does not necessarily lead to improved health. Quiz Ref IDClinical trials need to demonstrate that use of the test is associated with changes in physician management decisions, patient motivation and long-term behavioral changes, improved health outcomes, and/or reduced costs to the health care system. Using genomic tests to improve existing risk models would likely require the inclusion of many markers like 9p21.^73- 74 Such risk assessments may be more accurate but may not result in more appropriate treatments until the underlying mechanisms are known. Uncovering these mechanisms may provide insights into new or improved treatments and prevention activities.