Cardiovascular Disease Risk Prediction With and Without Knowledge of Genetic Variation at Chromosome 9p21.3

The success of risk prediction models for cardiovascular disease reflects an increasing understanding of the molecular basis of atherothrombosis. Aside from age, which integrates many biological activities and environmental exposures at once, other important components of risk prediction include plasma biomarkers for lipid metabolism, inflammation, thrombosis, and metabolic status (1). Current prediction models, however, cannot anticipate many cases of incident cardiovascular disease, motivating both the identification of new risk factors and the optimization of analytic methods for their use. Traditionally aggregated by the notion of family history, individual genetic variants may represent a class of risk factors with new prognostic information; they may be more removed from underlying disease processes than plasma risk factors or other clinical variables but may also be more general. For example, a recent exploration of the influence of lipid-associated variation on prediction of incident cardiovascular disease found residual predictive value for the genetic variation after adjustment for plasma lipid levels (2).

Genetic variation at the chromosome 9p21.3 region is a good candidate for adding more information to risk prediction. Variation at this locus has consistently been associated with coronary artery disease (3–5) and diabetes (6–9). In addition, the risk allele is carried by almost 75% of the white population and the lack of correlation between any of the disease-associated 9p21.3 genetic variants and major cardiovascular risk factors suggests novel influences on disease progression. To assess whether knowledge of variation at 9p21.3 improves global risk prediction, we examined the effect of adding genetic information from a single nucleotide polymorphism (SNP) in the 9p21.3 region to previously published prediction models for the WHS (Women's Health Study), which has a large, prospective cohort of initially healthy U.S. women who were followed over 10 years for incident cardiovascular events.

Methods

Study Population

Study participants were members of the WGHS (Women's Genome Health Study) (10), an ongoing prospective genetic evaluation study being conducted among initially healthy U.S. women who enrolled in the WHS, a trial of aspirin and vitamin E for the primary prevention of cardiovascular disease and cancer in women 45 years or older (11). Beginning in 1992, the WHS recruited female health professionals in the United States who had no major chronic disease at baseline, including cancer and cardiovascular disease, and followed them prospectively for incident myocardial infarction, stroke, coronary revascularization, and cardiovascular death. The institutional review board of the Brigham and Women's Hospital, Boston, Massachusetts, approved the study.

Among the WHS participants, 28 345 provided blood samples that were stored in liquid nitrogen until the time of analysis, as well as consent for ongoing analyses linking blood-derived observations with baseline risk factor profiles and incident disease events. Of these women, 23 226 had standard risk factor information available and were genotyped for the rs10757274 polymorphism. To reduce the potential for population stratification to affect our results, we included only the 22 129 (95.3%) white women.

Risk Factor Ascertainment

Baseline information on age, diabetes, smoking status, parental history of myocardial infarction before 60 years of age, blood pressure, and hypertension treatment were collected at study initiation. Plasma biomarkers were analyzed in a core laboratory facility, certified by the National Heart, Lung, and Blood Institute/Centers for Disease Control and Prevention Lipid Standardization Program, for total cholesterol, high-density lipoprotein cholesterol, apolipoprotein B-100, apolipoprotein A-I, high-sensitivity C-reactive protein, lipoprotein(a), and hemoglobin A_1c.

We determined genotypes for rs10757274 in the WGHS participants by using an oligonucleotide ligation procedure that combined polymerase chain reaction amplification of target sequences from 3 ng of genomic DNA with subsequent allele-specific oligonucleotide ligation (12). The ligation products of the 2 alleles were separated by hybridization to product-specific oligonucleotides, each coupled to spectrally distinct Luminex100 xMAP microspheres (Luminex, Austin, Texas). The captured products were fluorescently labeled with streptavidin R–phycoerythrin (Prozyme, San Leandro, California), sorted on the basis of microsphere spectrum, and detected by a Luminex100 instrument (12).

Outcome Ascertainment

Study participants were followed through March 2004 for total cardiovascular disease, which comprised incident myocardial infarction, ischemic stroke, coronary revascularization, and cardiovascular deaths. An end-points committee adjudicated events by using medical record review. Morbidity data were available on nearly all the women through 8 years of follow-up.

Risk Prediction Models

To assess the effect of variation at rs10757274 on global cardiovascular disease risk prediction, we considered covariates from 2 nongenetic risk prediction models. The first included the covariates from the Third Report of the National Cholesterol Education Program Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults (ATP III) risk score, as well as history of diabetes (noted as a high-risk equivalent) (13). The second used the covariates from the Reynolds Risk Score (1), a model that includes additional biomarker information, as well as data on family history. To provide direct comparability and the highest level of internal validity, we elected on an a priori basis to model all covariates, including rs10757274, in the same base population rather than use β-coefficients for the nongenetic covariates derived from previously published models. In so doing, we avoided the potential for bias that might occur by modeling the effect of the genetic data within the test cohort but using estimates of effect for the other covariates from a second, unrelated cohort.

Statistical Analysis

We used Cox proportional hazard models to generate crude and adjusted hazard ratios across genotypes and to test for trend. We used models with separate effects for each genotype in all analyses and generated adjusted survival curves by stratifying Cox proportional hazard models by genotype. We also used Cox models to generate estimates of predicted risk with and without genotype information, which we then assessed for accuracy. We saw no evidence of departures from proportionality in any of the models used. Our primary measure of discrimination was the Harrell c-index (14), a generalization of the area under the receiver-operating characteristic curve that allows for censored data. The c-index assesses the ability of the risk score to rank women who develop incident cardiovascular disease higher than women who do not. We assessed general calibration across deciles of predicted risk by using the Hosmer–Lemeshow goodness-of-fit test (15) to compare the average predicted risk with the Kaplan–Meier risk estimate within each decile and considered a chi-square value of 20 or higher (P < 0.01) to be poor calibration (16).

We assessed risk reclassification (1, 17) by sorting the predicted 10-year risk for each model into 4 categories (<5%, 5% to <10%, 10% to <20%, and ≥20%). We then compared the assigned categories for a pair of models. For each pair, we calculated the proportion of participants who were reclassified by the comparison model versus the reference model; we considered reclassification to be correct if the Kaplan–Meier risk estimate for the reclassified group was closer to the comparison category than the reference. We computed the Hosmer–Lemeshow statistic for the reclassification tables (18), which assesses agreement between the Kaplan–Meier risk estimate and predicted risk within the reclassified categories. We also computed the Net Reclassification Improvement (19), which compares the shifts in reclassified categories by observed outcome, and the Integrated Discrimination Improvement (19), which directly compares the average difference in predicted risk for women who go on to develop cardiovascular disease with women who do not for the 2 models, on the women who were not censored before 8 years.

Role of the Funding Source

This study was supported by grants from the National Heart, Lung, and Blood Institute and National Cancer Institue, National Institutes of Health; the Donald W. Reynolds Foundation; and the Leducq Foundation. Additional support for DNA extraction, reagents, and data analysis was provided by Roche Diagnostics and Amgen. Genotyping of the 9p21.3 variant was performed by Celera. The funding sources had no role in the design, conduct, or reporting of this study or the decision to submit the manuscript for publication.

Results

Of the 22 129 white women genotyped for this analysis, 5793 (26.2%) had no risk (G) alleles at rs10757374, 10 952 (49.5%) had 1 risk allele, and 5384 (24.3%) had 2 risk alleles. The number of risk alleles had a significant association with family history of premature myocardial infarction (P < 0.001 for trend) and a modest association with a history of diabetes, driven by the higher proportion of diabetic patients among those with 2 risk alleles (P = 0.04 for trend) (Table 1). We observed no such associations for the other standard cardiovascular disease risk factors or for the newer biomarkers. In addition, we found no association between genotype and other lipid measures, including lipoprotein(a), apolipoprotein A-I, and apolipoprotein B-100 (data not shown).

The crude association between rs10757274 genotype and total cardiovascular disease was not affected by adjustment for age or additional risk factors (Table 2), and the cardiovascular hazard associated with an increasing number of risk alleles was consistent across all components of the primary study end point. Specifically, polymorphism at rs10757274 was associated with an adjusted hazard ratio (HR) with incident cardiovascular disease of 1.25 (95% CI, 1.04 to 1.51) for the AG genotype and 1.32 (CI, 1.07 to 1.63) for the GG genotype. We observed similar associations with less power for coronary heart disease (AG, 1.28 [CI, 1.02 to 1.62]; GG, 1.30 [CI, 1.00 to 1.69]), myocardial infarction (AG, 1.33 [CI, 0.93 to 1.91]; GG, 1.28 [CI, 0.85 to 1.92]), and stroke (AG, 1.33 [CI, 0.96 to 1.85]; GG, 1.54 [CI, 1.07 to 2.21]). The stratified mean survival curves (Figure) show attenuation in the absolute risk from the use of the mean covariate values in the adjusted models, but no change in the relative risk between groups.

• Download figure
• Download PowerPoint

Knowledge of genotype had little effect on the coefficients for the other risk factors in either the ATP III or Reynolds Risk Score covariates (Table 3). Table 3 also shows that the addition of genotype information to either the ATP III covariates or the Reynolds Risk Score covariates had no effect on discrimination (c-index, 0.803 to 0.805 and 0.807 to 0.809, respectively). All models were calibrated as measured by the Hosmer–Lemeshow chi-square test.

We found a modest improvement in reclassification when we added genotype to the ATP III covariates (Table 4), with an additional 606 (2.7%) of the women reclassified. Of those, 526 (86.9%) were reclassified correctly, corresponding to a Net Reclassification Improvement of 2.7% (P = 0.02) and an Integrated Discrimination Improvement of 0.001 (P = 0.11). Both the original ATP III covariates and the ATP III covariates with the genotype information remained calibrated within the reclassification table (reclassification Hosmer–Lemeshow chi-square, 17.9 and 14.9, respectively).

By contrast, when we added genotype information to the Reynolds Risk Score covariates, 585 (2.6%) of the women were reclassified. However, of those, only 214 (36.6%) were reclassified correctly, corresponding to a Net Reclassification Improvement score of −0.2% (P = 0.59) and an Integrated Discrimination Improvement score of 0.0 (P = 0.18). Both the original Reynolds Risk Score covariates and the Reynolds Risk Score covariates with the genotype information remained calibrated within the reclassification table (reclassification Hosmer–Lemeshow chi-square, 12.6 and 12.7, respectively). We observed almost identical results in analyses that added genotype information to the published Reynolds Risk Score, rather than the Reynolds Risk Score covariates.

Discussion

By using recently developed prediction models and performance measures, we examined the incremental contribution of 9p21.3 variation to cardiovascular risk among more than 22 000 initially healthy white participants in the Women's Genome Health Study. Similar to a previous report examining this same variation among white men (20), we find a strong and significant association of 9p21.3 variation with cardiovascular disease in white women. However, in our data, knowledge of this genetic variation only marginally improved the classification of risk prediction in a model based on ATP III covariates and did not improve classification in a model that included family history and C-reactive protein (Reynolds Risk Score covariates). In both settings, the addition of genotype information had no appreciable effect on the c-index.

The rs10757274 SNP is one of a cluster of tightly linked SNPs at 9p21.3 that are associated with coronary artery disease (3–5), myocardial infarction (3, 6), and stroke (21); abdominal aortic aneurysm (6); and intracranial aneurysm (6). The linkage disequilibrium block spanning these SNPs extends 28 kb from rs10757274 toward the telomere about 55.7 kb from the 3′ end of the CDKN2B transcript, although weaker linkage disequilibrium continues through the CDKN2A/CDKN2B gene region. The major transcript of a third gene, MTAP, corresponds with an entirely separate linkage disequilibrium block still further toward the telomere, but an alternatively spliced variant may be transcribed into the linkage disequilibrium block containing CDKN2A/CDKN2B. Toward the centromere, the linkage disequilibrium block that includes the associations with the arterial diseases extends 28.3 kb bases from rs10757274 and is juxtaposed with a smaller linkage disequilibrium block of about 6.9 kb that includes an SNP (rs10822661) associated with diabetes (6–9).

Although the mechanism of association between the rs10757274 genotype and cardiovascular disease remains unclear, the magnitude of association observed in our study of women is consistent with reported associations for 9p21.3 SNPs (5). Our data also extend previous work that demonstrated significant association with incident stroke (21). Because we could not find correlations between polymorphisms in rs10757274 and standard lipids, apolipoprotein B-100, apolipoprotein A-I, lipoprotein(a), or high-sensitivity C-reactive protein, our data provide further evidence that the effect of this genetic variation is unlikely to reflect genetic differences in lipid levels or biomarkers of hemostasis and inflammation.

Despite high prevalence of the risk allele and consistent association of rs10757274 with vascular events, inclusion of genotype information did not improve global risk prediction; instead, the addition of genotype information to a model that included C-reactive protein and family history of cardiovascular disease worsened classification. Thus, although knowledge of variation in chromosome 9p21.3 may yield important pathophysiologic insights into atherothrombosis, it seems unlikely that screening for SNPs in this region will be clinical useful in daily practice.

Improved cardiovascular risk prediction has clear potential for public health but has been difficult to achieve, and the potential role of genetic information is just beginning to be explored. Whether a precisely measured gene panel will prove superior to knowledge of family history of cardiovascular disease is uncertain, although combination scores have had promising effects on prediction (2). As shown in this paper, and as expected from statistical work by Pepe and colleagues (22) and Cook (17), single polymorphisms alone, even common ones with consistent modest associations, are unlikely to improve prediction.

Our study benefits from a large sample size and extensive prospective follow-up. However, because of the small numbers of nonwhite participants, our results are limited to white women. Further studies are needed in other populations, both of the association between the rs10757274 genotype and cardiovascular disease and the resultant effect on risk prediction. Our study also evaluated only variation at chromosome 9p21.3 and thus does not exclude the possibility that multigene panels might correctly reclassify a larger proportion of persons. Our study did not examine the additional effect of genotype relative to the previously published risk scores, but rather examined the additional effect after refitting all covariates in the same population. Although this comparison potentially limits the generalizability of our results, it preserves their validity.

In conclusion, in a large prospective cohort of white women, we confirmed the association of 9p21.3 variation with total cardiovascular disease, coronary heart disease, myocardial infarction, and stroke. However, adding 9p21.3 genetic variation did not improve discrimination or classification of the predicted risk achieved with traditional risk factors, C-reactive protein, and family history of cardiovascular disease.