Coronary Heart Disease Risk Prediction in the Era of Genome-Wide Association Studies

Abstract

HomeCirculationVol. 121, No. 20Coronary Heart Disease Risk Prediction in the Era of Genome-Wide Association Studies Free AccessResearch ArticlePDF/EPUBAboutView PDFView EPUBSections ToolsAdd to favoritesDownload citationsTrack citationsPermissions ShareShare onFacebookTwitterLinked InMendeleyReddit Jump toFree AccessResearch ArticlePDF/EPUBCoronary Heart Disease Risk Prediction in the Era of Genome-Wide Association StudiesCurrent Status and What the Future Holds Steve E. Humphries, Fotios Drenos, Gie Ken-Dror and Philippa J. Talmud Steve E. HumphriesSteve E. Humphries From the Centre for Cardiovascular Genetics (S.E.H., F.D., G.K.-D., P.J.T.), British Heart Foundation Laboratories, Department of Medicine, University College London, London, United Kingdom, and UCL Genetics Institute (S.E.H.), Department of Genetics Environment and Evolution, University College London, London, United Kingdom. Search for more papers by this author , Fotios DrenosFotios Drenos From the Centre for Cardiovascular Genetics (S.E.H., F.D., G.K.-D., P.J.T.), British Heart Foundation Laboratories, Department of Medicine, University College London, London, United Kingdom, and UCL Genetics Institute (S.E.H.), Department of Genetics Environment and Evolution, University College London, London, United Kingdom. Search for more papers by this author , Gie Ken-DrorGie Ken-Dror From the Centre for Cardiovascular Genetics (S.E.H., F.D., G.K.-D., P.J.T.), British Heart Foundation Laboratories, Department of Medicine, University College London, London, United Kingdom, and UCL Genetics Institute (S.E.H.), Department of Genetics Environment and Evolution, University College London, London, United Kingdom. Search for more papers by this author and Philippa J. TalmudPhilippa J. Talmud From the Centre for Cardiovascular Genetics (S.E.H., F.D., G.K.-D., P.J.T.), British Heart Foundation Laboratories, Department of Medicine, University College London, London, United Kingdom, and UCL Genetics Institute (S.E.H.), Department of Genetics Environment and Evolution, University College London, London, United Kingdom. Search for more papers by this author Originally published25 May 2010https://doi.org/10.1161/CIRCULATIONAHA.109.914192Circulation. 2010;121:2235–2248For DNA-based tests that assess genetic predisposition to coronary heart disease (CHD) to be of clinical value, they need to provide information over and above conventional risk factors (CRFs) currently used in risk algorithms, such as the Framingham Risk Score,1 which incorporates CRFs such as age, gender, blood lipid concentrations, blood pressure, body mass index, family history, and smoking habit. To achieve this, several hurdles must be passed.The first challenge is to identify a set of common single-nucleotide polymorphisms (SNPs) at loci associated with CHD risk. Over the last 10 to 15 years, this has been done by use of a “candidate gene” approach through association studies in prospective analysis or case-control studies, ie, comparing SNP genotype or allele frequency between groups of individuals with CHD and healthy subjects. Several of the genes, chosen because of their key role in processes that predispose to atherosclerosis, have meta-analysis–confirmed effects on risk of CHD,2 the best example of which is the APOE gene, which encodes apolipoprotein E, with 3 common isoforms that are associated with strong effects on plasma lipids and more modest effects on risk of CHD.3 This “hypothesis-driven” search for useful genetic variants provides the foundation for the development of genetic CHD risk profiles, and in the last 2 years, it has been enhanced by technical advances that have allowed “hypothesis-free” genome-wide association studies (GWASs), primarily in a case-control setting. Although the list of identified CHD-risk loci and SNPs will clearly grow, we have at least the basis to start the examination of their potential clinical utility.The second set of challenges is to obtain a robust estimate of the size of the risk effects associated with these SNPs. This requires population-based prospective studies to avoid bias, because estimates in the case-control setting, although efficient for gene discovery, are a suboptimal design for the evaluation of predictive performance of a marker. In addition, information is needed on the risk-allele prevalence between countries and by race and ethnicity, as well as any differences in risk-effect size in these different groups and whether the effect is modified by gender or by the presence of other genes or environmental factors (ie, the context dependence of the effect). To achieve this, genotyping of large data sets will be required, and robust estimates will require that data be combined from several different studies.Third, the clinical utility of adding these genetic risk scores to the CRF algorithms must be examined, in most cases by use of a simplistic additive model, and the most appropriate clinical setting for its application must be explored given concerns about the psychological impact of DNA testing and confidentiality issues. The final set of challenges, given that the SNPs in the currently identified loci do not represent the full heritability estimate for CHD risk, involves determining how newly emerging data from post-GWAS research can be incorporated into risk algorithms. These research areas include the ability to identify rare or private mutations and the possible utility of measuring telomere length, specific copy number variations (CNVs), or epigenetic effects that result from DNA methylation.Already, some commercial companies are offering CHD risk genotyping over the Internet, based on the success of the GWASs; for the public to have access to this information is, we believe, premature, and in isolation from CRFs, it has little value. There are still few solid data with regard to stages 2 and 3, and the potential clinical utility of post-GWAS research, outlined in stage 4, is still very unclear. In this review, we examine the progress made in tackling these 4 challenges and discuss the research pathways that may make genetic testing for CHD potentially more likely; the issue of confidentiality and motivation associated with DNA testing is addressed briefly.Candidate-Gene SNPs Singly and in Combination for CHD Risk Prediction: Modeling StudiesThe candidate-gene approach to validate potential CHD risk genes is based on genes chosen because of their involvement in CHD-related (patho)physiology and metabolism, and although this approach has been questioned since the development of GWASs,4 in studies that are adequately powered, it still has validity. However, the use of single SNPs, underpowered studies, population stratification, and lack of replication has led to inconsistencies and poor reproducibility of results, which rightly resulted in candidate-gene studies getting “bad press.” As expected with any single genotype in a multifactorial disease like CHD, published risk estimates have been modest, in the range of 1.12 to 1.75 (reviewed in Casas et al2). Meta-analyses provide the possibility of more robust odds ratios (ORs), as illustrated by the meta-analyses of 15 SNPs in 12 genes that incorporated data from as many as 53 studies.5 The summary ORs of these meta-analyses were similar, ranging from 0.8 to 1.34. Meta-analyses are also not without bias, including publication bias, population stratification, lack of genotype blinding, control selection bias, and genotype errors, so that even with significant summary ORs, results should be viewed with caution.6 However, given statistically robust proof of CHD risk, the question arises of how best to use the genotype information in risk algorithms. One approach is to utilize the ORs from meta-analyses to derive a genetic risk function using a binomial distribution for the risk coefficient of each gene, weighted for the frequency of each risk allele. To illustrate this approach, we recently constructed an in silico risk model using 11 SNPs in 10 candidate genes (APOB, eNOS, APOE, ACE, PAI1, MTHFR, GPIIb-IIIa, PON1, LPL, and CETP) and the published (white) allele frequencies and predicted summary risk estimates from meta-analyses.7 Compared with those with 3 or 4 risk-associated alleles (50% of individuals), those with 6-allele (8.3%) and ≥7-allele (3%) risk genotypes had a significantly higher mean OR for CHD risk (OR 1.70, 95% confidence interval [CI] 1.14 to 2.25, and OR 4.51, 95% CI 2.89 to 7.04, respectively; see Drenos et al7 and the online-only Data Supplement). Put another way, compared with those in the lowest decile of genetic risk, those in the highest decile had an OR of developing CHD in the range of 3.05 (95% CI 2.32 to 4.14). When age and the risk alleles carried were taken into account, the mean 10-year probability of developing CHD for a 55-year-old man in the lowest decile of “genetic risk” was estimated to be 10% (95% CI 8.5% to 11.4%), whereas those in the 9th and 10th deciles had risk >20%. These results support the view that in combination, common SNPs with a modest impact on risk will have clinical utility, but it is evident that this modeled group of SNPs needs to be augmented by others, and it is unknown whether these will contribute to risk over and above the effect of CRFs. Space limitations preclude a comprehensive review of this field, but in a recent modeling paper8 that used the same 10 candidate genes described above, it was reported that although the discriminative power of these SNPs alone was poor, as evaluated by the area under the receiver operating characteristic curve (AROC; AROC=0.59), it would improve to 0.69 and 0.76, respectively, if an additional 40 or 90 SNPs of similar allele frequency and ORs were available. Because these values are similar to or greater than those achieved by current CHD prediction algorithms, these data support the view that risk prediction improvement will be achievable if enough genetic variants can be identified.Candidate-Gene SNPs in Combination: The Risk-Score ApproachThe alternative approach to combining SNP information is simply to assign a risk value of, for example, 0 if a subject is a noncarrier of a risk allele, 0.5 or 1 if a carrier, and 1 or 2 if homozygous for that allele, and then to calculate the overall score for each individual.9,10 Using this method and 12 SNPs in a similar group of genes as reported above,7 Yiannakouris et al9 observed that in a subset of the Greek component of the European Prospective Investigation Into Cancer and Nutrition (EPIC) case-control study, the mean gene score was higher in case subjects than in control subjects (P<0.002), and the OR for myocardial infarction associated with a score ≧3.0 (54.3% of the control subjects) was 1.55 (95% CI 1.02 to 2.37).Using this gene-score approach, Kathiresan et al10 determined 9 candidate SNP genotypes, 6 in genes involved in determining levels of low-density lipoprotein (LDL) cholesterol (LDL-C; APOB, APOE, HMGCR, LDLR, PCSK9, and ABCA1) and 3 involved in high-density lipoprotein (HDL) cholesterol (CETP, LIPC, and LPL), with an individual’s score ranging from 0 to 18 for LDL raising or HDL lowering alleles. These were examined in 2246 men with 137 CHD events and 3168 women with 101 events who participated in the Malmö Prospective CHD study and who were followed up for 10.6 years.10 As shown in Figure 1, the mean LDL/HDL ratio varied from a CHD-protective value of 2.5 in those with >6 risk alleles (2.8% of total) to a CHD risk ratio value of 3.4 in those with >13 risk alleles (5.4% of total). This was associated with a steep gradient of CHD risk from 2.7 to 11 CHD events per 1000 patient-years. Kathiresan et al10 then examined the ability of the SNPs to discriminate case subjects from control subjects. The AROC value of the CRFs was not improved significantly by the addition of the gene score, but risk classification did improve. Of those at intermediate risk (9%), 26% moved to a higher or lower risk category (improvement P=0.01). Interestingly, this effect remained significant after adjustment for baseline lipids, with an adjusted hazard ratio per unfavorable allele of 1.15 (95% CI 1.07 to 1.24, P=0.0003). What this study suggests is that a single lipid measure (that is included in a CRF algorithm) might be a poor estimator of an individual’s risk, whereas an individual’s genetic profile may better predict their lifetime exposure to high lipids and might thus provide risk prediction over and above the lipid measure itself. Such a concept is strongly supported by data on PCSK9, for which a single loss-of-function SNP (Arg47Leu) has been shown to result in 15% lower plasma LDL-C levels, presumably due to less degradation of LDL receptors and thus more functional receptors in the liver, which results in faster LDL-C clearance from the blood. This reduction in LDL-C would be predicted to result in a lower CHD risk of ≈23%, but the observed reduction in carriers was 47%.11 Similarly, in patients with mutations that cause familial hypercholesterolemia, the observed CHD risk is considerably greater than that predicted simply on the basis of plasma LDL-C levels, because of the lifetime accumulated LDL-C burden of such patients.12Download figureDownload PowerPointFigure 1. Graph of observed LDL/HDL ratio and CHD event rate in subjects from the Malmö Prospective CHD Study. For LDL-C, trend P<10−18; for HDL cholesterol, trend P<10−24. Adapted from Kathiresan et al.10 HR indicates hazard ratio; pyears, person-years.Identification of New Candidates for Inclusion in Risk AlgorithmsIn the last few years, the results from several GWASs have identified new loci involved in determination of the risk of CHD, which has attracted huge interest. The chromosome 9p21 (chr9p21) locus associated with CHD and myocardial infarction risk was the first of these, identified simultaneously in 3 GWASs published in 2007.13–15The original GWASs identified 2 “lead” SNPs, rs10757274 and rs1333049, and many subsequent studies have genotyped only 1 of these, which makes a direct comparison problematic. A meta-analysis of all data published to date is presented in Figure 2A16–21 for rs10757274 and Figure 2B22–26 for rs1333049 (the search strategy is detailed in the online-only Data Supplement). The 2 SNPs are in strong linkage disequilibrium (estimated r2 from data available in the report by Kathiresan et al27 is 0.88), and as can be seen in Figure 2A and 2B, both SNPs are strongly associated with CHD in white people, with similar effect sizes in case-control and prospective studies. As expected for any biomarker of modest effect, not all studies demonstrated a statistically significant effect owing to issues of small sample size and power and the play of chance, although effects may also have been modified by different study characteristics, such as the prevalence of smoking or other “environmental” CHD risk factors. However, there is no significant evidence for heterogeneity of effect for either SNP, with very similar overall per-allele CHD risk effects of 1.29 (95% CI 1.19 to 1.40) for rs10757274 and 1.29 (95% CI 1.24 to 1.35) for rs1333049. It is possible that the use of both or additional SNPs at this locus may refine and improve the identification of the “risk haplotype,”25 but further data are required to confirm this. Download figureDownload PowerPointFigure 2. Meta-analysis per allele ORs for CHD risk for (A) rs10757274 and (B) rs1333049. The linkage disequilibrium between the SNPs was estimated from data in Samani et al26 and kindly made available by the authors as D′=0.96 and r2=0.88 in unrelated subjects (Table 5 in Meng et al27). C, Cartoon of the chr9p21.3 locus showing the nearest genes (MTAP, ANRIL, CDKN2A and CDKN2B) and the location of the 2 GWAS-identified SNPs, rs10757274 and rs1333049, reported in the meta-analyses, with a Haploview linkage disequilibrium plot of the region around the 2 SNPs. ARIC indicates Atherosclerosis Risk In Communities; OHS1, OHS2, and OHS3, Ottawa Heart Study 1, 2, and 3; CCHS, Copenhagen City Heart Study; DHS, Dallas Heart Study; FH, familial hypercholesterolemia; WGHS, Women’s Genome Health Study; WTCCC, Wellcome Trust Case Control Consortium; MI, myocardial infarction; GerMIFS II, German Myocardial Infarction Family Study II; UK MI, United Kingdom Myocardial Infarction Study; MONICA/KORA, Monitoring of Trends and Determinants in Cardiovascular Disease/Cooperative Health Research in the Region Augsburg; PRIME, Prospective Epidemiological Study of Myocardial Infarction; and AMC-PAS, Academic Medical Center Amsterdam Premature Atherosclerosis Study.GWASs have identified several other CHD loci, and their chromosome locations, the frequency of the risk allele, and the reported size of the risk effect are shown in Table 1. Interestingly, although the CELSR2/SORT1/PSRC1 locus is associated with LDL-C levels,30,31 of which variation in SORT1 appears to be the most likely candidate,32 other loci show no association with measured phenotypes, and it appears that their mechanism of risk does not operate by influencing known CRF traits such as lipids or blood pressure; thus, the addition of SNPs such as these to the Framingham risk-score algorithm has the potential to improve its overall utility. Table 1. List of Other Chromosomal Localizations Associated With CHD or MI From GWAS That Have Been Replicated in More Than 1 GWASChromosome Localization/GeneSNP and Risk GenotypeRisk Allele FrequencySize of the Effect (95% CI)Overall PReferenceMI indicates myocardial infarction.Data include original discovery study and replication studies.*From overall analysis in 12 713 case subjects and 12 821 control subjects.28†From overall analysis in 11 550 case subjects and 11 205 control subjects.26‡From overall analysis in 19 407 case subjects and 21 366 control subjects.291p13.3 CELSR2/PCSR1/SORT1rs646776–T*0.811.19 (1.13–1.26)7.9×10−1228rs599839-A†0.281.13 (1.08–1.19)1.4×10−7261q41 MIA3rs17465637-C*0.721.14 (1.10–1.19)1.4×10−928rs3008621-G†0.161.10 (1.04–1.17)1.0×10−3262q36rs2943634-C†0.341.05 (1.0–1.1)0.03263q22.3 MRASrs9818870-T‡0.201.15 (1.11–1.19)7.4×10−13296q25 MTHFD1Lrs6922269-A*0.261.09 (1.0–1014)2.3×10−528rs6922269-A†0.261.05 (1.0–1.1)0.022610q11 CXCL12rs1746048-C*0.841.17 (1.11–1.24)7.4×10−928rs501120-T†0.131.11 (1.05–1.18)4.3×10−42612q24 HNF1A/C12orf43rs2259816-A‡0.361.08 (1.05–1.11)4.8×10−72915q22 SMAD3rs17228212-T*0.731.05 (1.01–1.09)0.0228rs17228212-C†0.261.00 (0.95–1.04)0.926The risk-effect sizes found in these GWASs are of the same magnitude (1.2 to 1.6) as those seen in the meta-analyses of candidate SNPs.2 As for the candidate-gene SNPs, replication of effects and meta-analyses of published GWAS SNP data are vital to validate the findings, even though the original GWAS reports included replication studies (often several of them). To detect these modest effects with a reasonable degree of statistical certainty, very large replication studies are required; for example, even in a combined cohort of 33 382 subjects with 1436 CHD events, statistically significant risk effects were only confirmed for the GWAS SNPs on chromosomes 1p, 1q, 9p, 10q, and 19q, whereas those on chromosomes 2p, 6q, and 15q had more modest effects that were not statistically significant.33 Despite the large number of individuals genotyped, these differences may reflect a genuine degree of heterogeneity and context dependency, but it may also be that the original studies identified loci that they were only marginally powered to detect, and these findings would be unlikely to be replicated in other similarly powered studies. Finally, it is also possible that the original findings were spurious owing to a type I error. Achievement of the numbers required to confirm or refute these modest effects will be possible with the establishment of international consortia. Over the next year, there will be genotype data from >200 000 individuals genotyped with the HumanCVD BeadChip (Illumina, Inc, San Diego, Calif), which includes many of the CHD GWAS hits, and this will also provide a source of replication.34The GWAS-identified loci reported in Table 1 only represent the locus nearest to the risk-associated SNP(s) and are not necessarily the actual genes involved, and to date, the actual risk-causing variants for any of these loci have not been identified. However, if the linkage disequilibrium or level of correlation between the risk SNP and the disease-causing DNA change is high (eg, >90%), then that SNP will be a good surrogate marker for the variant and could be used, even at this early stage, in a genetic risk-score algorithm. Clearly, further work is needed to identify the functional variants, because their inclusion in the risk algorithm, instead of simply a marker in linkage disequilibrium, will improve risk prediction and reduce uncertainty. One of the current challenges in molecular genetics is to identify these functional variants.The chr9p21 risk SNPs lie in a gene-poor region, and the nearest genes (CDKN2A-ARF-CDKN2B) are >100-kilobases (kb) upstream from the risk SNPs (Figure 2C).13–15CDKN2A/CDKN2B are involved in cell cycle control, and thus, alterations in their expression could be postulated to lead to senescence and apoptosis, both of which are processes involved in plaque progression and rupture. However, more detailed analysis of the region between these genes and the risk SNPs has suggested an alternative candidate. In the PROCARDIS study (Precocious Coronary Artery Disease), susceptibility to coronary artery disease was encoded by 2 common haplotypes that span the 53-kb region that overlaps with ANRIL, a gene that encodes an antisense noncoding RNA.35 This is a member of a gene family involved in transcriptional control that overlaps and regulates CDKN2B36 and is expressed in atheromatous human vessels in vascular endothelial cells, monocyte-derived macrophages, and coronary smooth muscle cells, all of which are involved in atherosclerosis.35 A recent paper (using a mouse model with 58 kb of chr 9p21 deleted) has presented data which suggested that ANRIL expression is not the most likely mechanism, and identified a cis-acting element that influenced expression of CDKN2A/2B and thus cell apoptosis.35a This chr9p21 region is clearly a disease “hot spot,” being associated with risk of heart failure,37 type 2 diabetes mellitus,38 abdominal aortic aneurysms,39 stroke,37 and periodontal disease,40 whereas deletion of this whole region is implicated in certain cancers.41 The fact that common variation in this gene region is involved in such a wide range of diseases suggests that this locus encodes 1 or more key players in cell homeostatic processes that are involved in this set of complex multifactorial diseases and raises the possibility that influencing the expression at this locus may have important therapeutic consequences.In Silico Modeling of Effect of Combined Risk SNPsWe have modeled the use of 7 of these “novel” GWAS SNPs in the risk-score algorithm in combination with the 11 candidate-gene SNPs discussed above7 using the published risk-allele frequency in whites. The predicted distribution of individuals with different numbers of the combined 17 risk alleles is shown in Figure 3. The most common group will have 5 risk alleles, with 10% having only 3 and 5% having 2 or fewer, whereas 6.8% have 8 risk alleles, 2.7% have 9, and 1% have 10 or more. When the reported estimates of CHD risk per allele are used, compared with risk in the commonest group, those with 3 risk alleles have a roughly 33% lower and those with 2 or fewer risk alleles have 50% lower CHD risk. By contrast, those with 8 risk alleles have a risk of >1.9, those with 9 have a risk approaching 2.5, and those with 10 have a risk of >3.1; by simulation in 10 000 subjects, these effects of 2 or fewer and 10 or more risk alleles are statistically significant (online-only Data Supplement). It is of relevance that the risk associated with being a current cigarette smoker is roughly 2-fold that of a nonsmoker, and so we can predict from this that 2% to 3% of the population will be at a clinically important higher genetic risk and roughly 4% will be at lower risk on the basis of the combined information from these SNPs. Although it is clear that the risk of smoking can be reduced by quitting, there are appropriate interventions for those who have a high genetic risk profile, such as more aggressive treatment of modifiable risk factors such as plasma cholesterol levels and blood pressure by pharmacological intervention, lifestyle interventions to reduce obesity and stress, and modification of diet and alcohol intake. Although not yet proven, it is reasonable to assume that these risk factor modifications will be equally effective in subjects at genetically high risk as in subjects in the general population, as has been found, for example, with the reduction in CHD mortality seen with lipid-lowering therapies in patients with the monogenic disorder of familial hypercholesterolemia. However, although it appears very likely that these SNPs will add to risk prediction over and above the CRF-based Framingham risk score, further data are required to confirm this. Download figureDownload PowerPointFigure 3. Estimated CHD risk found in individuals carrying different numbers of risk SNPs by use of published meta-analysis risk estimates in Casas et al2; see Tables 1 and 2.Clinical Utility of Genotype Risk Stratification: Chromosome 9p21 SNP as an ExampleProspective studies of incident as opposed to prevalent CHD are required to assess the utility of both biomarkers and genotypes with regard to CHD prediction. Case-control studies are efficient for gene discovery, but because they usually recruit more severely diseased or younger patients, they are unrepresentative of the disease cases in the general population, they may provide biased information on population allele frequency, and they cannot give unbiased estimates of attributable risk or the effect of genes on other important risk factors for cardiovascular disease (CVD). Most importantly, only prospective studies allow an accurate estimation of the absolute risk associated with a genotype or permit evaluation of the predictive utility of genetic information over and above the impact of classic risk factors. We therefore investigated whether the addition of a chr9p21 SNP genotype improved the prediction of CHD events by CRFs in the Framingham risk-score algorithm in 2742 healthy middle-aged men from the prospective Northwick Park Heart Study II (NPHSII) who were followed up for an average of 14 years with 270 CHD events.17 The rs10757274 G allele was associated with incident CHD with similar effect size to that observed in case-control studies, (hazard ratio of 1.60 in GG compared with AA men, 95% CI 1.12 to 2.28). The population-attributable fraction for CHD explained by the SNP was 26.2% (95% CI 7.1 to 41.1), which was independent of CRFs and reported family history of early CHD. Although this high population-attributable fraction estimate would suggest that this variant is likely to be of clinical utility, this is not an appropriate statistic measure to demonstrate this. To examine this, the ability to risk stratify, or “discrimination,” was evaluated by the AROC, and when genotype was added to the model, perhaps surprisingly given its large effect in univariate analysis, the AROC for CRFs alone of 0.62 (95% CI 0.58 to 0.66) was nonsignificantly (P=0.14) increased by 3% to 0.64 (95% CI 0.60 to 0.68). Similarly, in the Women’s Genome Health Study of 22 129 white women followed up for ≈10 years with 715 CVD events,18 the same chr9p21 SNP was significantly associated with CHD risk in univariate analysis (hazard ratio in carriers 1.25, 95% CI 1.04 to 1.51), but the addition of this genotype to a CRF-based algorithm did not improve the AROC. However, it is now increasingly recognized that prediction can only be improved significantly by the inclusion of factors that are both common and have very large effects,42 and that a single genotype (or biomarker) associated with ORs in the region of 1.2 to 1.6 will not on its own significantly improve risk prediction for polygenic, multifactorial CHD. To improve an already high AROC, biomarkers (or single SNPs or SNPs in combination) with large effects that are also common are needed, and there is diminishing added value in the higher AROC ranges for variants with similar effects, as we demonstrated in modeling the added effects of SNPs with similar frequency and risk size as the chr9p21 variant.17 We modeled the effect on CHD risk of up to 10 hypothetical, randomly assigned gene variants with allele frequencies and risk similar to those of rs10757274. The addition of 1 further SNP with similar characteristics increased the AROC significantly (P<0.03), whereas the inclusion of 2 or more SNPs had a greater effect (P<0.001), with the addition of further SNPs having smaller incremental effect. The AROC for 10 SNPs was 0.76.17However, the more relevant clinical use for such genetic information is in risk calibration, which is the ability of the genotype to improve the stratification of individuals into risk categories (ie, whether genotype information increased the number of men correctly reclassified into those developing CHD or remaining event free during follow-up). The potential clinical utility of this genotype for risk stratification in the NPHSII men was also examined. Although treating subjects with a low Framingham risk score will not be cost-effective because of the low event rate, the majority of CHD events occur in subjects with intermediate risk scores, because this is the most common group. The National Institute for Clinical Excellence, which sets standards of clinical treatment in the United Kingdom, recommends that subjects with a 10-year risk of CVD of >20% should be treated with stat