Multilocus Genetic Risk Scores for Coronary Heart Disease PredictionSignificance
Objective—Current guidelines do not support the use of genetic profiles in risk assessment of coronary heart disease (CHD). However, new single nucleotide polymorphisms associated with CHD and intermediate cardiovascular traits have recently been discovered. We aimed to compare several multilocus genetic risk score (MGRS) in terms of association with CHD and to evaluate clinical use.
Approach and Results—We investigated 6 Swedish prospective cohort studies with 10 612 participants free of CHD at baseline. We developed 1 overall MGRS based on 395 single nucleotide polymorphisms reported as being associated with cardiovascular traits, 1 CHD-specific MGRS, including 46 single nucleotide polymorphisms, and 6 trait-specific MGRS for each established CHD risk factors. Both the overall and the CHD-specific MGRS were significantly associated with CHD risk (781 incident events; hazard ratios for fourth versus first quartile, 1.54 and 1.52; P<0.001) and improved risk classification beyond established risk factors (net reclassification improvement, 4.2% and 4.9%; P=0.006 and 0.017). Discrimination improvement was modest (C-index improvement, 0.004). A polygene MGRS performed worse than the CHD-specific MGRS. We estimate that 1 additional CHD event for every 318 people screened at intermediate risk could be saved by measuring the CHD-specific genetic score in addition to the established risk factors.
Conclusions—Our results indicate that genetic information could be of some clinical value for prediction of CHD, although further studies are needed to address aspects, such as feasibility, ethics, and cost efficiency of genetic profiling in the primary prevention setting.
Since the dissemination of the HapMap project in 2005,1 genetic researchers have performed a large number of genome-wide association studies to identify genetic determinants of complex diseases. In the cardiovascular field, recent examples include the publication from the CARDIoGRAMplusC4D consortium, which increased the number of loci robustly associated with coronary heart disease (CHD) in individuals of Northern European descent to 46,2 and studies of related traits, such as lipid fractions3 or fasting glucose and insulin.4 The improved biological understanding of these traits has yet to be followed by clinical applications of the discoveries. Although direct-to-consumer tests of recently discovered genetic markers are already available on the market, especially in the United States, both US5 and European6 guidelines for prevention of cardiovascular diseases in clinical practice advise against prognostic use of DNA-based tests in the primary prevention setting.
See accompanying article on page 2049
Several studies have investigated whether the introduction of a multilocus genetic risk score (MGRS) in addition to an established CHD risk algorithm, such as the Framingham Heart Study (FHS) risk score,7 would improve disease prediction. Using 13 CHD-associated single nucleotide polymorphisms (SNPs), Ripatti et al8 showed that an MGRS was highly associated with CHD, but was unable to improve the risk classification beyond what was achieved with traditional risk factors. Paynter et al9 considered a larger genetic score (101 SNPs), including SNPs associated with cardiovascular disease–related phenotypes, and failed to demonstrate a significant association with CHD. However, to the best of our knowledge, no previous studies have examined the use of an MGRS, including the most recently reported CHD-related SNPs. Furthermore, an alternative scoring approach has been proposed, which includes all SNPs associated with the outcome in an external population up to a certain probability value, instead of using only published genome-wide significant SNPs. This so-called polygene approach, which has shown promising results for other traits in cross-sectional studies,10 has so far not been tested for CHD prediction.
In the present study, including 6 longitudinal, population-based Swedish cohorts of >10 000 participants, our primary aim was to compare an overall MGRS and a CHD-specific MGRS in terms of disease association and clinical use. Our secondary aims were to investigate the clinical use of a polygene MGRS and the associations of trait-specific MGRS with CHD and established cardiovascular risk factors.
Materials and Methods
Materials and Methods are available in the online-only Supplement.
Genetic Score Generation Procedures
In Figure 1A, we have schematically outlined the process for selection of SNPs to be included in the MGRS from the National Human Genome Research Institute GWA study catalog.11 Briefly, 119 traits that are related to CHD in a direct or more broad sense (Table I in the online-only Data Supplement) were selected from the National Human Genome Research Institute catalog by a medical doctor (E.I.) with expertise in cardiovascular epidemiology that was blinded to the results of the association analyses in our data. All genome-wide association studies investigating these traits underwent an accurate quality control to assess the validity of the SNPs reported in the catalog. After quality control, 615 single SNPs from 92 studies investigating 49 traits were used to calculate the scores. We eliminate correlated SNPs in high linkage disequilibrium using the pruning technique as implemented in PLINK,12 and we created 3 weighted MGRS using the log(odds ratio) for CHD association in Wellcome Trust Case Control Consortium13 (score 1 and 3) or in a published article2 (score 2) to calculate weights and 7 unweighted trait-specific MGRS.
An overall MGRS obtained from 395 SNPs associated with CHD or CHD-related traits (Tables I and II in the online-only Data Supplement);
A CHD-specific MGRS obtained from 46 SNPs that have been found associated with CHD by the CARDIoGRAMplusC4D consortium, which is the largest genome-wide association studies meta-analysis on CHD to date2;
A polygene MGRS obtained as the weighted sum of risk alleles for SNPs with P<0.2 in the Wellcome Trust Case Control Consortium study. This threshold corresponds to the highest increment in the C-statistic over a basic model, as shown graphically in Figure I in the online-only Data Supplement;
An unweighted trait-specific MGRS for each of 6 established cardiovascular risk factors: body mass index (BMI; 37 SNPs), high-density lipoprotein (HDL)-cholesterol (47 SNPs), systolic blood pressure (25 SNPs), total cholesterol (TC; 34 SNPs), smoking (7 SNPs), type 2 diabetes mellitus (40 SNPs), and 1 FHS MGRS, including 180 SNPs associated with any of the aforementioned traits.
A total of 10 612 participants, free of CHD at baseline and with complete information on FHS risk score components and BMI, were included in our study. A total of 781 CHD events (539 in men and 242 in women) were observed during a median follow-up of 4.3 years (interquartile range, 3.6–5.8). In Table 1, we report the descriptive statistics of our sample, together with the associations of the FHS risk score components and BMI with incident CHD.
Association With CHD and Risk Factors
In a single SNP analysis (Table I in the online-only Data Supplement), MECOM rs419076 and IGFBP3 rs7784776 showed the strongest associations with CHD after adjustment for FHS risk factors. However, the associations were not significant after multiple-testing correction in the univariate analysis. Both the overall MGRS and the CHD-specific MGRS were highly significantly associated with CHD. Among the risk factors, the CHD-specific MGRS was associated only with TC; interestingly, no association was detected with the FHS. Each trait-specific MGRS was significantly associated with the corresponding trait, except for the smoking MGRS (Table 2). The FHS MGRS was associated with all the risk factors for CHD, except for smoking, but not directly with CHD. The HDL-cholesterol MGRS was strongly associated with BMI and TC. No association with CHD was found for the polygene MGRS. Once the SNPs included in the CHD-specific score were removed from the overall MGRS, this score was still significantly associated with CHD (hazard ratio, 1.091; P=0.008).
Reclassification, Discrimination, and Calibration
Participants who were in the upper quartile of the distribution of the overall MGRS had 1.54× increased risk for CHD compared with individuals in the lowest quartile (P<0.001; Table 3). The overall MGRS significantly improved risk classification beyond established FHS risk factors (net reclassification improvement [NRI], 4.2%; P=0.006), but the discrimination improvement was modest (C-index improvement, 0.002). Higher reclassification and discrimination were observed for the CHD-specific MGRS (NRI, 4.9%; C-index improvement, 0.004). Participants’ distribution within risk categories and reclassification after addition of the CHD-specific MGRS are reported in Figure II in the online-only Data Supplement. Intermediate- and high-risk subjects evidenced the largest changes in individual risk, when the CHD-specific MGRS was added to the FHS risk factors (Figure III in the online-only Data Supplement). Good calibration was observed for all scores (Table 3). The TC MGRS had the highest R2, explaining 6.2% of the total variance for this phenotype in our sample (Figure IV in the online-only Data Supplement). When calculating the discriminatory abilities of the CHD-specific MGRS without adjusting for established risk factors, we obtained a C-index of 0.54.
Number of Events and Event-Free Life Years Prevented
Among the 1272 (Figure II in the online-only Data Supplement) participants with a 10-year risk between 10% and 20%, 83 (6.5%) participants would have been reclassified as high risk (>20%), when adding the CHD-specific score and, therefore, would have been eligible for statins treatment according to the adult treatment panel-III guidelines.14 Twenty of these participants experienced an event during the 10 years follow-up. That is, assuming a risk reduction of 20% for individuals treated with statins, the targeted assessment of the genetic risk score among intermediate risk subjects could help to prevent ≈4 (ie, 0.20×20) additional CHD events during a 10-year period, which corresponds to 1 avoided event for every 318 people screened (ie, 1272/4; see online-only Data Supplement for details).
Furthermore, if the CHD-specific genetic score was measured on the entire study population in addition to the established risk factors, 3.15 [confidence interval, −1.30; 7.60] event-free life years per 1000 people screened would have been saved (Table III in the online-only Supplement).
In the present longitudinal study of 10 612 participants, we investigated 10 MGRS for association with CHD and established cardiovascular risk factors. Three scores were further investigated in terms of clinical use (ie, the ability to discriminate between individuals with and without the disease and to correctly classify them in clinical categories of risk). We found that both a large comprehensive genetic score and a CHD-specific score were able to significantly improve the correct classification of study participants. In addition, we showed that a polygene MGRS, developed in an external data set, including all SNPs up to a nonsignificant P value threshold, did not outperform a literature-based score. Finally, we found that the genetic risk score for HDL-cholesterol was associated both with BMI and TC, potentially indicating pleiotropic effects of HDL-associated loci. The association between the CHD-specific MGRS and the 10-year predicted risk calculated from established risk factors (FHS) was not significant. Indeed, only 14 of 44 CHD-specific loci have been observed in a previous large study2 to be associated with lipid-related traits, type 2 diabetes mellitus, or blood pressure. Similarly, in an older study,15 only 1 of 14 CHD-associated SNPs was associated with a wide range of established risk factors and novel biomarkers. These previous results, together with our findings, support the hypothesis that most of the CHD loci are not involved in pathways perturbing currently known risk factors, making a genetic score of CHD-specific loci a good candidate for future improvements of current prediction algorithms, and CHD loci an interesting starting point for further studies into the pathophysiology of atherosclerosis.
Our results indicate that use of genetic profiling, including SNPs associated with intermediate traits, might improve allocation of patients to correct risk strata. Although we calculated prediction measures in a relatively small- and high-risk subsample of the study, in which we had long enough follow-up time (n=3014), the NRI (4.9%; P=0.01) was significant. This proportion of correct reclassification is clinically relevant, and is larger than that observed for HDL-cholesterol in our study (NRI, 2.7%) and for C-reactive protein and fibrinogen in a recent meta-analysis (NRI, 1.52% and 0.83%, respectively).16 However, caution should be exercised when comparing estimates from a large meta-analysis with that observed in a single study. In a hypothetical population similar to that investigated in this study, the routine use of genetic profiling in primary prevention of CHD among individuals at intermediate risk could prevent 1 event for every 300 to 400 people screened. If the risk assessment with genetic profiling is performed on the entire population, and not only on individual at intermediate risk, 3.15 [−1.30; 7.60] years free of CHD events could be saved per 1000 people undergoing the risk assessment. Clinicians and health policy makers should evaluate whether these benefits outbalance the costs of implementing such testing in clinical practice.
Genetic profiling has some advantages compared with other biomarkers that are routinely used in clinical practice. For example, genetic information remains stable throughout life and is, therefore, not sensitive to regression dilution bias. Hence, although the strength of the genotype–phenotype associations has been suggested to be modified by age,17 genetic markers are likely to be predictive throughout life. This might allow risk prediction to be performed much earlier in life, to allow for earlier primary prevention in high-risk individuals. However, to investigate this aspect, study populations with younger participants and longer follow-up are needed than what was available for the present study. Another potential advantage is that no invasive blood sampling is needed, because DNA is routinely extracted from saliva. However, several important obstacles remain to be considered before genetic profiling can be introduced in routine healthcare. These include demonstration of usefulness and cost-effectiveness for risk prediction in several independent samples and the fact that genetic markers, in contrast to modifiable risk markers, are not useful for assessing efficacy of treatment or other risk-reduction strategies.
The alternative scoring approach (polygene approach), which includes all SNPs associated with the outcome in an external population up to a certain probability value, has shown promising results for other traits and has been proposed to increase the variance explained for common complex diseases.18 For the first time, we applied this approach to CHD. In our study, this score was not associated with CHD, probably because of the relatively small data set used for generating the score, resulting in lower accuracy of the weights. Theoretical frameworks have predicted polygenic scores to substantially increase the discriminatory ability of the model, with sample size larger than those currently available from meta-analysis consortia.19,20 Therefore, future studies using larger development data sets are needed to understand whether a large number of SNPs with small effect size not exceeding the genome-wide threshold could be able to improve the clinical use beyond established genetic variants. The polygene score and the overall MGRS embrace a similar approach; they include a large number of SNPs, both true signals and noise, trying to capture the polygene architecture of CHD. The SNP selection strategy for these 2 scores was, however, different. The polygene score uses a probability value threshold for SNP selection and, therefore, the predictive power is much dependent on the sample size of the study where the selection was made. In contrast, the overall MGRS uses a priori information from the literature and, therefore, the prediction ability is determined by the quality of the literature selection. The overall MGRS was significantly associated with CHD, even after the exclusion of CHD-specific SNPs, indicating that a selection strategy based on literature-based information can be useful to improve the prediction ability of genetic risk scores. Moreover, we performed 2 sensitivity analyses to investigate whether the inclusion of a large number of SNPs in the overall MGRS might have caused a spurious association with CHD. First, we plotted the association results across the 6 investigated studies to address consistency (Figure V in the online-only Supplement). Second, we created 100 random overall MGRS and studied their association with CHD in TwinGene (Figure VI in the online-only Supplement). The association was consistent across studies and almost identical in the 2 larger studies. The average association of the 100 random scores was significantly lower (P=0.01) than the associations observed in the original score. These analyses suggest that the association between the overall MGRS and CHD is unlikely to be explained by chance.
Previous studies did not find a significant association between an overall MGRS and a CHD.21,22 A smaller number of traits included in the score, different outcomes definitions, and heterogeneity between study populations might explain such differences. In addition, strengths of our study include the large number of incident CHD events investigated, as well as the up-to-date literature-based MGRS, including genetic variants from the most recent studies. Moreover, we used information from an external study population to assign weights in our scoring. This is in contrast to previous studies that have used association coefficients reported in the literature8; an approach that biases the score attributable to the winner’s curse. We also acknowledge several limitations to our study. We do not have any information on family history. However, this information is not included in the FHS risk score, and a previous report indicates that inclusion of family history does not influence the magnitude of the association between a CHD-specific MGRS and CHD.8,21 Moreover, although family history remains an important contributing factor to risk prediction, it has recently been shown that genetic scores can substantially increase the discriminatory abilities of the model above family history alone.20 However, larger sample sizes for deriving these scores are needed.
Furthermore, our study was undertaken in Swedish middle-aged to elderly individuals; hence, the generalizability to other ethnicities or age groups is unknown. We expect early onset CHD events to be more influenced by genetic factors compared with late-onset CHD. Therefore, the predictive ability of a genetic score in younger individuals is likely to be higher.23 Finally, we only included myocardial infarction and unstable angina in our outcome, but not stable angina, because the validity of this diagnosis in Swedish National Patient Register is unknown. Differences in the definition of the cardiovascular outcomes across studies might limit the comparability and generalizability of the results.
In conclusion, using data from 6 Swedish prospective cohort studies with 10 612 healthy participants from the community, we have investigated the clinical use of genetic scores in primary prevention of cardiovascular diseases. Current efforts to discover additional genetic loci associated with CHD and related traits, as well as the use of sequencing approaches along with integration of other omics technologies, are likely to further improve the performance of existing predictive equations. Our results indicate that genetic information could be of some clinical value for prediction of CHD, although further studies are needed to address aspects such as feasibility, ethics, and cost efficiency of genetic profiling in the primary prevention setting.
We thank Dr Rapsomaniki for the methodological contribution.
Sources of Funding
This study was supported by grants from the Swedish Foundation for Strategic Research (ICA08-0047), the Swedish Research Council (project grants no. 2012-1397 and 2009–2298), the Swedish Heart-Lung Foundation (2012-0197, 2010-0401 and 2008-0326), the National Institutes of Health (AG04563, AG10175, AG08861, AG08724, AG028555, and DK 066134), the Swedish Society of Medicine, the Knut and Alice Wallenberg Foundation, the Torsten and Ragnar Söderberg Foundation, and the Strategic Cardiovascular Program of Karolinska Institutet and the Stockholm County Council.
The online-only Data Supplement is available with this article at http://atvb.ahajournals.org/lookup/suppl/doi:10.1161/ATVBAHA.113.301218/-/DC1.
- Received January 21, 2013.
- Accepted May 5, 2013.
- © 2013 American Heart Association, Inc.
- Greenland P,
- Alpert JS,
- Beller GA,
- et al
- Graham I,
- Atar D,
- Borch-Johnsen K,
- et al
- Wilson PW,
- D’Agostino RB,
- Levy D,
- Belanger AM,
- Silbershatz H,
- Kannel WB
- Hindorff LA,
- Sethupathy P,
- Junkins HA,
- Ramos EM,
- Mehta JP,
- Collins FS,
- Manolio TA
- 14.↵Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults. Executive summary of the third report of the national cholesterol education program (NCEP) expert panel on detection, evaluation, and treatment of high blood cholesterol in adults (adult treatment panel iii). JAMA. 2001;285:2486–2497.
- Angelakopoulou A,
- Shah T,
- Sofat R,
- et al
- Evans DM,
- Visscher PM,
- Wray NR
- Thanassoulis G,
- Peloso GM,
- Pencina MJ,
- Hoffmann U,
- Fox CS,
- Cupples LA,
- Levy D,
- D’Agostino RB,
- Hwang SJ,
- O’Donnell CJ
- Vaarhorst AA,
- Lu Y,
- Heijmans BT,
- et al
- Nora JJ,
- Lortscher RH,
- Spangler RD,
- Nora AH,
- Kimberling WJ
Both US and European guidelines for prevention of cardiovascular diseases in clinical practice advise against prognostic use of DNA-based tests in the primary prevention setting. In this study, we show that genetic information could be of some clinical value for prediction of coronary artery disease. By measuring the coronary artery disease–specific genetic score in addition to the established risk factors, 1 additional coronary artery disease event for every 318 people screened at intermediate risk can be saved. Both a score including 46 well-established coronary artery disease–associated SNPs and a score including 395 SNPs associated with cardiovascular risk factors have similar prediction performances.