Dr J. Gustav Smith, Department of Cardiology, Faculty of Medicine, Lund University, Skåne University Hospital, SE-221 85, Lund, Sweden. (fax: +46 46 15 78 57; e-mail: email@example.com).
Smith JG, Almgren P, Engström G, Hedblad B, Platonov PG, Newton-Cheh C, Melander O (Department of Cardiology, Lund University, Lund, Sweden; Department of Clinical Sciences, Lund University, Malmö, Sweden; Broad Institute of Harvard and Massachusetts Institute of Technology, Cambridge, MA, USA; and Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA). Genetic polymorphisms for estimating risk of atrial fibrillation: a literature-based meta-analysis. J Intern Med 2012; doi: 10.1111/j.1365-2796.2012.02563.x
Background. Genetic polymorphisms associated with common aetiologically complex diseases have recently been identified through genome-wide association studies. Direct-to-consumer genetic testing for such polymorphisms, with provision of absolute genetic risk estimates, is marketed by several commercial companies. Polymorphisms associated with atrial fibrillation (AF) have shown relatively large risk estimates, but the robustness of such estimates across populations and study designs has not been investigated.
Design. A systematic literature review with meta-analysis and assessment of between-study heterogeneity was carried out for single-nucleotide polymorphisms (SNPs) in the six genetic regions associated with AF in genome-wide or candidate gene studies.
Results. Data were identified from 18 samples of European ancestry (n =12 100 cases, 115,702 controls) for the single-nucleotide polymorphisms (SNP) on chromosome 4q25 (rs220733), from 16 samples (n =12 694 cases, 132 602 controls) for the SNP on 16q22 (rs2106261) and from four samples (n =5272 cases, 59 725 controls) for the SNP in KCNH2 (rs1805123). Only the publications in which the associations were initially reported were identified for SNPs on 1q21 and in GJA5 and IL6R, why meta-analyses were not performed for those SNPs. In overall random-effects meta-analyses, association with AF was observed for both SNPs on chromosomes 4q25 [odds ratio (OR), 1.67; 95% CI, 1.50–1.86, P =2 × 10−21] and 16q22 (OR, 1.21; 95% CI, 1.13–1.29, P =1 × 10−8) from genome-wide studies, but not the SNP in KCNH2 from candidate gene studies (P =0.15). There was substantial effect heterogeneity across case–control and cross-sectional studies for both polymorphisms (I2 =0.50–0.78, P <0.05), but not across prospective cohort studies (I2 =0.39, P =0.15). Both polymorphisms were robustly associated with AF for each study design individually (P <0.05).
Conclusions. In meta-analyses including up to 150 000 individuals, polymorphisms in two genetic regions were robustly associated with AF across all study designs but with substantial context-dependency of risk estimates.
Since genetic association studies became feasible on a genome-wide scale in 2006 reproducible associations between common variants (polymorphisms) and many complex diseases have been established. Whereas rare genetic variants (mutations), exclusive to individual families, can have an almost deterministic impact on disease, polymorphisms confer a probabilistic increment in risk, set against other clinical and environmental risk factors in causing disease. Direct-to-consumer genetic testing for such disease-associated polymorphisms is currently marketed by commercial companies, providing the customer with an estimate of absolute genetic risk, derived from reported average population risk and genotypic risk estimates. However, absolute risk estimates vary widely between companies, and research on the clinical utility of genetic testing is only in its infancy with a number of issues still to be addressed .
First, limited data have been published on the consistency of genotypic risk estimates across studies and populations. As genome-wide association (GWA) studies are tools for the discovery of novel genetic susceptibility regions, narrow sample selection criteria are often utilized to improve statistical power, including the use of cases with early disease onset or with a family history of disease or the use of ‘super-controls’, who have not developed disease despite reaching a relatively old age. Such strategies can potentially inflate risk estimates, and the results of GWA studies may therefore not be directly generalizable to genetic prediction on the population level. Second, the clinical validity of polymorphisms for prediction has for most diseases been shown to be limited owing to relatively small conferred risks. For example, for coronary artery disease, one of the most well-studied conditions, risk estimates for associated polymorphisms have been shown to be similar across studies [2, 3]. However, these risk estimates are modest (odds ratio ∼1.2 per risk allele, corresponding to an approximate risk increase of 20% compared to individuals without the risk allele) and contribute minimally to conventional risk factors in prediction, as quantified by the area under the receiver operating characteristic curve [3, 4] which scales with the magnitude of risk estimates .
Relatively higher risk estimates with odds ratios up to 2.46, corresponding to a risk increase of 146% per risk allele, have been reported for polymorphisms associated with atrial fibrillation (AF) [6–10]. AF is a common cardiac condition and major risk factor for stroke, heart failure and death, which is known to have a heritable component [11–13]. To date, polymorphisms at six genomic regions have been reproducibly associated with AF: (i) chromosome 4q25, located 150 kb from the closest gene, that is, a transcription factor (PITX2) involved in cardiac development ; (ii) chromosome 16q22, located in an intron of another transcription factor of unknown function, expressed in cardiac tissue (ZFHX3) [7, 8]; (iii) an amino acid-altering variant in KCNH2, the gene encoding one of the major cardiac voltage-gated potassium channels ; (iv) a variant located in an intron of the IL6R gene, encoding the receptor for the cytokine interleukin-6, ; (v) a polymorphism on chromosome 1q21, located in an intron of a cardiac potassium channel (KCNN3) ; and (vi) another polymorphism in the promoter region upstream of the GJA5 gene, encoding a component of cardiac gap junctions . The latter two polymorphisms have been specifically associated with lone AF, that is, AF presenting at a young age in the absence of precipitating factors such as heart failure, valvular disease and hyperthyroidism. However, the consistency of risk estimates has not been evaluated for these six polymorphisms.
Thus, we performed a systematic literature review of studies to investigate the collective evidence of association between these polymorphisms and AF and evaluate the robustness of risk estimates across studies of various design.
Materials and methods
Systematic literature review
A systematic literature search was performed in PubMed for studies testing the association between AF and single-nucleotide polymorphisms (SNPs) in the six genetic regions for which reproducible association has been reported: chromosomes 4q25 (rs2200733), 16q22 (rs2106261) and 1q21 (rs13376333) identified in GWA studies [6–9], and a missense single-nucleotide polymorphisms (SNP) (rs1805123) in the KCNH2 gene and SNPs in GJA5 (rs10465885) and IL6R (rs4845625) identified in candidate gene studies [10, 14, 15]. The following search criterion was used: ‘atrial fibrillation’ AND (4q25 OR 16q22 OR 1q21 OR PITX2 OR ZFHX3 OR KCNN3 OR KCNH2 OR GJA5 OR IL6R OR rs2200733 OR rs2106261 OR rs13376333 OR rs1805123 OR rs10465885 OR rs4845625). Reference lists of included articles were hand-searched. We excluded studies reported in languages other than English or those reporting only on interventional subgroups, such as patients undergoing coronary artery bypass graft surgery or cardiac catheter ablation procedures. In several studies, consortium results with multiple independent populations were reported. These samples were listed and analysed as separate datasets. Samples included in the results of more than one study were only included once, but population-based cohort samples could contribute to both cross-sectional and prospective analyses.
Meta-analyses were performed separately for samples of different primary ancestries (European, Asian and African). Heterogeneity was assessed using Cochran’s Q test for heterogeneity, computed as the sum of the squared deviations of each study’s effect from the weighted mean over the study variance, and the I2 test, the percentage of total variation across studies that is because of heterogeneity rather than chance (I2 =[Q − df]/Q, where ‘Q’ is Cochran’s Q statistic and ‘df’ refers to number of degrees of freedom) [16, 17]. For each genetic polymorphism, effect estimates were combined across all studies and per study design (case–control, cross-sectional and prospective cohort) using random-effects meta-analysis as described by DerSimonian and Laird  or fixed-effects meta-analysis with inverse variance weights. To assess the potential role of study design or sample age as the source of any observed heterogeneity, additional heterogeneity analyses were performed by study design and sample mean age was meta-regressed on point estimates for cross-sectional and prospective cohort studies. Meta-regression was performed separately for cross-sectional and prospective cohort studies and was not performed for case–control studies as there were substantial differences in age distribution between cases and controls in most of the latter studies. Sample inclusion bias was evaluated for each SNP both overall and for each study design using funnel plots of individual study samples and Egger’s linear regression asymmetry test, where the normalized effect estimate is regressed against the inverse of its standard error and the hypothesis that the intercept of the regression line differs from the origin is tested . Several genetic inheritance models were explored (additive, recessive and dominant models) in the largest published sample to study the relation between these genetic polymorphisms and AF (the prospective Malmö Diet and Cancer study; MDCS) . Genetic models were compared using two measures of global fit: the Akaike information criterion (AIC)  and the Schwarz Bayesian information criterion (BIC) . Statistical analyses were performed in SAS version 9.2 (SAS Institute, Cary, NC, USA) or Stata version 11.1 (StataCorp, College Station, TX, USA).
Systematic literature review
From the literature review, we identified five publications, including five cross-sectional, seven case–control and six prospective European samples with a total of 12 100 cases and 115 702 controls, that tested the association between the SNP on chromosome 4q25 and AF (Fig. 1) [6, 7, 20, 23, 24]. For the SNP on chromosome 16q22, three publications were identified, including five cross-sectional, five case–control and six prospective samples with a total of 12 694 cases and 132 602 controls (Fig. 2) [7, 8, 20]. For the SNP in KCNH2, three publications including one cross-sectional, two case–control and one prospective European samples with 5272 cases and 59 725 controls were identified (Fig. 3) [10, 20, 25]. For the SNPs on chromosome 1q21 and in IL6R and GJA5, we only identified the publications in which the associations were initially described [9, 14, 15], why meta-analyses were not performed for these SNPs.
Few studies using samples of primarily non-European ancestry were identified: three Asian samples from three publications [6, 26, 27] evaluated the association between the SNP on chromosome 4q25 and AF, with a total of 916 cases and 3845 controls. One study reporting association between the SNP on 16q22 and AF in a Chinese sample of 650 cases and 1447 controls (P =0.001) was also identified . Amongst African Americans, association of 4q25 with AF was reported in two studies [14, 29] and of 16q22 with AF in one .
Meta-analysis in samples of European ancestry
Characteristics of European samples in previous studies included in the meta-analysis are shown in Table 1. Distribution of sex and age at baseline was reported for all six population-based samples. Population-based samples had a similar proportion of male participants but varied widely in distribution of baseline age, ranging from a sample mean of 58.1 (SD, 7.6) years in the MDCS to 76.5 (SD, 5.5) years in the Age, Gene/Environment Susceptibility study. Of the 12 case–control samples, six did not report distribution of sex and age at diagnosis for all participants. The distribution of age and sex varied widely between cases and controls in most case–control studies. The age at diagnosis in cases ranged from 54.2 (SD, 13.7) years in a case–control study from Nashville, USA, to 74.4 (SD, 8.7) years in a study from Sweden. In individuals of European ancestry, the minor allele was the same for all cohorts (4q25, T; 16q22, A; KCNH2, C) and allele frequencies were similar across cohorts as shown in Table S1, with weighted average minor allele frequencies of 0.11 for 4q25, 0.18 for 16q22 and 0.22 for KCNH2.
Table 1. Baseline characteristics of samples of European ancestry included in meta-analyses
Sample size, n (cases/controls)
Case age, years (mean, SD)
Control age, years (mean, SD)
Case sex (% male)
Control sex (% male)
For case–control studies, characteristics at diagnosis for cases are presented separately from characteristics at recruitment for controls. Case–control samples are referred to based on the geographic region of origin. Baseline characteristics are presented for the entire cohort for cross-sectional and prospective studies. AGES, Age, Gene/Environment Susceptibility study; ARIC, Atherosclerosis Risk in Communities study; CHS, Cardiovascular Health Study; FHS, Framingham Heart Study; MDCS, Malmö Diet and Cancer Study; NA, not available; RS, Rotterdam Study. *Sample characteristics from Sinner et al. 2011 .
The association of 4q25 and 16q22 with AF was robust when analysed in random-effects meta-analyses both overall (4q25, P =2 × 10−21; 16q22, P =1 × 10−8) and by study design as shown in Figs 1 and 2. No association was observed in a random-effects meta-analysis of KCNH2 with AF (P =0.15) as shown in Fig. 3.
Significant heterogeneity of relative risk estimates across samples was observed for all three SNPs (I2 =57–82%). When assessed by study design, heterogeneity for 4q25 was observed in case–control (I2 =71%) and cross-sectional (I2 =78%) but not prospective samples (I2 =39%, P =0.15). Heterogeneity for 16q22 was observed for case–control samples (I2 =65%). When meta-regressed on risk estimates for cross-sectional samples, no association was observed with mean sample age for the SNP on chromosome 4q25 (β [−0.01] per year, 95% CI [−0.135] − 0.115, P =0.81) or 16q22 (β 0.02 per year, 95% CI [−0.04] − 0.07, P =0.43).
No sample inclusion bias was detected for 4q25 or 16q22 (all P >0.05) overall or for any study design. Funnel plots including all samples are shown in Figs S1 and S2.
Meta-analyses in samples of Asian and African ancestry
Minor allele frequencies were different in Asian or African samples compared to European samples, as shown in Table S2. Of importance, the C-allele for the SNP on chromosome 4q25 was more highly prevalent in Asian or African cohorts and was consistently the major allele in populations of Asian ancestry whilst it was the minor allele in those of European or African descent. However, the C-allele was also significantly associated with increased risk of AF in both individuals of Asian (P =2 × 10−4) and African (P =0.002) ancestry, as shown in Table 2. No significant heterogeneity was detected, but the number of studies in these populations was small. Only one study reported association between the SNP on chromosome 16q22 and AF in Asian individuals, and only one small study reported the lack of association in an African cohort.
Table 2. Results in samples of Asian and African ancestry
Sample size, n (cases/controls)
All studies were case–control studies, with the exception of the two population-based, African American cohorts that were analysed jointly in . Results for the single-nucleotide polymorphisms (SNPs) rs2200733 on chromosome 4q25 and rs2106261 on 16q22 are presented as odds ratio per risk allele with corresponding 95% confidence interval and P-value. Combined results from fixed-effects meta-analyses are presented where multiple samples were available, including odds ratio with corresponding 95% confidence interval, P-value and heterogeneity statistics including the I2 statistic and P-value for Cochran’s Q test.
In the prospective MDCS, the lowest AIC and BIC statistics were consistently observed for additive genetic models for all three polymorphisms indicating best model fit, with the exception that the dominant model showed similar statistics to the additive model for 4q25 (Table 3).
Table 3. Genetic inheritance models
Global model fit for each genetic model in the prospective Malmö Diet and Cancer study is presented as the Akaike information criterion (AIC) and the Schwarz Bayesian information criterion (BIC).
In these meta-analyses including up to 150 000 individuals, robust association with AF was observed for two genetic polymorphisms, both across all studies and within each individual study design. Both polymorphisms are common with minor allele frequencies of 0.11 and 0.18 in populations of European ancestry. Relative risk estimates varied widely in magnitude across published studies and across case–control and cross-sectional studies, with 50–78% of variation because of heterogeneity rather than chance, but not across prospective cohort studies. Overall odds ratios were modest: 1.67 per risk allele for the polymorphism on chromosome 4q25 and 1.21 for the polymorphism on 16q22. The very high odds ratios reported in some studies were confined to individual case–control or cross-sectional samples.
Few studies have addressed between-study heterogeneity of genetic risk estimates. Our findings are in contrast to those of previous studies in which minimal heterogeneity of polymorphisms associated with coronary artery disease was observed [2, 3]; this highlights that GWA studies constitute a useful tool for the discovery of polymorphisms associated with disease but that risk estimates need to be considered in the context of the population in which the polymorphisms were identified. The origin of the observed heterogeneity remains uncertain, but is likely to relate to the aetiological heterogeneity of the AF diagnosis. Although a stronger association with a family history of AF has been reported for AF presenting at a young age [11, 12], we did not detect any effect interaction with age in meta-regression analyses. A particularly strong association with family history has been reported in individuals presenting with lone AF , and the findings of a previous study support a larger effect of the association for at least the polymorphism on chromosome 4q25 with lone AF . In the light of these observations, the wide variation in effect estimates in case–control studies was not entirely unexpected, as they vary greatly in exclusion and inclusion criteria and ascertainment method. For example, cases with concomitant heart failure, valvular disease or hyperthyroidism have been excluded from some studies. The observation of heterogeneity across cross-sectional studies was more unexpected. Cohorts differ in age distribution, but whereas age-dependent genotypic effects have been described for genetic polymorphisms associated with other conditions , meta-regression analyses did not support any association between sample age and effect size for AF. Effect heterogeneity across populations could also be due to varying correlation of genotyped polymorphisms with putative, ungenotyped polymorphisms constituting the actual causal variants underlying the associations with AF. However, a recent fine-mapping study of the 4q25 region did not find any stronger association than with the polymorphism genotyped here , suggestive that it might be the actual causal variant. It appears more likely that the observed heterogeneity across cross-sectional studies is explained by differences in population characteristics based on differences in recruitment, exclusion criteria and end-point ascertainment. For example, participants with prevalent HF, valvular disease or coronary artery disease were excluded from genotyping in the Cardiovascular Health Study. These observations have implications for the interpretation of risk estimates from genetic studies of AF and other complex diseases.
It has been suggested that information about family history and genetic polymorphisms should be added to current prediction models for AF in the general population . Although guidelines from international arrhythmology societies do not recommend genotyping for these polymorphisms , direct-to-consumer genotyping is available from commercial companies. Our findings suggest the feasibility of genetic risk prediction based on risk estimates from prospective cohort studies. Heterogeneity across prospective cohort studies was limited despite different age distributions at baseline, ascertainment methods and follow-up times. However, risk estimates are relatively modest and unlikely to improve predictive accuracy. Commercial companies offering genotyping need to consider carefully the context of estimates provided to customers. Furthermore, because of the lack of effective preventive interventions, tests for these polymorphisms are unlikely to be clinically useful.
The results of our meta-analysis do not support an association between AF and the K897T missense variant in KCNH2. Although replication was claimed in the initial report, the association could not be replicated in another recent large case–control sample with similar age distribution and clinical characteristics to the sample in which the association was first identified  or in a large prospective population-based study from our group . Using the Genetic Power Calculator (http://pngu.mgh.harvard.edu/~purcell/gpc/), we estimated that the power to detect (P <0.05) an additive odds ratio of 1.25 as in the original report  or of 1.10 as in the meta-analysis approached 100% for both effect sizes, based on the current sample size and observed minor allele frequency for the K897T variant. Thus, the findings of the present meta-analysis do not support the large effect described in the initial report , but the possibility of a smaller effect cannot be excluded.
When this meta-analysis was performed, only the initial discovery studies had examined the association of SNPs on 1q21 and GJA5 with lone AF, and the association of SNPs near IL6R with AF. A number of SNPs associated with PR interval have also recently been associated with AF, although this finding has not been replicated . Even more recently, the number of loci associated with AF has increased from six to 12 following the report of a large meta-analysis . Future studies will be needed to evaluate the robustness of effect estimates for these SNPs and the association with AF in the general population. Furthermore, only limited data were available in populations of primarily non-European ancestry. Additional studies in such populations are warranted.
In conclusion, two genetic polymorphisms are robustly associated with AF irrespective of study design; however, effect estimates are consistent across prospective cohort studies, but vary considerably across case–control and cross-sectional studies. These findings have general implications for the interpretation of risk estimates from GWA studies and establish the feasibility of individual risk prediction of AF based on genetic polymorphisms in a prospective setting, although effect estimates are modest.
Conflict of interest statement
No conflict of interest was declared.
Drs Smith, Hedblad, Engström, Melander and Platonov gratefully acknowledge financial support from the Swedish Heart-Lung Foundation. Dr Smith was also supported by Skåne University Hospital, the Medical Faculty of Lund University and the Thorsten Westerström Foundation. Dr Newton-Cheh was supported by NIH grant K23-HL-080025, a Doris Duke Charitable Foundation Clinical Scientist Development Award, and a Burroughs Wellcome Fund Career Award for Medical Scientists. Drs Hedblad and Melander were supported by the Swedish Medical Research Council. Dr Melander was supported by grants from the European Research Council (StG-282255), the Medical Faculty of Lund University, Skåne University Hospital in Malmö, the Albert Påhlsson Research Foundation, the Crafoord Foundation, the Swedish National Health Service, the Hulda and Conrad Mossfelt Foundation, the Ernhold Lundströms Research Foundation, the King Gustaf V and Queen Victoria Fund, the Lennart Hanssons Memorial Fund, the Marianne and Marcus Wallenberg Foundation and the Knut and Alice Wallenberg Foundation. Dr Platonov was supported by the Swedish National Health Service, Skåne University Hospital and the Craaford Foundation.