Multilocus genetic risk score associates with ischemic stroke in case-control and prospective cohort studies

Abstract

BACKGROUND AND PURPOSE

Genome-wide association studies have revealed multiple common variants associated with known risk factors for ischemic stroke (IS). However, their aggregate effect on risk is uncertain. We aimed to generate a multilocus genetic risk score (GRS) for IS based on genome-wide association studies data from clinical-based samples and to establish its external validity in prospective population-based cohorts.

METHODS

Three thousand five hundred forty-eight clinic-based IS cases and 6399 controls from the Wellcome Trust Case Control Consortium 2 were used for derivation of the GRS. Subjects from the METASTROKE consortium served as a replication sample. The validation sample consisted of 22 751 participants from the Cohorts for Heart and Aging Research in Genomic Epidemiology consortium. We selected variants that had reached genome-wide significance in previous association studies on established risk factors for IS.

RESULTS

A combined GRS for atrial fibrillation, coronary artery disease, hypertension, and systolic blood pressure significantly associated with IS both in the case-control samples and in the prospective population-based studies. Subjects in the top quintile of the combined GRS had >2-fold increased risk of IS compared with subjects in the lowest quintile. Addition of the combined GRS to a simple model based on sex significantly improved the prediction of IS in the combined clinic-based samples but not in the population-based studies, and there was no significant improvement in net reclassification.

CONCLUSIONS

A multilocus GRS based on common variants for established cardiovascular risk factors was significantly associated with IS both in clinic-based samples and in the general population. However, the improvement in clinical risk prediction was found to be small.

Introduction

Stroke is the leading cause of permanent disability and the third most common cause of death in high-income nations.^1,2 Approximately 80% of stroke cases are caused by ischemia, with large artery atherosclerosis and cardioembolism from atrial fibrillation (AF) being among the most common mechanisms.

Ischemic stroke (IS) is highly heritable.^3,4 Probably reflecting some of the heritability, recent genome-wide association studies (GWAS) have identified common variants in several genomic regions that are associated with IS⁵ (L.L. Kilarski et al, unpublished data, 2013) or specific stroke subtypes such as large artery stroke.^6,7 Furthermore, GWAS have identified multiple variants in multiple genomic regions that are associated with established risk factors for IS, including AF,⁸ coronary artery disease,^9,10 and hypertension.^11,12 Several of these variants were also found to be associated with IS risk typically through their association with specific stroke mechanisms, such as large artery disease^6,13–15 or cardioembolism,^16,17 although some loci were also found to be associated with IS as a whole⁵ (L.L. Kilarski et al, unpublished data, 2013). In general, the observed increase in risk associated with individual variants was found to be small usually in the range of 1.2 to 1.6. However, because most risk alleles are common in people of European ancestry, their effect on a population level is likely to be substantial. Current risk prediction models, based on conventional risk factors,¹⁸ perform reasonably well and are used to guide clinical decision making. Yet, efforts to add emerging factors into risk prediction scores continue^19–21 because even incremental improvements in predictive performance might lead to clinically meaningful changes in risk classification. Combining the effects of genetic risk variants into multilocus genetic risk scores (GRSs) may aid in risk prediction. In fact, recent studies have demonstrated significant improvements in coronary risk prediction using multilocus GRS for coronary artery disease.^22–25 Despite the increase in direct-to-consumer testing for genetic variants for stroke,²⁶ the use of adding genetic variants to stroke risk prediction has not been evaluated systematically.

The purpose of this study therefore was to construct a multilocus GRS based on common variants previously shown to reach genome-wide significance for an association with known risk factors for stroke. Single nucleotide polymorphisms (SNPs) were identified on the basis of systematic literature and database review using predefined criteria. We hypothesized that a GRS derived from clinic-based case–control samples would replicate in independent clinic-based case–control samples and improve the ability to predict IS in population-based cohorts.

Methods

Study Sample

The clinic-based case–control sample for derivation of the GRS consisted of 3548 IS cases and 6399 controls from the Wellcome Trust Case Control Consortium 2.⁶The clinic-based replication sample included 3856 IS cases and 4069 controls from the METASTROKE consortium.¹⁵ The prospective population-based cohorts included 22 751 participants from the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) consortium.²⁷ A detailed description of these cohorts together with information on the methods used for genotyping is provided in Table 1 and the online-only Data Supplement.

Table 1.

Description of Case–Control and Population-Based Cohorts Used in the Analysis

	Age, y	Women, %	No. of Cases	No. of Controls	No. of Incident Cases	Total No. of Subjects	Ancestry	Genotyping
Clinic-based derivation samples
WTCCC2-UK	72.2±12.5	48.5	2374	5562	…	7936	CEU	Illumina 660 (cases)
								Illumina 1.2M-Duo (controls)
WTCCC2-Munich	66.6±12.9	42.1	1174	837	…	2011	CEU	Illumina 660 (cases)
								Illumina 550 (controls)
Clinic-based replication samples
ASGC	72.9±13.2	45.5	1177	1244	…	2421	CEU	Illumina 610
MGH-GASROS	66.6±14.5	42.9	767	395	…	1162	CEU	Affymetrix 6.0
BRAINS	74.4±7.2	46.7	394	444	…	838	CEU	Illumina 610
GEOS	41.0±6.9	41.1	448	498	…	946	CEU	Illumina HumanOmni1
ISGS/SWISS	66.5±13.6	48.4	1070	1488	…	2558	CEU	Illumina 650 (cases)
								Illumina 550 (controls)
Prospective population-based cohorts
ARIC	54.3±5.7	52.8	…	…	437 (4.6%)	9349	CEU	Affymetrix 6.0
CHS	72.3±5.4	60.9	…	…	453 (13.9%)	3268	CEU	Illumina 370
FHS	66.1±12.6	54.7	…	…	198 (4.5%)	4371	CEU	Affymetrix 550
Rotterdam	69.1±9.0	40.6	…	…	467 (8.1%)	5763	CEU	Illumina 550

ARIC indicates The Atherosclerosis Risk in Communities study; ASGC, Australian Stroke Genetics Collaborative; BRAINS, Bio-Repository of DNA in stroke; CEU, Caucasian European; CHS, Cardiovascular Health Study; FHS, Framingham Heart Study; GEOS, Genetics of Early Onset Stroke; ISGS, Ischemic Stroke Genetics Study; MGH-GASROS, Massachusetts General Hospital—Genes Associated with Stroke Risk and Outcomes Study; SWISS, Siblings With Ischemic Stroke Study; WTCCC2-UK, The Wellcome Trust Case Control Consortium II UK; and WTCCC2-Munich, The Wellcome Trust Case Control Consortium II Munich.

Selection of Genetic Variants

We selected SNPs from GWAS on modifiable risk factors for IS as defined by the American Heart Association stroke statistics and the American Heart Association/American Stroke Association guideline on primary prevention of stroke^28,29 that were published before January 2012. This included GWAS on the following risk factors: hypertension, diastolic blood pressure, systolic blood pressure, smoking/tobacco use, type 1 diabetes mellitus, type 2 diabetes mellitus, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, overweight and obesity, AF, hypertriglyceridemia, coronary artery disease, myocardial infarction, alcohol abuse, uric acid, elevated circulating urate levels, and hyperhomocysteinemia (Table I in the online-only Data Supplement).

The National Human Genome Research Institute GWAS catalogue³⁰ was used as the primary source for published GWAS. We considered studies assessing >100 000 markers in the discovery stage and variants reaching a global P value of <1E–5.^31,32 Findings from candidate gene studies were excluded as were studies based on somatic mutations (eg, in cancer cells) compared with naturally occurring mutations. We further searched PubMed using the following MeSH terms “risk factor” AND “GWAS” OR “genome wide” OR “genome wide association.” The resulting citations and abstracts were reviewed manually as were cited articles in the selected publications. Studies on non-Central European populations and studies where the effect allele was not reported were disregarded. A full list of the 521 SNPs included in the analysis is provided in Table I in the online-only Data Supplement.

Weighting Scheme and Calculation of Risk Scores

SNPs were weighted by their estimated effect sizes (β) provided in the original reports (Table I in the online-only Data Supplement). For SNPs derived from quantitative trait studies, we arbitrarily used a uniform weight of 0.1, corresponding to an odds ratio (OR) of 1.105 (weak to moderate effect size) while accounting for directionality of effects. This was performed to account for the different scales and measures that were used in the original studies to measure the effects of SNPs on the respective traits.

Weighted multilocus GRSs (wGRSs) for individual risk factors (eg, hypertension) were calculated using the – score function implemented in PLINK³³ and an additive model. Risk profile scores were derived by adding the number of risk alleles multiplied by the weight of the risk variant. SNPs with missing information were excluded from the model. Scores were expressed as the mean score per SNP in the set. The combined GRSs (cGRSs) were calculated by adding Z score transformed wGRS. Z scores were used to account for the variable numbers of risk alleles constituting the wGRS.

Statistical Analysis

For the case–control studies, risk profiles were analyzed by generalized linear models using logistic regression in R with phenotype (case or control) as the outcome variable and the risk profile score as the predictor variable. In the absence of additional information in the control cohorts, sex was used as the only covariate. To account for intrinsic genetic differences between cohorts, we added an indicator variable for recruiting site in all analyses of the merged derivation and replication sample. Step-wise logistic regression was performed using the stepAIC function in the MASS library in R. All normalized wGRSs were included in the analysis to discover the optimal set of wGRS for inclusion in the cGRS by exact Akaike information criterion. Odds ratios for the wGRS and cGRS are reported as an increase of risk per improvement of 1 SD of the respective GRS. The variance in case/control status explained by the score statistic was estimated as the difference in variance using Nagelkerke pseudo-R² between a model including the cGRS, sex, and study site (full model) versus the covariates alone (reduced model). To evaluate the potential value of the cGRS in risk prediction, we compared the receiver operating characteristic curves of models with and without cGRS (lroc package for logistic regression in R). C-statistics were computed to assess the gain in predictive power of the cGRS and of the full versus reduced model.

For the prospective population-based cohorts, we used Cox regression models to evaluate the association of the cGRS with the incident IS and an R² measure to estimate variance explained by the model. The statistical significance of change in the area under the receiver operating characteristic curve (AUC) between models was tested with the correlated C-index approach. C-statistics were meta-analyzed using the point estimates with SEs as input followed by inverse-variance meta-analysis. Significance of the full versus reduced model was evaluated using a method reported by Hanley and McNeil,³⁴ which accounts for the correlation between both models. This correlation between the AUCs was obtained using the metacor package for R. The continuous net reclassification indices was calculated using published methods.³⁵

Results

The clinic-based derivation sample included 3548 IS cases and 6399 controls from the Wellcome Trust Case Control Consortium 2 (WTCCC2; merged WTCCC2-UK and WTCCC2-Munich). Three thousand eight hundred fifty-six clinic-based cases and 4069 controls from METASTROKE served as a replication sample. A total of 22 751 participants from CHARGE (the Atherosclerosis Risk in Communities Study [ARIC], Cardiovascular Health Study [CHS], Framingham Heart Study [FHS], and Rotterdam) were included in the prospective cohort analyses. Median follow-up in the prospective cohorts was 14.4 years (interquartile range, 8.5). One thousand five hundred fifty-five (6.8%) incident cases of IS occurred during follow-up. Table 1 provides background characteristics for the clinic-based case–control samples and the population-based cohorts.

Weighted GRS for Individual Risk Factors

We first analyzed wGRS for individual risk factors (n=16) for IS in the merged clinic-based WTCCC2-UK and WTCCC2-Munich sample encompassing a total of 3548 cases and 6399 controls. Significant associations were found with wGRS for AF, coronary artery disease, diastolic blood pressure, hypertension, and systolic blood pressure (Table 2). Point estimates for OR in the merged derivation samples varied between 1.06 and 1.09 with similar effect sizes in the 2 derivation samples. Density plots for the wGRS are displayed in Figure 1. Analysis of the clinic-based replication and merged derivation and replication sample confirmed associations with the wGRS (Table II in the online-only Data Supplement) except for the wGRS for hypertension, which was not significant in the replication sample but reached significance in the merged derivation and replication sample.

Table 2.

Results for the wGRS for Individual Risk Factors in the Derivation Samples

	No. of Variants in wGRS	WTCCC2-UK Derivation Sample		WTCCC2-Munich Derivation Sample		Merged Derivation Sample
Risk Factor	Odds Ratio (95% CI)	P Value	Odds Ratio (95% CI)	P Value	Odds Ratio (95% CI)	P Value
Established RF28
AF	14	1.082 (1.031–1.134)	1.16E-03*	1.061 (0.970–1.160)	1.99E-01	1.076 (1.032–1.122)	6.19E-04*
BMI+weight	95	1.007 (0.960–1.057)	7.68E-01	1.000 (0.915–1.093)	1.00	1.005 (0.963–1.050)	8.09E-01
CAD	51	1.115 (1.062–1.170)	1.04E-05*	1.017 (0.930–1.111)	7.20E-01	1.091 (1.046–1.139)	5.84E-05*
DBP	37	1.074 (1.024–1.127)	3.50E-03*	1.032 (0.944–1.128)	4.93E-01	1.064 (1.020–1.110)	4.21E-03*
HDL	55	0.996 (0.949–1.045)	8.64E-01	1.012 (0.926–1.107)	7.89E-01	0.999 (0.957–1.042)	9.54E-01
HTN	19	1.048 (0.999–1.099)	5.39E-02	1.091 (0.998–1.194)	5.59E-02	1.059 (1.014–1.105)	8.25E-03*
LDL	39	1.029 (0.981–1.080)	2.37E-01	1.026 (0.938–1.121)	5.81E-01	1.030 (0.987–1.074)	1.75E-01
MI	12	1.064 (1.014–1.116)	1.19E-02*	0.977 (0.894–1.069)	6.16E-01	1.043 (0.999–1.088)	5.27E-02
SBP	29	1.077 (1.027–1.131)	2.40E-03*	1.037 (0.949–1.134)	4.19E-01	1.069 (1.024–1.115)	2.13E-03*
Smoking	11	0.996 (0.950–1.045)	8.80E-01	0.923 (0.845–1.010)	8.03E-02	0.980 (0.939–1.022)	3.47E-01
T1D	42	1.001 (0.954–1.050)	9.73E-01	1.010 (0.924–1.104)	8.27E-01	1.002 (0.961–1.045)	9.20E-01
T2D	62	1.010 (0.962–1.059)	6.94E-01	0.968 (0.886–1.059)	4.78E-01	1.001 (0.960–1.044)	9.63E-01
Triglyceride levels	49	0.995 (0.948–1.044)	6.84E-01	1.039 (0.951–1.136)	3.99E-01	1.005 (0.964–1.049)	8.08E-01
Less well-documented RF28
Alcohol dependence	26	1.025 (0.977–1.076)	3.10E-01	0.885 (0.809–0.967)	7.35E-03*	0.991 (0.950–1.034)	6.90E-01
Homocysteine levels	34	1.010 (0.963–1.060)	6.73E-01	1.004 (0.919–1.010)	9.28E-01	1.010 (0.968–1.053)	6.54E-01
Urate levels	21	0.990 (0.944–1.039)	8.23E-01	1.036 (0.948–1.133)	4.33E-01	1.000 (0.958–1.044)	9.98E-01

AF indicates atrial fibrillation; BMI, body mass index; CAD, coronary artery disease; CI, confidence interval; DBP, diastolic blood pressure; HDL, high-density lipoprotein; HTN, hypertension; LDL, low-density lipoprotein; MI, myocardial infarction; RF, risk factor; SBP, systolic blood pressure; wGRS, weighted genetic risk scores; WTCCC2-UK, The Wellcome Trust Case Control Consortium II UK; and WTCCC2-Munich, The Wellcome Trust Case Control Consortium II Munich.

^*Statistically significant risk factors at the P<0.05 level.

Figure 1.

Distribution of genetic risk score (GRS) in patients with ischemic stroke and controls in the merged derivation sample (WTCCC2-UK and WTCCC2-Munich). Shown are weighted GRSs (wGRSs) for atrial fibrillation, coronary artery disease, diastolic blood pressure, hypertension, and systolic blood pressure. Bottom right, The combined GRS (cGRS). Red line represents the distribution of the risk score in cases; and black line, the distribution of the risk score in controls. Note that the wGRS for AF is constituted by variants from 2 risk loci, which explains the bimodal distribution.

Combined Risk Score

We next tested combinations of the 16 wGRS for their association with IS. In step-wise logistic regression, a cGRS generated from wGRS for AF, coronary artery disease, hypertension, and systolic blood pressure encompassing a total of 113 variants (Table III in the online-only Data Supplement) showed the strongest association with IS in the merged WTCCC2-UK and WTCCC2-Munich sample (OR, 1.07 [95% confidence interval {CI}, 1.04–1.09]; P=5.78E–10; Table 3; Figure 1). The cGRS was found to replicate in the clinic-based replication sample (OR, 1.03 [95% CI, 1.01–1.04]; P=1.67E–03). Combining the derivation and replication samples and adding an indicator variable for recruiting site resulted in an OR similar to that found in the derivation sample (OR, 1.06 [95% CI, 1.04–1.08]; P=4.9E–07; Table 3).

Table 3.

Results for the cGRS^* in the Clinic-Based Derivation and Replication Samples and in the Population-Based Sample

Sample	Odds Ratio (95% CI)	P Value	ΔR2, %	ΔAUC, % (PValue)
Clinic-based derivation (WTCCC2-UK+WTCCC2-Munich; 3548 cases and 6399 controls)	1.065 (1.044–1.087)	5.78E-10	0.502	1.71 (1.5E-06)
Clinic-based replication (3856 cases and 4069 controls)	1.026 (1.009–1.043)	1.67E-03	0.041	0.19 (0.11)
Combined clinic-based derivation and replication (7404 cases and 10 468 controls)	1.059 (1.036–1.083)	4.87E-07	0.179	0.42 (1.8E-06)
Population-based (CHARGE; 1554 incident cases among 22 276 participants)	1.027 (1.005–1.049)	1.57E-02	N/A	0.11 (0.649)

AUC indicates area under the receiver operating characteristic curve; cGRS, combined genetic risk score; CHARGE, Cohorts for Heart and Aging Research in Genomic Epidemiology; CI, confidence interval; N/A, not assessed; WTCCC2-UK, The Wellcome Trust Case Control Consortium II UK; and WTCCC2-Munich, The Wellcome Trust Case Control Consortium II Munich.

^*Generated from wGRS for atrial fibrillation, coronary artery disease, hypertension, and systolic blood pressure. The score included a total of 113 variants (for details see Table III in the online only Data Supplement).

When we divided the cGRS in the derivation sample into quintiles, we found a linear increase in IS risk across quintiles (Cochran–Armitage test for linear trend across quintiles P=1E–13; Figure 2A). Compared with subjects in the third quintile (reference), subjects in the top quintile were estimated to have a 1.42× increased risk of IS (95% CI, 1.03–1.95), whereas subjects in the bottom quintile were estimated to have a 0.35-fold risk of IS (95% CI, 0.26–0.55). Similar results were obtained in the combined clinic-based derivation and replication sample (Figure 2B), although effect sizes were smaller (bottom quintile, 0.59 [95% CI, 0.41–0.84]; top quintile, 1.32 [95% CI, 0.97–1.78]).

Figure 2.

Odds ratios for risk categories defined using the combined genetic risk score. A, Clinic-based derivation sample. B, Merged clinic-based derivation and replication sample. The primary (left) y axis displays the percentage of cases and controls in each quintile, the secondary (right) y axis displays the odds ratio (OR) associated with each quintile where the middle (third) quintile serves as a reference.

Improvement of Prediction Quality

To determine the predictive value of the cGRS, we calculated the improvement in prediction quality of the cGRS compared with a model that included sex and study site (reduced model). The predictive strength, that is, the proportion to which the model explains variation in the data set, was 8.06% when using the cGRS in addition to all covariates (full model) and 7.56% when using only the covariates (reduced model). Hence, the cGRS resulted in an increase of 0.5% in predictive power (Table 2). This difference in explained variance was lower in the replication samples alone (0.041%) and was 0.179% in the combined clinic-based derivation and replication sample. The improvement in the c-statistic (AUC) for the cGRS was significant both in the clinic-based derivation sample (AUC_{full model}=62.75%, AUC_{reduced model}=61.04%, ΔAUC=1.71%; P=1.5E–06) and in the combined derivation and replication samples (AUC_{full model}=64.51%, AUC_{reduced model}=64.09%, ΔAUC=0.42%; P=1.8E–06; Table 3; Figure 3A and 3B).

Figure 3.

Receiver operating characteristic curves for models predicting a diagnosis of ischemic stroke in the derivation sample (A) and in the merged clinic-based derivation and replication sample (B). The reduced model of only covariates is shown in green, the full model including the combined genetic risk score (cGRS) in purple. The blue line represents the cGRS without covariates. The black dashed diagonal line represents a random prediction.

Validation of the cGRS in Prospective Population-Based Cohorts

To ascertain the validity and predictive power of the cGRS in stroke-free individuals, we meta-analyzed data from the prospective population-based ARIC, CHS, FHS, and Rotterdam cohorts. The pooled hazard ratio of the cGRS for incident IS was 1.03 (95% CI, 1.01–1.05; P=0.016) with no significant heterogeneity between studies (I²=19.9%; Figure I in the online-only Data Supplement).

AUC analyses in individual cohorts revealed largely variable c-statistics (AUC_{full model}=57.6%–72.2%), with no significant improvements in risk prediction in a combined meta-analysis (ΔAUC=0.11%; P=0.649; Table 3). The improvement of R² with the full model compared with the reduced model ranged between 0.1% and 0.3% in individual cohorts.

We also assessed net reclassification results for the population-based cohorts. The continuous boundary-less net reclassification indices in individual cohorts ranged between −0.0101 and 0.1108, with no significant improvement in total reclassification.

Discussion

We found that a multilocus GRS composed of 113 common variants predicted IS risk. The cGRS was derived from clinic-based samples, replicated in independent samples, and validated using incident cases from prospective population-based cohorts. Subjects in the top quintile of the cGRS had >2-fold increased risk of IS when compared with subjects in the lowest quintile. However, the improvement in risk prediction by adding the cGRS to a simple model with sex and study site alone was small and not significant in the prospective validation cohort.

Considering >500 variants that had been shown previously to be associated with known risk factors for IS, we found weighted GRS for AF, coronary artery disease, hypertension, and systolic and diastolic blood pressure to show the strongest association with IS in the clinic-based derivation sample. These risk factors match with risk factors showing the strongest predictive value in conventional risk prediction models for stroke^18,36 and with recent studies that have shown associations of individual risk alleles for AF, coronary artery disease, and hypertension with IS as a whole or with specific stroke subtypes.^5,13,15,16Together, these findings suggest that multiple genetic variants at multiple chromosomal loci influence IS risk, possibly via known risk factors for IS. We did not examine our cGRS in relation to established risk scores such as the Framingham Risk Score for Stroke because information on risk factors was incomplete in the case–control samples and because one would not expect to see significant improvements over and above a score that contains actual information on the presence of these risk factors. It has been suggested that GRS may be most useful earlier in life, that is, before phenotypic variation in the risk factors incorporated into conventional risk prediction scores typically manifest.²²However, our results suggest that the gain in predictive power by adding the GRS to information on sex alone is limited.

Our findings agree with studies in other conditions, including breast cancer,³⁷diabetes mellitus,^38,39 coronary artery disease,^23,40 or multiple sclerosis,⁴¹ that found limited improvement in risk prediction with GRS based on GWAS discoveries. Our approach differs from most studies, in which we also considered variants that have to date not been associated with the clinical phenotype itself but instead reached genome-wide significance for association with known risk factors for the phenotype of interest, in this case IS. This enabled us to consider a much larger number of variants than were included in previous efforts to generate GRS. The validity of this approach is demonstrated by our finding that wGRS for systolic blood pressure, diastolic blood pressure, and hypertension all significantly associated with IS in the clinic-based case–control samples. Nevertheless, the improvement in risk prediction was small.

The limited use of our specific GRS for IS and of GRS in many other conditions might relate to several factors. First, because of the presence of multiple common alleles with small effects, almost all individuals carry some risk alleles. Second, the majority of individuals have multiple risk alleles close to the mean number of risk alleles in the overall population with the minority having extreme numbers of risk alleles as also reflected by the distribution of the cGRS in the current study (Figure 2B). Third, effect sizes between risk variants vary, which means that individuals having the same risk score may differ with regard to genetic risk unless GRSs were weighted as was the case in the current study. Finally, risk estimates on some of the variants may be imprecise as the number of studies is still relatively small. This might be interpreted as a failure of personalized medicine using genomics. On the contrary, the low level of genetic variance explained by common risk alleles identified to date suggests that much of the genetic predisposition is still undiscovered.^42,43 Future studies will need to determine whether iterations of GRS incorporating additional common variants as well as low-frequency variants derived from whole-exome or whole-genome sequencing lead to clinically useful improvements in risk prediction.

The magnitude of effect of the cGRS in the combined case–control sample was higher than in the prospective population-based cohorts. This might be because of optimization of the cGRS in the derivation sample. Alternatively, the observed differences in OR might reflect different genetic architectures in clinic-based and incident cases from population-based cohorts. Compared with incident stroke cases from prospective population-based cohorts, clinic-based cases usually are younger^7,44,45 and have more vascular risk factors including AF⁴⁵ and hypertension.⁴⁶ Another reason might be the relatively low incidence of stroke in the population-based cohorts resulting in reduced power. Of note, however, the cGRS was significantly associated with both prevalent and incident IS in the current study.

This study has several methodological strengths including replication of the cGRS in an independent case–control sample and validation in community-based samples that had been followed for extended time periods and provided a large number of incident events. Together, these samples represent one of the largest collections of IS cases with genome-wide data available to date.¹⁵ However, this study also has limitations. First, our GRS was based on SNPs that had been selected on the basis of predefined criteria. Thus, many variants with an effect on IS may have been excluded. We did not include variants published after January 2012 because of logistic challenges in obtaining calculated scores from all the replication and validation samples. This also includes variants from recently published loci associated with large artery stroke^7,15 and coronary artery disease.¹⁰ Thus, our estimate on the predictive value and the explained variance of the cGRS likely is an underestimate of multilocus GRSs in IS. However, we consider it unlikely that the results would have been materially different with inclusion of those additional loci. Second, some quantitative traits could not be weighted for effect sizes because effects in the original studies had been reported on different scales. This likely resulted in an underestimation of effect sizes. Third, we did not analyze GRS for stroke subtypes because this information was not available in all validation cohorts. Finally, our sample consisted of white subjects of European descent. Thus our results may not be generalizable to other ethnicities.

We think our findings have clinical relevance because genetic testing is increasingly available and marketed to the public. We find no clinical use in constructing a multilocus panel of SNPs for stroke risk that extended to include variants acting on intermediate phenotypes such as hypertension or AF. It has been suggested previously that testing AF variants in well-defined populations such as patients with a cryptogenic stroke might aid in selecting subjects for further diagnostic procedures.^16,47 The current study was not designed to address this. However, we found a wGRS for AF to have relatively small effect sizes in otherwise unselected clinic-based samples. Regardless of these results, any strategies aimed at testing common genetic variants to inform clinical decision making would need to be rigorously tested before moving to clinical practice.

In conclusion, we found a multilocus cGRS derived from GWAS for established risk factors for IS to be significantly associated with IS risk. Odds ratios in the highest and lowest quintile of the cGRS differed substantially. However, the power of the cGRS in predicting IS risk and hence its clinical use was found to be limited. Future alternative strategies of constructing GRS for IS and combining GRS with risk factor profiles and clinical information might eventually result in better risk prediction.