Genome-wide association study identifies a variant in HDAC9 associated with large vessel ischemic stroke.

Genetic factors have been implicated in stroke risk, but few replicated associations have been reported. We conducted a genome-wide association study (GWAS) for ischemic stroke and its subtypes in 3,548 affected individuals and 5,972 controls, all of European ancestry. Replication of potential signals was performed in 5,859 affected individuals and 6,281 controls. We replicated previous associations for cardioembolic stroke near PITX2 and ZFHX3 and for large vessel stroke at a 9p21 locus. We identified a new association for large vessel stroke within HDAC9 (encoding histone deacetylase 9) on chromosome 7p21.1 (including further replication in an additional 735 affected individuals and 28,583 controls) (rs11984041; combined P = 1.87 × 10(-11); odds ratio (OR) = 1.42, 95% confidence interval (CI) = 1.28-1.57). All four loci exhibited evidence for heterogeneity of effect across the stroke subtypes, with some and possibly all affecting risk for only one subtype. This suggests distinct genetic architectures for different stroke subtypes.

Cerebrovascular disease (stroke) is one of the three most common causes of death and the major cause of adult chronic disability (1). Stroke represents an increasing health problem throughout the world as the proportion of elderly increases, and is an important cause of dementia and age-related cognitive decline. While conventional risk factors such as hypertension account for a significant proportion of stroke risk, much remains unexplained (2). Twin and family history studies suggest genetic factors are responsible for some of this unexplained risk (3). Stroke is a syndrome rather than a single disease, and subtypes of stroke are caused by a number of different specific disease processes. About 80% of stroke is ischemic; the three most common ischemic stroke subtypes are large vessel, cardioembolic and small vessel (lacunar) stroke. Genetic epidemiological studies show heterogeneity between stroke subtypes, the large vessel subtype being more strongly associated with family history (4). SNPs associated with atrial fibrillation were found only to be significantly associated with cardioembolic stroke (5,6), and a 9p21 variant initially associated with coronary artery disease and atherosclerosis only associated with large vessel stroke (7). This suggests that different genetic variants can predispose to different subtypes of ischemic stroke.

To date there have been few genome wide association studies (GWAS) in ischemic stroke and few replicable associations have been identified (8). To further understand the genetic basis of ischemic stroke, we undertook a GWAS as part of the Wellcome Trust Case Control Consortium 2 (WTCCC2). We hypothesised that associations might be present only with specific stroke subtypes. To investigate this, cases were classified into stroke subtypes according to the pathophysiological TOAST classification (9), using clinical assessment as well as brain and vascular imaging where available (see Online Methods). Association analyses were performed on all ischemic stroke combined (including individuals not further classified by stroke subtype), and also the three major stroke subtypes: large vessel, small vessel and cardioembolic stroke. Discovery samples were of European ancestry and were genotyped on Illumina arrays (see Online Methods). Following quality control, the discovery set consisted of 3,548 cases (2,374 British, 1,174 German) and 5,972 controls (5,175 British WTCCC2 common controls, and 797 German controls) genotyped on an overlapping set of 495,851 autosomal SNPs (Table 1 and Online Methods). Within the British and German data, cases and controls were well matched for ancestry (see Online Methods and Supplementary Figure 1). We therefore performed association analysis separately in the two groups and combined them using a fixed effect meta-analysis approach. A two-stage replication study was performed in 5,859 cases (3,863 European, 1,996 American) and 6,281 controls (4,554 European, 1,727 American) all of self-reported European ancestry. (Table 1 and Online Methods). Full details of the cohorts are available in the Supplementary material. and Supplementary Table 1.

Table 1

Post quality control breakdown of case and control by cohort and ischaemic stroke subtype

All: all ischaemic stroke, LVD: large vessel stroke, SVD: small vessel stroke, CE: cardioembolic stroke. Note that not all strokes are classified into a subtype.

		All strokes	LVD	CE	SVD	Controls
DISCOVERY	Munich	1174	346	330	106	797
	UK[1]	2374	498	460	474	5175
	Total	3548	844	790	580	5972
STAGE 1 REPLICATION - EUROPEAN	Krakow	1214	152	362	170	551
	Leuven	418	63	154	52	650
	Lund	428	21	139	97	465
	Munich[2]	54	19	16	5	310
	UK[3]	1749	306	303	490	2578
	Total	3863	561	974	814	4554
STAGE 2 REPLICATION - US	Boston	533	150	206	56	522
	Cincinnati	438	67	106	90	257
	GEOS	419	37	90	54	498
	ISGS	606	121	156	111	450
	Total	1996	375	558	311	1727
STAGE 1 + STAGE 2 REPLICATION	Total	5859	936	1532	1125	6281
DISCOVERY + REPLICATION	Total	9407	1780	2322	1705	12253

The UK discovery cohort was made of three British cohorts from London, Oxford and Edinburgh and used the shared WTCCC2 controls.

The Munich replication samples comprised some samples planned for the discovery GWAS where there was insufficient DNA for GWAS but sufficient for replication. It used controls from a German cohort enrolled in the PROCARDIS trial.

The UK replication cohorts included samples from Aberdeen, Glasgow and Imperial as well as some samples planned for the discovery GWAS where there was insufficient DNA for GWAS but sufficient for replication (see methods). The UK replication cohorts used shared POBI controls genotypes as part of the WTCCC2.

Table 2 shows results at previously reported loci and Figure 1 shows the association analysis results across the autosomes. We replicated an association between cardioembolic stroke and variants close to the PITX2 gene and also a SNP in the ZFHX3 gene, both of which were initially associated with atrial fibrillation, a well recognised risk factor for stroke (5,6,10). We also replicated a previously reported association between large vessel stroke and the 9p21 region (7). As we, and others, already reported (11,12), we did not confirm the previously published association between all stroke and variants in the 12p13 region (13, 14).

Table 2

Association signals at the newly associated locus (upper tier) and at loci previously reported as associated with stroke or one of the stroke subtype (lower tier)

All: all ischaemic stroke, LVD: large vessel stroke, CE: cardioembolic stroke. For PITX2 and ZFHX3, results are given for one SNP reported in the literature and the SNP that showed the strongest association signal in the discovery samples. For the 9p21 region, the SNP reported in the literature is also the one showing the strongest association signal in the discovery cohort. There is some overlap between samples in this study and previous published studies of associations with above loci (5,6,10,13). A SNP in PRKCH, associated with stroke in Japanese populations (23) was very rare (<0.5%) in our Caucasian population so we had no power to perform an analysis of association with this SNP.

Chr	rsID	Position[6]	Candidategene	Strokesubtype	Riskallele	RAF[7]	Discovery	Stage 1&2	Stage3	Combined
							P-value	P-value (one-sided)	P-value (one-side)	P-value
							OR(95% CI)	OR (95% CI)	OR (95% CI)	OR(95%CI)
7p21.1	rs11984041[1,2]	18,998,460	HDAC9	LVD	A	0.09	1.07E-05	7.90E-05	2.25E-04	1.87E-11
							1.50 (1.25-1.79)	1.38 (1.17-1.63)	1.39 (1.15-1.68)	1.42 (1.28-1.57)
4q25	rs2200733[2,3,4]	111,929,618	PITX2	CE	A	0.10	3.64E-06	3.99E-04	-	5.06E-8
							1.49 (1.26-1.77)	1.24 (1.09-1.41)	-	1.32 (1.20-1.46)
	rs19065993	111,932,135			A	0.19	3.45E-08	3.16E-04	-	1.39E-09
							1.45 (1.27-1.66)	1.19 (1.08-1.32)	-	1.28 (1.18-1.39)
9p21.3	rs23832073	22,105,959	CDKN2A,CDKN2B	LVD	G	0.51	2.35E-03	2.03E-03	-	2.93E-05
							1.18 (1.06-1.31)	1.16 (1.05-1.28)	-	1.17 (1.09-1.25)
12p13.33	rs11833579[3,5]	645,460	NINJ2	All	G	0.75	9.65E-01	5.25E-01	-	9.81E-01
							1.00 (0.92-1.08)	1.00 (0.94-1.06)	-	1.00 (0.95-1.05)
16q22.3	rs7193343[3]	71,586,661	ZFHX3	CE	A	0.16	1.94E-05	-	-	-
							1.36 (1.18-1.57)	-	-	-
	rs12932445[2]	71,627,389			G	0.17	3.91E-07	4.84E-02	-	1.44E-05
							1.44 (1.25-1.66)	1.09 (0.98-1.21)	-	1.20 (1.11-1.31)

Krakow replication samples were not considered because of Hardy-Weinberg test p-value < 5×10−4 in controls.

SNP imputed in GEOS replication samples.

SNP reported in the literature.

ISGS replication samples were not considered because of Hardy-Weinberg test p-value < 5×10−4 in controls.

SNP imputed in the British discovery samples and not genotyped nor imputed in the discovery and replication German controls.

NCBI human genome build 36 coordinates.

Risk allele frequency computed in the British discovery controls

Figure 1

Genome-wide association results at autosomal SNPs in combined British and Germany discovery samples

All ischemic stroke (top panel), large vessel stroke (second panel), small vessel stroke (third panel) and cardioembolic stroke (final panel). Loci previously reported in the literature for particular stoke subtypes (Table 2) are shown in black, with the new HDAC9 locus shown in red. In addition the combined p-values for discovery study and stages one and two of the replication study, at the top SNPs for these loci, are marked with diamonds.

Thirty-eight previously unreported loci showed potential association for all stroke or one of the stroke subtypes in the discovery samples, and we further investigated these loci in the European replication samples by genotyping 43 SNPs covering these loci as well as 7 SNPs to cover the previously reported loci (Supplementary Table 2). Thirteen of these previously unreported loci and the previously reported loci were taken forward to replication in the American samples with genotyping of 20 SNPs covering these regions (Supplementary Table 3). Most replication samples were genotyped using Sequenom assays; for those previously typed with GWAS chips we used genotype imputation where the SNP was not directly typed (see Supplementary Tables 2 and 3 and Online Methods). A SNP at chromosome 7p21.1 (rs11984041) showed evidence of association with large vessel stroke in the discovery data (P=1.07×10−5) and in the joint European and US replication data in the same direction (one-sided P=7.9×10−5). As a further check, we investigated this SNP in three further collections of large vessel cases and matched controls (735 cases, 28583 controls in total), which we refer to as Stage 3 replication (see Online Methods for details). The Stage 3 data also showed evidence in the same direction (one-sided P=2.25×10−4). Together, the combined discovery and three-stage replication data provide strong evidence for association (P=1.87×10−11) and suggest each copy of the A allele increases risk of large vessel stroke by approximately 1.4 fold (Table 2 and Figure 2). This SNP is within the final intron of the gene HDAC9. The risk allele (A) frequency was 9.29% and 8.78% in the UK and German discovery controls respectively.

Figure 2

Forest plot for the associations between rs11984041 and large vessel stroke in discovery and replication collections

The blue lines show the 95% confidence intervals of the log odds-ratio (OR) for each cohort, with the area of each square proportional to the inverse of the standard error. The diamonds indicate the 95% confidence interval for, from top to bottom, the discovery summary (combined British and German discovery collections), combined across collections within each of the three replication stages, the replication summary (combined across all three replication stages) and the overall summary (all discovery and replication collections combined). Evidence was combined across collections via an inverse-variance weighted fixed-effect meta-analysis. There was no evidence of heterogeneity of effect across collections (P = 0.92).

Standard statistical tests of association between rs11984041 and each of cardioembolic and small vessel stroke are not significant (discovery plus 2-stage replication p = 0.12, OR = 1.10, 95% CI = 0.98 – 1.23, and p = 0.06, OR = 1.13, 95% CI = 1.00 – 1.28 respectively). A non-significant result could simply be due to a lack of power: lack of significance in itself cannot rule out an effect in these subtypes. We investigated this potential genetic heterogeneity further by formally comparing different statistical models for the effect of the SNP on the different stroke subtypes. The models we compared were: (i) a model in which the variant has no effect on risk for any of the subtypes (“null” model); (ii) a model in which the SNP has the same effect on each subtype (“same effects” model); (iii) three models, in each of which the SNP has an effect on one subtype, and no effect for the other two subtypes (“LVD”, “SVD” and “CE” models respectively for the effect only in large vessel, small vessel, and cardioembolic stroke); and (iv) a “correlated effects” model allowing different, but correlated, effects for each subtype. We undertook the model comparison in a Bayesian statistical framework (see Online Methods for details), for our new association around HDAC9, as well as for the previously reported associations we confirmed as listed in Table 2. The results, based on the discovery and the first two stages of the replication, are shown in Figure 3.

Figure 3

Genetic heterogeneity of different stroke sub-types for the 4 loci with significant associations: HDAC9, PITX2, 9p21(CDKN2A/CDKN2B) and ZFHX3

Bar plots show the posterior probabilities on the models of association: no effect in any subtype (Null), same effect in all subtypes, correlated effects across subtypes or subtype specific effects (see main text). Models are a priori assumed to be equally likely. Bayes factors, which compare the evidence (marginal likelihood) between any pair of models, can be calculated as the ratio of the posterior probability assigned to each model as reported under each bar of the plot. Accompanying density plots show the marginal posterior distribution on the odds ratio of the risk allele for each stroke subtype assuming a model of correlated effects (see methods for specification of priors). These analyses were performed using both discovery and replication samples (stage 1 & 2).

For rs11984041 at HDAC9 there is very strong evidence against the null model and both the SVD and CE models (unsurprisingly given we ascertained this SNP on the basis of evidence for an effect in LVD) and also strong evidence against the model in which the SNP has the same effect in each subtype, thus demonstrating genetic heterogeneity across stroke subtypes at this SNP. The greatest posterior weight rests on the model in which there is only an effect for large vessel disease, with some weight on the correlated effects model, and in this model the posterior distributions on effect size for SVD and CE stroke are concentrated on much smaller effect sizes than for LVD.

In our data, heterogeneity is also seen at rs2383207 in the 9p21 region, a locus associated with heart disease and related phenotypes, and previously associated with large vessel stroke. Most support is for the model in which the effect sizes for the three stroke subtypes are correlated but there is also substantial weight on the model in which there is only an effect for large vessel stroke. The same analyses in our data for the top SNPs in the regions previously associated with cardioembolic stroke (PITX2 region, rs1906599, and ZFHX3 regions, rs12932445) show strong support for the model in which these SNPs only affect risk for cardioembolic stroke. Together these analyses provide compelling evidence for heterogeneity of genetic effects between stroke subtypes.

The association with rs11984041, in the gene HDAC9, implicates a novel region of the genome in an individual’s susceptibility to stroke. Any association with stroke could be mediated via associations with intermediate cardiovascular risk factors that themselves increase large vessel stroke risk. Our study design does not allow a direct assessment of this, as such risk factors were not available for control individuals. However, to date no associations have been reported between rs11984041 or correlated SNPs and hypertension (15), hyperlipidaemia (16), or diabetes (17) from large-scale GWAS of these risk factors.

Association of genetic variants surrounding HDAC9 are represented in Figure 4. All variants showing an association signal reside within a peak between two recombination hotspots and encompass the tail end of HDAC9. The downstream genes TWIST1 and FERD3L are physically relatively close to the identified peak and cannot be excluded as possible mechanisms via which genetic variants may exert cis-effects on the large vessel stroke phenotype. HDAC9 is a member of a large family of genes that encode proteins responsible for deacetylation of histones, and therefore regulation of chromatin structure and gene transcription (18). HDAC9 is ubiquitously expressed, with high levels of expression in cardiac tissue, muscle and brain (19). Although known as histone deacetylases, these proteins also act on other substrates (20) and lead to both upregulation and downregulation of genes (21).

Figure 4

Plot of association signals around rs11984041 for large vessel stroke in the combined British and German discovery samples. SNPs are coloured based on their correlation (r2) with the labelled hit SNP which has the smallest P-value in the region. r2 is calculated from the WTCCC2 control data. The bottom section of each plot shows the fine scale recombination rates estimated from HapMap data, and genes are marked by horizontal red lines. Arrows on the horizontal red lines denote direction of transcription, and black rectangles are exons.

The mechanism by which variants in the HDAC9 region increase large vessel stroke risk is not immediately clear. The specific association with this stroke subtype would be consistent with the association acting via accelerating atherosclerosis. The HDAC9 protein inhibits myogenesis and is involved in heart development (19) although deleterious effects on systemic arteries have not yet been reported. Alternatively it could increase risk by altering brain ischaemic responses and therefore have effects on neuronal survival. The protein has been shown to protect neurons from apoptosis, both by inhibiting JUN phosphorylation by MAPK10 and by repressing JUN transcription. HDAC inhibitors have been postulated as a treatment for stroke (22).

It is not uninformative that a large GWAS (~3,500 cases, ~6,000 controls) failed to find any novel associations for the combined phenotype of ischemic stroke. It may be that the genetic architecture of the disease involves fewer variants of more moderate effect than many other diseases, and/or that these happen not to be well tagged by the Illumina 660-W chip used in the study. On the other hand, as our data demonstrate, all the known loci exhibit genetic heterogeneity across the stroke subtypes, with at least some, and possibly all, affecting only a single subtype. This supports the possibility that distinct subtypes of the disease have differing genetic architectures. However this is based on only four loci and does not exclude the possibility that future loci associated with stroke may predispose to all ischaemic stroke. Clinical classification of disease into subtypes is not perfect. Since errors in classification would reduce power to detect heterogeneity, our findings of homogeneity within classes indirectly reinforces the value of current classification methods. Because GWAS studies to date, including the one reported here, have had relatively small sample sizes for each disease subtype (and hence are underpowered for common variants of small effect), it remains possible, and indeed a priori likely, that the range of effect sizes for each subtype will be similar to those for other common diseases. This suggests that future genetic studies should study adequate sample sizes for particular subtypes of ischaemic stroke, rather than for the disease as a whole.

In summary, in this largest GWAS study of ischemic stroke conducted to date, we identified a novel association with the HDAC9 gene region in large vessel stroke with an estimated effect size which is at the larger end for GWAS loci (OR 1.38, 95% CI 1.22-1.57 from replication data). We also replicated three other known loci, and showed genetic heterogeneity across subtypes of the disease for all four stroke loci. This genetic heterogeneity seems likely to reflect heterogeneity in the underlying pathogenic mechanisms, and reinforces the need for separate consideration of stroke subtypes in the research and clinical context.