Genome-wide association study reveals genetic risk underlying Parkinson's disease.

We performed a genome-wide association study (GWAS) in 1,713 individuals of European ancestry with Parkinson's disease (PD) and 3,978 controls. After replication in 3,361 cases and 4,573 controls, we observed two strong association signals, one in the gene encoding alpha-synuclein (SNCA; rs2736990, OR = 1.23, P = 2.24 x 10(-16)) and another at the MAPT locus (rs393152, OR = 0.77, P = 1.95 x 10(-16)). We exchanged data with colleagues performing a GWAS in Japanese PD cases. Association to PD at SNCA was replicated in the Japanese GWAS, confirming this as a major risk locus across populations. We replicated the effect of a new locus detected in the Japanese cohort (PARK16, rs823128, OR = 0.66, P = 7.29 x 10(-8)) and provide supporting evidence that common variation around LRRK2 modulates risk for PD (rs1491923, OR = 1.14, P = 1.55 x 10(-5)). These data demonstrate an unequivocal role for common genetic variants in the etiology of typical PD and suggest population-specific genetic heterogeneity in this disease.

Advances in genotyping technology have allowed rapid genome-wide screening of common variants in large populations launching a new era in the investigation of the genetic basis of complex diseases. So far GWAS have contributed little in the field of PD2–4; likely because previous studies lacked power to detect effects of the size expected in these diseases.

The present study was designed as a two-stage GWAS. Characteristics of the cohorts are shown in Table 1. For stage I, genotyping was performed using Infinium BeadChips (Illumina, Inc.). Following quality control 463,185 SNPs were analyzed in 1,713 PD cases and 3,978 controls. To assess the homogeneity of our cohort, pair-wise Identity by State distances (IBS) were calculated using HapMap data as a reference. These analyses revealed that our samples share common Caucasian ancestry (supplementary figures 1 and 2). We chose not to use genomic control as false positive association, possibly caused by population substructure (lambda=1.17), would be controlled for by our two-stage design. Power calculations showed our sample had 80% power to detect variants conferring an odds ratio (OR) of 1.3 with an allele frequency of 10% (supplementary figure 3).

Table 1

Study characteristics of cases and controls in stage I and II

		cases			controls
		Sample size	aao mean (s.d.)	male/female ratio	Sample size	aae mean (s.d.)	male/female ratio
Stage I	USA	988	55.9 (15.1)	1.09	3071	62 (15.6)	0.96
	Germany	757	56 (11.64)	1.49	976	NA	1.08
Stage II	USA	1528	62.5 (8.55)	2.44	2044	63 (15.6)	2.45
	Germany	1100	61 (11.32)	1.37	2168	57 (10.54)	1.4
	UK	824	59 (12.3)	3.5	544	NA	0.57

aao = age at onset; aae = age at examination; s.d. = standard deviation; NA = not available

Each SNP was tested for association using a Cochran-Armitage genotypic trend model. Four SNPs on chromosome 4q22 within the SNCA locus exceeded Bonferroni corrected genome-wide significance threshold in stage I (most significant p=5.69×10−9, rs2736990; Figure 1, Table 2). Three SNPs at the MAPT locus on chromosome 17q21 also surpassed genome-wide significance in stage I (most significant p=5.05×10−8, rs199533). A logistic regression analysis adjusted for the two first components of the multidimensional scaling values following calculation of pair-wise IBS distance suggests that the significant results obtained are not biased by population substructure (supplementary material).

Figure 1

Graphical representation of p values in stage I and stage II. A) In stage I, p values are log transformed (y-axis) and plotted against chromosomes (x-axis). The red line indicates the Bonferroni threshold. Signals indicated in red are on chromosome 4 and chromosome 17 and surpass Bonferroni threshold for genome wide significance. B) log transformed p-values of Stage II SNPs (y-axis) are plotted against chromosomes (x-axis). Signals indicated in red are on chromosome 4 and chromosome 17 and surpass Bonferroni threshold for multiple testing

Table 2

The top three SNPs of the 2 loci that surpass Bonferroni threshold for multiple testing in both stages and the top three SNPs in the LRRK2 locus and the additional loci in chromosomes 1 and 4.

locus				Stage I				Stage II				Stage I + II combined
SNP	Chr.	Position	Alleles (minor/major)	MAF_U	MAF_A	p-value	cOR	MAF_U	MAF_A	p value	cOR	MAF_U	MAF_A	p value	cOR
Genome-wide significant loci
rs393152	17	41074926	G/A	0.22	0.18	1.42E-07	0.76	0.22	0.18	1.69E-09	0.78	0.18	0.22	1.95E-16	0.77
rs199533	17	42184098	T/C	0.20	0.16	5.05E-08	0.75	0.20	0.17	7.95E-08	0.79	0.17	0.20	1.09E-14	0.78
rs17563986	17	41347100	C/T	0.22	0.18	3.44E-07	0.77	0.21	0.18	3.21E-08	0.79	0.18	0.22	1.67E-14	0.78
rs2736990	4	90897564	C/T	0.46	0.52	5.69E-09	1.27	0.46	0.51	5.52E-09	1.20	0.51	0.46	2.24E-16	1.23
rs3857059	4	90894261	G/A	0.07	0.10	3.60E-08	1.49	0.07	0.10	3.39E-08	1.43	0.10	0.07	3.74E-15	1.48
rs11931074	4	90858538	T/G	0.07	0.10	4.78E-08	1.49	0.08	0.10	3.75E-08	1.42	0.10	0.07	1.62E-14	1.46
Other loci
rs823128	1	203980001	C/T	0.04	0.03	1.90E-04	0.64	0.04	0.03	1.96E-04	0.68	0.03	0.04	7.29E-08	0.66
rs11240572	1	204074636	T/G	0.04	0.02	1.30E-04	0.63	0.04	0.03	1.81E-03	0.72	0.03	0.04	6.11E-07	0.67
rs823156	1	204031263	C/T	0.18	0.16	4.31E-03	0.85	0.17	0.16	0.06	0.93	0.16	0.18	7.60E-04	0.89
rs12646913	4	15348374	G/A	0.08	0.07	0.041	0.85	0.08	0.07	0.32	0.97	0.07	0.08	0.03	0.92
rs4698412	4	15346446	G/A	0.44	0.43	0.085	0.93	0.45	0.43	0.14	0.95	0.43	0.44	0.03	0.94
rs12502586	4	15335662	A/G	0.10	0.12	6.50E-03	1.20	0.10	0.10	0.89	1.00	0.11	0.10	0.07	1.08
rs11564162	12	38729159	C/T	0.21	0.18	4.00E-05	0.78	0.21	0.19	0.10	0.93	0.19	0.21	9.52E-05	0.87
rs2896905	12	38779683	T/C	0.40	0.35	5.03E-06	0.82	0.38	0.38	0.73	1.01	0.37	0.39	7.81E-03	0.93
rs1491923	12	38877384	C/T	0.31	0.34	2.20E-04	1.20	0.31	0.33	0.01	1.10	0.33	0.31	1.55E-05	1.14

Position from NCBI genome build 36; Chr. = Chromosome; MAF U = minor allele frequency in controls; MAF A = minor allele frequency in cases; p-value = nominal p-value of trend test; cOR = common OR.

Replication comprised genotyping of 384 SNPs selected based on the p-value observed in stage I (supplementary table 1). Genotyping was performed in an independent cohort of 3,452 cases and 4,756 controls from the US, Germany and Britain. Taking into account the results obtained from pairwise Identity by State distances calculations and considering that genetic heterogeneity and allelic heterogeneity are not likely to produce type I and type II errors when pooling white North American and white North European populations, we decided to analyze all Stage II samples together. Following quality control filtering, 345 SNPs were analyzed in 3,361 cases and 4,573 controls. Twenty-one SNPs within the SNCA and MAPT loci surpassed Bonferroni threshold for significance (p<0.000145; Table 2; supplementary table 2).

Three hundred and forty five stage II SNPs passed our quality control procedures (Supplementary table 1). Notably, we observed clusters of SNPs showing improved association signals when combining our stage I and stage II datasets (supplementary figure 4). Although some of these SNPs are at loci that contain biologically plausible candidate genes for PD, they do not reach genome-wide significance and thus we have resisted drawing conclusions from these data; however, of particular note is a cluster of 7 SNPs in chromosome 10q24.32, with p-values below 1×10−3 (Supplementary figure 5). These and other variants showing a consistent, moderate association across stages warrant independent replication.

To further delineate the signals on SNCA and MAPT, allelic association of significant SNPs was tested, conditioned on alleles of other significant SNPs at the same locus. No independent signals were identified, suggesting that a single detectible pathogenic variant exists at each locus (supplementary table 3); however without complete sequence data across these loci we cannot rule out independent effects. We did not find evidence for epistasis between SNCA and MAPT risk alleles (supplementary material).

Analysis of the linkage disequilibrium (LD) structure across the SNCA locus revealed two blocks of LD (Figure 2A). The 3′ block contains three of the four significantly associated SNPs, suggesting that the causal variant is located in the 3′ region of the SNCA gene. This is strengthened by analysis of the haplotype frequencies at this locus (supplementary figure 6) and previous studies5. The REP1 microsatellite in SNCA was previously associated with PD5 and its pathological effect has been suggested to be mediated by gene expression6. Analysis of REP1 genotype data in 1,774 samples from the US cohort revealed the risk allele of REP1 is in LD with the 3′ risk alleles identified here (r2=0.365 with rs3857059; supplemental figure 7A), thus the association identified at the REP1 locus and the SNPs identified here may be the result of residual LD between these loci. This is supported by logistic regression analysis conditioned on REP1 genotypes, showing that association at REP1 is not independent from the association identified here (supplementary material). We recently reported a significant association of SNCA SNPs with another synucleinopathy, multiple system atrophy (p=5.5×10−12, MSA)7; comparison of these data reveals disparate SNCA risk SNPs in MSA and PD, a finding that may shed light on the exact pathogenic substrate and molecular etiology of these disorders (supplementary table 4).

Figure 2

Association and recombination rates across SNCA, MAPT, PARK16 and LRRK2. −log10 p values are shown for stage I and II analyses, annotated transcripts are shown across the top of each plot. Red dotted line indicates threshold for genome wide significance in stage I and orange line indicates threshold for significance of stage II.

As expected, one large highly inter-correlated block of high LD was observed across the MAPT locus (Figure 2B). Available genotype data of the H1/H2 haplotypes in this region showed that the risk alleles of the associated SNPs are in LD with the H1 haplotype (r2=0.761 with rs393152; supplementary figure 7B). It is unclear from the current data whether the MAPT risk haplotype identified here corresponds to the subhaplotype associated with corticobasal degeneration (CBD) and progressive supranuclear palsy (PSP)8. Because of the LD structure we cannot rule out other genes at this locus as the pathogenically relevant genes; however, MAPT is biologically the most plausible candidate.

Following data exchange with colleagues performing a PD GWAS in Japan we chose to study two loci implicated in Asian PD on chromosomes 1q32 and 4p15. In our stage I data, the most significant p-values at these 12 SNPs were 1.3×10−4 and 6.5×10−3 (1q32 and 4p15.3 respectively). The signal at 1q32 was significant enough to carry through to stage II replication, but this SNP had been excluded because of the low minor allele frequency in controls (0.03). Genotyping of these 12 SNPs was performed in an available subset of our replication cohort comprising 2,816 cases and 3,401 controls. The signal on chromosome 1q32 was replicated in this cohort (rs823128, p=1.9×10−4; Table 2). While this failed to surpass Bonferroni correction, the p-value across stages was highly significant (rs823128, p=7.3×10−8) and it is worth noting that the significance improved for all SNPs at this locus when combining stage I and II results (Figure 2C). For these reasons and because the association at this locus was consistently detected in the Asian cohort1 we are confident this signal represents a true association and this has been designated PARK16. Although we failed to replicate the signal on chromosome 4p15, which included only one gene, BST1, the low minor allele frequency of the associated SNPs in Caucasian individuals may have affected our ability to observe association.

Mutations in SNCA and MAPT have been associated with autosomal dominant forms of parkinsonism9,10. Given this, it is interesting that we observed association proximal to LRRK2, which also contains mutations causing autosomal dominant PD11,12. In stage I 23 SNPs upstream of LRRK2, and 12 SNPs within LRRK2 were associated with PD (lowest p=5.03×10−6 in rs2896905, located in SLC2A13, 0.27Mb from LRRK2). Of these, 3 SNPs surpassed our threshold for replication and were analyzed in stage II. Only one SNP, 0.17Mb upstream of LRRK2, remained associated with PD after stage II (rs1491923, p=0.01; supplementary material; Figure 2D). While this did not surpass our threshold for multiple testing, the combined stage I and II p-values revealed a compelling association (p=1.6×10−5). Interestingly, the other 2 replicated SNPs at this locus were also nominally associated with PD after combining stage I and stage II datasets (p-values of 9.5×10−5 and 7.8×10−3; rs11564612 and rs2896905 respectively). Data from the Asian cohort also revealed a significant association with PD at this locus1. SLC2A13, neighboring LRRK2, cannot be ruled out as the gene of effect at this locus, however, LRRK2 is clearly a more plausible candidate.

Although mutations and copy number variants of SNCA are the cause of rare familial forms of PD10,13, association of common variants has been more controversial. This study provides unequivocal evidence that variation in SNCA contributes to the etiology of sporadic PD. The clustering of associated SNPs in the 3′ UTR suggests that the causal variant might affect post-transcriptional RNA processing or RNA stability, possibly mediated by miRNA binding sites or by alternative splicing.

Association of the H1 haplotype at the MAPT locus with PSP and CBD has been described and replicated8; however, association studies of variants at MAPT in PD have produced conflicting results14,15. Our data provide unequivocal evidence for an association of the MAPT locus with PD. This is surprising given the classic separation of synucleinopathies and tauopathies, although a cross-talk between molecular pathways characterized by different aggregating proteins has been repeatedly suggested. While there are additional genes at the MAPT locus, the role of MAPT in neurodegenerative diseases is well established and this association is biologically plausible.

We provide evidence for an association of PD with variability proximal to LRRK2 and at a novel locus at 1q32 (PARK16). The kinase activity of lrrk2 is an attractive therapeutic target; our data suggests that this protein is also relevant to the etiology in sporadic PD patients without mutations. The PARK16 locus spans 5 transcripts, SNORA72, NUCKS1, RAB7L1, SLC41A1 and PM20D1. It will be important to define the immediate biological consequences of all four risk loci identified here. It is notable that three of the most significant loci identified here contain genes known to be mutated in Mendelian forms of parkinsonism. This supports the notion that rare familial disease is etiologically related to typical sporadic PD, and suggests that genes that contain common risk variants are excellent candidates to contain rare disease causing mutations. One might also predict that deep sequencing of these loci will reveal rare mutations that alter risk for, rather than cause, disease. It is also interesting that two of the four loci discussed here, are risk factors for other neurodegenerative diseases, including MSA (SNCA), PSP (MAPT) and CBD (MAPT).

The combined population attributable risk associated with the identified loci, considering the genotypic counts of those most associated SNPs in our cohort, is approximately 25% (supplemental methods). Our study was a retrospective case-control study and the frequency of the risk variants detected might not reflect the frequencies of true causal variants; thus these values should be interpreted with caution.

In an attempt to define a biological consequence of risk variants, we mined data produced within our laboratory. In this work genome wide genotyping and expression profiling of >22,000 transcripts was performed in 133 human frontal cortex samples, allowing us to determine SNPs significantly associated with expression level. These data revealed a strong association between genotype at the risk alleles of the MAPT locus and expression levels of both MAPT and LRRC37A (figure 3), in addition to less significant association with expression levels of other transcripts at this locus, but did not reveal association between risk SNPs and expression levels of proximal genes at the SNCA, LRRK2 or PARK16 loci (Supplementary table 5, figure 8). Notably the alleles at the MAPT locus associated with increase risk of PD are associated with increased expression of MAPT in the human brain.

Figure 3

Expression quantitative trait loci across the MAPT locus measured in 133 human frontal cortex samples; panel A shows association between genotypes and transcript levels across the MAPT locus. In this analysis the allelic load at genotyped polymorphisms across the locus is tested for association with transcript levels of each gene across the locus. The results of the analysis are shown as log transformed p values color-coded to match the transcript of interest. Notably genotypes across this locus are associated with MAPT (pink) and LRRC37A (blue) levels. Boxplots in panels B and C illustrate a dose relationship between allele load at the most significantly associated PD SNP (rs393152) and expression of MAPT (B) and LRRC37A (C), p values 4.1×10−6 and 1.7×10−13 respectively. Notably, the allele associated with higher risk for PD, A, is associated with high levels of MAPT expression and low levels of LRRC37A expression.

We show for the first time a clear role for common genetic variability in the risk of developing PD. Further we describe a possible population specific genetic heterogeneity in this disorder, since the association to MAPT was absent in the Asian cohort. This observation has potential implications for the analysis of complex traits across populations; such genetic heterogeneity, particularly at minor risk loci, has the potential to mask true associations when analyses are performed across populations. With the discovery of the PARK16 locus in the Asian population, this highlights the power of comparing GWAS across different populations. A further increase in the number and size of cohorts for GWAS in PD will likely reveal additional common genetic risk loci and these, in turn, will improve understanding and ultimately treatment of this devastating disorder.

METHODS

Methods and associated references appear online. Note: Supplementary information is available on the Nature Genetics website.