Introduction

Migraine is a common neurological disorder that is characterized by severe attacks of headache accompanied by symptoms such as nausea, vomiting and photophobia and phonophobia.1 Two main types of migraine are distinguished based on the presence of an aura that can precede the headache: migraine with aura (MA) or without aura (MO). Although MA and MO have been considered distinct disease entities,2, 3 it is now more and more accepted that they represent different manifestations of the same disease.4, 5, 6

Genetic studies in familial hemiplegic migraine (FHM), a rare monogenic subtype of MA that is considered a suitable model for common migraine,7 revealed three genes (CACNA1A,8 ATP1A29 and SCN1A10) that are involved in ion and neurotransmitter transport in the brain. Despite considerable efforts, linkage and candidate-gene association studies in common migraine have had limited success, with only a few consistently replicated linkage findings.11, 12, 13, 14, 15, 16 A recent genome-wide association study (GWAS), using data from migraine patients who were recruited through headache clinics, found evidence for a role of the metadherin (MTDH) gene in common migraine.17 The associated SNP in that study affects MTDH gene expression and thereby indirectly regulates the expression of the glutamate transporter gene SLC1A2 (also known as EAAT2 or GLT-1), encoding a major glutamate transporter in the brain. This fits in well with the theory that increased glutamate release or reduced glutamate uptake increases the risk of migraine attacks.18, 19, 20, 21, 22

Here, we present a GWA meta-analysis for common migraine by the Dutch Icelandic migraine genetics consortium (DICE). This is the first population-based GWAS for common migraine, including 2446 migraine cases and 8534 controls from six Dutch and Icelandic samples. For replication, two population-based samples of Dutch and one of Australian origin were available. De novo genotyping was performed in the two Dutch replication cohorts (N=769 and 337 cases; 940 and 826 controls, respectively). In addition, an in silico replication study was performed in the Australian replication cohort (N=1851 cases, 4008 controls).

Methods

Populations: subjects, phenotypes and genotyping

The five Dutch samples that were used for the meta-analysis came from the Erasmus Rucphen Family (ERF) study,23, 24 The Netherlands Study of Depression and Anxiety (NESDA),25 The Netherlands Twin Registry (‘NTR1’ and ‘NTR2’)26 and the Rotterdam study,27 and included 330, 756, 378, 276 and 349 migraine cases, respectively. The Icelandic sample came from the AGES-Reykjavik Study and included 357 migraine cases.28 In addition to the 2446 migraine cases, 8534 non-migraine controls (2862 Icelandic and 5672 Dutch controls) from the respective cohorts were included (for details, see Table 1 ). All individuals came from population-based samples and were unrelated, with the exception of the ERF participants, who are part of a genetically isolated population in the Southwest of the Netherlands. Data on migraine symptomatology were collected by means of questionnaires (ie, AGES, NESDA, NTR1–2, Rotterdam), or a combination of questionnaires and telephone interview follow-up (ie, ERF).

Table 1 Sample descriptives

Three additional independent samples were available for replication; two Dutch population-based samples (the GEM sample29 and a third sample from the NTR), and one Australian sample, the Australian Twin Migraine (ATM) GWA study.30, 31 The Dutch GEM sample included 769 migraine cases and 940 non-migraine controls. The NTR replication sample consisted of 337 cases and 826 non-migraine controls, and the Australian sample consisted of 5859 unrelated individuals (1851 migraine cases, 1631 non-migraine controls and 2377 additional unselected controls).

Genotyping was performed using a variety of SNP genotyping platforms. To ensure sufficient overlap between studies?, genotypes for 2.5 million HapMap SNPs were imputed using MACH32 or IMPUTE33 software. An overview of the samples, including details on sample size, genotyping and imputations, is provided in Table 1. More details on the background of the studies, phenotyping strategies and genotyping procedures can be found in the Supplementary data.

GWA and meta-analysis

In each sample, a logistic regression association test was carried out. Next, a meta-analysis was performed combining the GWA results of the six samples (total number of individuals: 10 890) using the METAL program (http://www.sph.umich.edu/csg/abecasis/metal/). As different phenotype definitions were used in the different samples, the effect sizes may not be directly comparable between studies. Therefore, a pooled Z-score approach was used. With the pooled Z-score method, an overall Z-score is calculated based on the summed Z-scores from the individual studies, weighted by each study's sample size. The weights are calculated as the square root of (Nstudy/Ntotal), and the squared weights sum to one. The direction of effect is indicated by the sign of the Z-score. To ensure that meta-analysis results were based on SNP data of a large enough number of individuals, 184 350 SNPs that were available for <70% of all participants were excluded from the meta-analysis. This left a total of 2 394 913 autosomal SNPs for analysis. Annotation of meta-analysis results was performed with WGA viewer version 1.26E (Dongliang Ge and David B Goldstein; http://people.genome.duke.edu/~dg48/WGAViewer/).34 P-values<5 × 10−8 were considered genome-wide significant.35

Replication studies

A replication study was performed with direct genotyping in the GEM and the NTR replication samples. The top SNP was genotyped with a TaqMan assay in both GEM and the NTR replication sample. In addition, another 18 SNPs with a P-value<1 × 10−5 were selected based on informativeness given the LD structure. These SNPs were genotyped in the GEM sample using an in-house Sequenom iPLEX Mass-ARRAY platform (Sequenom Inc., San Diego, CA, USA). Logistic regression was performed to test for association between these SNPs and migraine status. Third, all DICE SNPs with a P-value<1 × 10−4 in the meta-analysis were selected, and for these SNPs, an in silico replication study was performed in the ATM GWA data set. Finally, the 19 SNPs that were genotyped in GEM and the NTR replication sample and measured or imputed in the ATM GWA replication sample were meta-analysed together with the discovery data sets. A more detailed description of the genotyping procedures and association analyses can be found in the Supplementary data.

Post hoc analyses

Text mining

Relationships between genes (emerging from the meta-analysis) and migraine were studied using the Anni text-mining program (Anni version 2.1; http://www.biosemantics.org/anni).36 For details see the Supplementary data.

Comparison of results with migraine genes and loci from previous studies

Genome-wide linkage studies and association studies for migraine were identified with a literature search in PubMed. We examined which SNPs with a P-value<1 × 10−4 coincided with a region containing a published migraine linkage peak. In addition, a selection of migraine candidate genes was made and inspected in our meta-analysis data set by calculating a gene-based P-value for each of the selected genes using the VEGAS program.37 More details can be found in the Supplementary methods.

Results

Meta-analysis

GWA analyses were performed in the six population-based samples and the results were meta-analyzed using a pooled Z-score approach. As shown in Figure 1, no systematic deviation from the expected distribution of P-values was observed in the Q–Q plot, which is reflected by a genomic inflation factor (λ) of 1.022. A total of 32 SNPs had a P-value<1 × 10−5 (Table 2 ). None of these SNPs exceeded the threshold for genome-wide significance (Figure 2). A total of 10 SNPs were located within genes; 8 in the metastasis associated in colon cancer 1 (MACC1) gene (7p21), 1 in the immunoglobulin lambda-like polypeptide 1 (IGLL1) gene (22q11) and 1 in the nerve growth factor receptor (NGFR) gene (17q21–q22). The most significant result was obtained for SNP rs9908234 (P=8.00 × 10−8) in the NGFR gene, with the strongest evidence coming from the AGES and NESDA studies (Supplementary Table S1). Data for 17 additional SNPs in this gene were available, but none of these were associated with migraine (all P-values>0.05). These SNPs were not in LD with rs9908234. Next, we performed text mining with the Anni program. The concept ‘migraine’ was matched against the genes located within or close to our top SNPs (P-value<1 × 10−4). Remarkably, the NGFR gene surfaced as the best migraine candidate gene.

Figure 1
figure 1

Q–Q plot showing the expected and observed distribution of P-values in the meta-analysis that included the five Dutch samples and the Icelandic sample. The genomic inflation factor (λ) for the meta-analysis was 1.022.

Table 2 Selected SNPs with P-values<1 × 10−5 in the meta-analysis
Figure 2
figure 2

Manhattan plot showing the P-values by chromosome for the meta-analysis.

Replication analysis of the top SNPs

From the 32 top SNPs with P-values<1 × 10−5, we selected 19 SNPs for genotyping in the GEM sample. The selection was made such that the genotyped SNPs were maximally informative given the LD between them. The top SNP rs9908234, located in the NGFR gene, was genotyped in one additional replication sample from the NTR. The association observed in the discovery samples could not be replicated for rs9908234 (GEM: OR=0.86, P=0.31; NTR replication sample: OR=0.89, P=0.579; see Supplementary Table S2). The findings for the other 18 SNPs were not replicated either. None showed a P-value<0.05 in the GEM sample: the smallest P-value observed was 0.10, but this effect was in the opposite direction compared with the meta-analysis.

An in silico replication study was performed in the ATM GWA sample. This analysis included all SNPs with P-values <1 × 10−4 in the DICE meta-analysis. In the ATM GWA sample, there were data for 327 out of 340 SNPs with P-values<1 × 10−4. None of these SNPs had a P-value<0.01 in the ATM GWA data set. A total of 11 SNPs had P-values between 0.01 and 0.05, but for only 3 SNPs, the effect was in the same direction as in the DICE cohorts (rs6919479, P=0.045; rs9363693, P=0.045; and rs9294736, P=0.037; all on chromosome 6). These results were not significant after correction for multiple testing.

Finally, a new meta-analysis was carried out in the DICE discovery samples, the GEM sample and the ATM GWA sample, for the 19 SNPs genotyped in GEM. For rs9908234, the meta-analysis also included the NTR replication cohort. The P-values for these SNPs did not decrease compared with the first meta-analysis (Supplementary Table S2).

Comparison of meta-analysis results with previous genetic findings in migraine

The large sample size of the present study provided a unique opportunity to further investigate previous findings from linkage and candidate gene studies on a larger scale, and to try and replicate the findings recently reported in a large clinic-based GWAS for migraine.17

First, we investigated whether there were any SNPs with P-values<10−4 that were located in previously identified migraine linkage regions (Supplementary Table S3). Five SNPs were located on chromosome 10q22–q23, a locus that has been reported for migraine several times.11, 15, 16 However, none were located in or near a gene that could easily be linked to migraine pathophysiology. Interestingly, one SNP (rs1972860, P=6.02 × 10−5) was located in the glutamate receptor, ionotropic, delta 2 (GRID2) gene on chromosome 4q22, a region reported in several different migraine linkage studies.11, 12, 13, 14

In addition, we performed a gene-based association test for selected candidate genes for migraine (Table 3 ). Seven candidate genes were selected based on the results of previous candidate gene association studies for common migraine. Furthermore, a recently published GWAS of clinic-based migraine identified an SNP (rs1835740) that was located between two interesting candidate genes: the MTHD gene and the plasma glutamate carboxypeptidase (PGCP) gene.17 An eQTL analysis revealed that rs1835740 most likely affects migraine through cis-regulation of MTHD, which in turn downregulates SLC1A2, a gene that encodes an important glutamate transporter in the brain. Therefore, we selected MTHD, PGCP and SLC1A2 as candidate genes, and also inspected SNP rs1835740 and two nearby correlated SNPs (rs982502 and rs2436046). Finally, the three FHM genes (CACNA1A, ATP1A2 and SCN1A) were included in the analysis.

Table 3 Results of the meta-analysis in previously identified candidate genes and genes previously implicated in familial hemiplegic migraine

Gene-based tests were performed for each of the selected candidate genes, using the meta-analysis results of all SNPs tested in the respective genes (Table 3). A gene-based test result was considered significant at an alpha level of 0.05, with Bonferroni correction for 13 tests, which corresponds to a gene-based P-value of 0.05/13=0.0038. None of the genes identified through candidate gene association studies was significantly associated with migraine in the meta-analysis. Although there were nominally significant SNPs in the LTA, ESR1 and INSR genes, results were not significant after correction for the number of SNPs tested within the respective genes. The PGCP and SLC1A2 genes also had several nominally significant SNPs, but again were not significant in the gene-based test. However, in the MTHD gene, 19 of the 28 tested SNPs had a P-value<0.01 in the meta-analysis (Supplementary Table S4). The gene-based P-value for MTDH was 0.002, which remained significant after Bonferroni correction. The SNP that showed association in the clinic-based GWA study (rs1835740)17 did not show significant association with migraine in the meta-analysis (P=0.64). Two nearby SNPs (rs982502 and rs2436046) reported in the same GWAS were also not associated with migraine in the present study.

Finally, we tested the three FHM genes, and found several nominally significant SNPs within CACNA1A and ATP1A2. The gene-based test for CACNA1A (best SNP rs3764615, P=0.004) was not significant (P=0.30). The gene-based P-value for ATP1A2 (best SNP rs2854248, P=3.62 × 10−4) was 0.006.

Discussion

This study describes the first meta-analysis of GWAS for population-based migraine, and contains a total of 2446 migraine cases and 8534 controls. The best P-value was obtained for SNP rs9908234, which is located in the NGFR gene. A replication study was performed in two Dutch replication cohorts that were available for wet replication; the GEM cohort (769 cases, 940 controls) and the NTR replication cohort (337 cases, 826 controls). In addition, the ATM GWA cohort (1851 cases, 1631 controls) was available for in silico replication. Although the NGFR gene is an interesting candidate gene for migraine, the association of NGFR with migraine could not be replicated in these cohorts. A total of 18 additional top SNPs (P-value<10−5) from the meta-analysis were tested in the GEM cohort and the ATM GWA cohort, but none could be replicated successfully.

There are several possible explanations for the lack of replication. First, several different genotyping platforms were used, which made imputation necessary to ensure sufficient overlap between the studies. Also, two different programs (MACH and IMPUTE) were used for imputation. However, given that MACH and IMPUTE use very similar imputation algorithms, and have been reported to be very similar in imputation accuracy,38 we do not expect this to have a major effect on our results. Second, there were some differences between the samples in the precision of the migraine diagnoses, and in most samples, a clinical migraine diagnosis was not available. This is often the case in population-based studies because, for reasons of efficiency, diagnoses are commonly made with (short) headache questionnaires. Unlike in clinic-based studies, they are not usually further evaluated with more extensive questionnaires or interviews by specialized physicians. Less accuracy of diagnosis may result in reduced power to detect association. The phenotypic differences also extend to the control groups, as all non-migraine individuals were included as controls. These differences between studies mean that effect sizes may not be directly comparable. To address this, a pooled Z-score meta-analysis was performed. This type of analysis does not require a direct comparison of effect sizes.39 Third, population-based cohorts also include many patients who have less severe migraine and a lower attack frequency. This means that they might be a genetically more heterogeneous group than patients from clinic-based cohorts. In addition, they are likely to have a lower genetic risk of migraine than the more severely affected patients in clinic-based cohorts. As a consequence, population-based studies may require a larger number of patients for sufficient power. Given that this study replicates previous findings, but does not produce genome-wide significant results, insufficient power (possibly because of the reasons above) seems the most likely explanation for the lack of replication of our top results. A lack of power makes it difficult to distinguish between true associations and false-positive findings in the original meta-analysis. Therefore, when the discovery samples have insufficient power, SNPs selected for replication based on small P-values may not replicate (even in sufficiently large replication samples) because they are false positives. Finally, it should be mentioned that the NESDA sample differed from the other samples because the majority of NESDA participants were selected for major depression. Because of the comorbidity of migraine and major depression, there is a higher prevalence of migraine in this sample than in the other samples. However, given that the percentage of MDD was similar in the migraine cases and the controls (94.3 vs 86.6%), any associations detected in this sample will be related to migraine and not to MDD.

In the present study, we also investigated SNP rs1835740 that was found to be significantly associated with MA in the first GWA study of clinic-based populations.17 This SNP is located on 8q21 between the MTDH and PGCP genes. The SNP itself was not associated with migraine in our study, but our gene-based analyses provided modest support for an association of MTDH with migraine.

In summary, although this study does not provide genome-wide significant association of an SNP with migraine, it provides suggestive evidence for an association with the MTDH gene, which is involved in the glutamate pathway, previously hypothesized to have a role in migraine based on findings in FHM.40 Clearly, even though a large number of patients and controls were included, the present study suffered from a lack of power. In addition to simply increasing the sample size, additional strategies aimed at minimizing phenotypic and genetic heterogeneity may be necessary. Strategies to achieve this can include the identification of reliable biomarkers or stratification of samples based on phenotypic similarity (eg, by looking at trait components,11 specific symptoms16, 41 and/or comorbid pathology). In addition, in future studies it may be worth focusing specifically on the glutamate pathway to assess whether genetic variants affecting glutamate levels are systematically associated with migraine.