Genome-wide association study identifies five susceptibility loci for glioma

Abstract

To identify risk variants for glioma, we conducted a meta-analysis of two genome-wide association studies by genotyping 550K tagging SNPs in a total of 1,878 cases and 3,670 controls, with validation in three additional independent series totaling 2,545 cases and 2,953 controls. We identified five risk loci for glioma at 5p15.33 (rs2736100, TERT; P = 1.50 × 10⁻¹⁷), 8q24.21 (rs4295627, CCDC26; P = 2.34 × 10⁻¹⁸), 9p21.3 (rs4977756, CDKN2A-CDKN2B; P = 7.24 × 10⁻¹⁵), 20q13.33 (rs6010620, RTEL1; P = 2.52 × 10⁻¹²) and 11q23.3 (rs498872, PHLDB1; P = 1.07 × 10⁻⁸). These data show that common low-penetrance susceptibility alleles contribute to the risk of developing glioma and provide insight into disease causation of this primary brain tumor.

Gliomas account for ∼80% of malignant primary brain tumors (PBT). In the United States, ∼21,000 individuals are diagnosed with glioma annually¹ and for most the prognosis is dismal¹. No lifestyle exposure has consistently been linked to glioma risk except ionizing radiation, which accounts for few cases¹. Predicated on the hypothesis that part of the twofold increased glioma risk in relatives of individuals with PBT² is a consequence of the co-inheritance of multiple low-risk variants, we conducted two genome-wide association (GWA) studies. Pooling data from both scans and following up the best-supported associations, we have identified five glioma risk loci.

The two GWA studies were conducted in the UK and US (UK-GWA and US-GWA studies, respectively; Fig. 1). In both, genotyping of cases was conducted using Illumina Infinium HD Human610-Quad BeadChips. The UK-GWA study was based on genotyping 636 glioma cases ascertained through INTERPHONE³. After application of quality control criteria, genotypes were available for 631 cases (Fig. 1). For controls, we used Illumina Hap550K-BeadChip genotypes on 1,438 individuals from the British 1958 Birth Cohort⁴. The US-GWA study was based on genotyping 1,281 glioma cases ascertained by The University of Texas M.D. Anderson Cancer Center. After applying quality control criteria, genotypes were available for 1,247 cases (Fig. 1). For controls, we used Illumina Hap550K-BeadChip genotypes on 2,243 individuals from the Cancer Genetic Markers of Susceptibility studies^5,6.

Figure 1.

Subjects and single-SNP exclusion schema for genome-wide association studies.

Across both case series, 572,571 SNPs were satisfactorily genotyped (99.4%), with mean individual sample call rates of 99.8%. We excluded 32 individuals because of non-Western European ancestry and two because of cryptic relatedness (Fig. 1 and Supplementary Fig. 1a). For the 1,878 cases and 3,670 controls in the combined data, 521,318 SNP genotypes were available. We considered only the 454,576 autosomal SNPs having call rates >95% in all series, showing mimimal departure from Hardy-Weinberg equilibrium (HWE; P > 10⁻⁵ in controls) and minor allele frequencies (MAF) > 1% in cases and controls (Fig. 1).

Comparison of the observed and expected distributions showed little evidence for an inflation of the test statistics in the datasets (inflation factor⁷ λ = 1.02 and 1.05 in UK-GWA and US-GWA, respectively, based on the 90% least significant SNPs; Supplementary Fig. 1b), thereby excluding significant hidden population substructure or differential genotype calling. Using standard methods and data from both studies, we derived joint odds ratios (ORs) and confidence intervals for each SNP and associated P values.

To identify true risk alleles among the 34 SNPs showing evidence of an association at P < 10⁻⁵, we conducted replication in three case-control series involving a total of 5,498 individuals that passed quality control (French series: 1,392 cases, 1,602 controls; German series: 504 cases, 573 controls; Swedish series: 649 cases, 778 controls). We satisfactorily genotyped 31 of the 34 SNPs in all series. In the French series, rs9656979 and rs7257116 were not genotyped. However, an association between these SNPs and glioma risk was not supported in either the German or Swedish series (Supplementary Table 1). Although rs7300686 was not typed in the German or Swedish series, an association with risk was not supported in the French series (Supplementary Table 1). Hence, failure to genotype all additional series for the three SNPs has not affected study findings. Across the five data series, there was a significant association between rs7124728 and glioma risk, but this was not supported in the three replication series (P > 0.1). rs4944840 maps 57 kb from rs7124728 and both SNPs are in strong linkage disequilibrium (LD; D′ = 0.89, r² = 0.71). No association between rs4944840 and glioma risk was seen in either GWA study (P = 0.13 UK-GWA, P = 0.03 US-GWA). There was a significant difference in MAF of rs7124728 between both GWA cases and replication cases but not between controls, suggesting a genotyping error with the Illumina610-Quad BeadChip.

Excluding rs7124728, in the combined analyses, 14 of the 31 SNPs representing five genomic regions satisfied an accepted threshold for genome-wide statistical significance (that is, P < 5 × 10⁻⁷; Table 1 and Supplementary Table 1). Under a fixed-effects model, the strongest signal was attained at rs4295627 (OR = 1.36, 95% CI: 1.29–1.43; P = 2.34 × 10⁻¹⁸), which localizes to 8q24.21 (130,754,639 bp; Fig. 2 and Table 1). There was, however, evidence of between-study heterogeneity (P_het = 0.01) ascribable to the association being modest in the Swedish series. Under a random-effects model, the OR was 1.36 (95% CI: 1.19–1.58; P = 2.01 × 10⁻⁵) using all data and 1.45 (95% CI: 1.32–1.57; P = 1.02 × 10⁻⁸) excluding the Swedish study. rs4295627 maps to intron 3 of CCDC26, encoding a retinoic acid modulator of differentiation and death⁸. Retinoic acid induces caspase-8 transcription through phosphorylation of CREB and increases apoptosis induced by death stimuli in neuroblastoma cells⁹ and in glioblastoma cells with downregulation of telomerase activity¹⁰. Some evidence suggests a second 8q24.21 locus defined by rs891835, mapping to intron 1 of CCDC26 207 kb telomeric to rs4295627 (130,560,934 bp; OR = 1.24, 95% CI: 1.17–1.30; P = 7.54 × 10⁻¹¹) and showing low LD with rs4295627 (D′ = 0.59 and r² = 0.19 and D′ = 0.58, r² = 0.26 in HapMap and GWA scan, respectively). The OR for rs4295627 with adjustment for rs891835 was 1.30 (95% CI: 1.20–1.41; P = 2.80 × 10⁻¹⁰). Similarly, adjusting rs891835 for rs4295627 provided evidence of an association, albeit at a nominal significance level (OR = 1.08, 95% CI: 1.00–1.17; P = 0.03). There was also an increasing trend in OR with an increasing number of risk alleles (P = 2.67 × 10⁻¹³) and the frequency of the two common haplotypes differed between cases and controls (Supplementary Table 2). When we imputed 8q24.21 genotypes using genome-wide data, rs1077236, mapping to 130,709,683 bp, provided a marginally better association signal than rs4295627 (P = 2.75 × 10⁻⁹ and 1.60 × 10⁻⁸, respectively). Hence, the possibility remains that rs4295627 and rs891835, although only weakly associated, are in LD with one or more causal variants in this region.

Table 1. Summary results for SNPs associated with glioma in GWA and replication samples.

						GWA studies		Replication studies		Combined
SNP	Chr.	Genea	Location (bp)	Ancestral allele frequency	Risk alleleb	OR (95% CI)	P	OR (95% CI)	P	OR (95% CI)	P	P het
rs2736100	5	TERT	1,339,516	0.51	G	1.20 (1.10–1.33)	2.21 × 10−6	1.33 (1.20–1.49)	2.87 × 10−13	1.27 (1.19–1.37)	1.50 × 10−17	0.18
rs2853676	5	TERT	1,341,547	0.27	A	1.22 (1.14–1.31)	5.30 × 10−6	1.30 (1.21–1.38)	1.06 × 10−9	1.26 (1.20–1.32)	4.21 × 10−14	0.67
rs10464870	8	CCDC26	130,547,005	0.21	C	1.24 (1.15–1.34)	3.90 × 10−6	1.22 (1.13–1.31)	1.77 × 10−5	1.23 (1.17–1.30)	3.04 × 10−10	0.05
rs891835	8	CCDC26	130,560,934	0.22	G	1.24 (1.15–1.33)	3.92 × 10−6	1.24 (1.15–1.33)	4.43 × 10−6	1.24 (1.17–1.30)	7.54 × 10−11	0.01
rs6470745	8	CCDC26	130,711,103	0.20	G	1.30 (1.20–1.39)	5.79 × 10−8	1.31 (1.22–1.41)	9.09 × 10−9	1.30 (1.24–1.37)	2.77 × 10−15	0.01
rs16904140	8	CCDC26	130,734,825	0.21	A	1.25 (1.16–1.35)	1.41 × 10−6	1.28 (1.19–1.37)	1.14 × 10−7	1.27 (1.20–1.33)	7.88 × 10−13	0.01
rs4295627	8	CCDC26	130,754,639	0.17	G	1.33 (1.23–1.42)	1.47 × 10−8	1.39 (1.30–1.49)	2.20 × 10−11	1.36 (1.29–1.43)	2.34 × 10−18	0.01
rs1063192	9	CDKN2A/B	21,993,367	0.44	C	1.21 (1.13–1.29)	1.44 × 10−6	1.21 (1.14–1.29)	6.97 × 10−7	1.21 (1.16–1.27)	4.61 × 10−12	0.81
rs2157719	9	CDKN2A/B	22,023,366	0.57	G	1.22 (1.11–1.35)	6.80 × 10−7	1.22 (1.11–1.33)	4.42 × 10−7	1.22 (1.14–1.30)	1.41 × 10−12	0.68
rs1412829	9	CDKN2A/B	22,033,926	0.42	C	1.22 (1.14–1.30)	7.23 × 10−7	1.23 (1.15–1.30)	1.80 × 10−7	1.22 (1.17–1.28)	6.23 × 10−13	0.67
rs4977756	9	CDKN2A/B	22,058,652	0.40	G	1.25 (1.17–1.32)	2.39 × 10−8	1.24 (1.16–1.31)	5.90 × 10−8	1.24 (1.19–1.30)	7.24 × 10−15	0.94
rs498872	11	PHLDB1	117,982,577	0.31	T	1.26 (1.17–1.34)	1.03 × 10−7	1.12 (1.04–1.20)	4.56 × 10−3	1.18 (1.13–1.24)	1.07 × 10−8	0.04
rs6010620	20	RTEL1	61,780,283	0.23	G	1.28 (1.18–1.38)	8.38 × 10−7	1.28 (1.18–1.38)	6.49 × 10−7	1.28 (1.21–1.35)	2.52 × 10−12	0.38
rs2297440	20	RTEL1	61,782,743	0.22	C	1.28 (1.18–1.38)	1.01 × 10−6	1.26 (1.16–1.35)	4.44 × 10−6	1.27 (1.20–1.34)	2.06 × 10−11	0.40

^aGene(s) mapping within 20 kb of each SNP are listed.

^bAncestral allele marked in bold.

Figure 2.

LD structure and association results for the five confirmed glioma-associated regions. (a) 8q24.21. (b) 5p15.33. (c) 9p21.3. (d) 20q13.33. (e) 11q23.3. Chromosomal positions and genes based on NCBI build 36 coordinates. Armitage trend test P values (as −log₁₀ values; left y axis) are shown for SNPs analyzed in the GWA studies in blue and for the combined analysis including the replication series in red. Recombination rates in HapMap CEU across the region are shown in black (right y axis). Also shown are the relative positions of genes mapping to each region of association. Exons of genes have been redrawn to show the relative positions in the gene, and therefore maps are not to physical scale.

Variation at 8q24.21 has been implicated in colorectal¹¹, breast¹², prostate⁶ and bladder cancer risk¹³. The genomic regions underlying these associations map to the 128- to 129-Mb interval, distinct from the glioma signal. Closer to the 8q24.21 glioma signal is rs987525, a risk factor for cleft lip¹⁴. Notably, nonsyndromic cleft lip has been reported to be a risk factor for PBT¹⁵. Although rs987525 maps 740 kb centromeric to rs4295627 and LD between rs987525 and rs4295627 is weak (D′ = 0.46, r² = 0.01), it is possible that these SNPs define a common disease locus.

The second strongest evidence for an association was for rs2736100 (OR = 1.27, 95% CI: 1.19–1.37; P = 1.50 × 10⁻¹⁷) mapping to 5p15.33, which localizes to intron 2 of TERT (1,339,516 bp; Fig. 2 and Table 1). TERT is the reverse transcriptase component of telomerase, essential for telomerase activity in maintaining telomeres and cell immortalization. Although 5p15.33 copy number change is uncommon in glioblastoma multiforme (GBM; Supplementary Fig. 2a), TERT expression correlates with glioma grade and prognosis¹⁶. Several lines of evidence raise the possibility of a second locus defined by rs2853676, which also maps to intron 2 of TERT (1,341,547 bp; OR = 1.26, 95% CI: 1.20–1.32; P = 4.21 × 10⁻¹⁴) and is in weak LD with rs2736100 (D′ = 0.67 and r² = 0.15 and D′ = 0.81 and r² = 0.24 in HapMap and GWA scan, respectively). The OR for rs2736100 with adjustment for rs2853676 was 1.19 (95% CI: 1.12–1.27; P = 5.33 × 10⁻⁸) and adjusting rs2853676 for rs2736100 also provided evidence of an association, with little change in the risk estimate (OR = 1.15, 95% CI: 1.08–1.23; P = 6.90 × 10⁻⁵). There was an increasing trend in OR with an increasing number of risk alleles (P = 1.19 × 10⁻¹⁹). Comparison of haplotype frequencies provided evidence of two haplotypes differing between cases and controls (Supplementary Table 2). Recent data has implicated variation at 5p15.33 (TERT-CLPTM1L) defined by rs2736098 with risk of multiple tumor types¹⁷. As LD between rs2736098 and both rs2736100 and rs2853676 is poor (D′ = 0.48 and r² = 0.11 and D′ = 0.28 and r² = 0.02 in HapMap and GWA scan, respectively), whether the 5p15.33 association with glioma is the same remains to be established.

The third strongest evidence for an association was for rs4977756 (OR = 1.24, 95% CI: 1.19–1.30; P = 7.24 × 10⁻¹⁵), mapping 59 kb telomeric to CDKN2B (22,058,652 bp; Fig. 2 and Table 1) within a 122-kb region of LD at 9p21.3. This region encompasses the CDKN2A-CDKN2B tumor suppressor genes. CDKN2A encodes p16(INK4A) a negative regulator of cyclin-dependent kinases and p14(ARF1), an activator of p53. CDKN2A-CDKN2B has an established role in glioma, with homozygous deletion in CDKN2A detectable in ∼50% of tumors^18,19 and loss of expression linked to poor prognosis¹⁶. Furthermore, germline mutation of CDKN2A-CDKN2B causes the melanoma-astrocytoma syndrome^20,21. Regulation of p16/p14ARF is important for sensitivity to ionizing radiation, the only environmental factor strongly linked to gliomagenesis²². These data provide evidence that common variation in addition to rare mutations in CDKN2A-CDKN2B contributes to glioma predisposition.

A 68-kb core region of 9p21.3 incorporating the association signal influences risk of type 2 diabetes (T2D)²³ and coronary heart disease (CHD)²⁴. rs564398, a risk factor for T2D and CHD, is in LD with rs4977756 (D′ = 0.77, r² = 0.50). A possible explanation for a common etiology is that these variants affect determination of birth weight, linked as a risk factor to all three diseases^25–27.

The fourth strongest evidence for an association was for rs6010620 (OR = 1.28, 95% CI: 1.21–1.35; P = 2.52 × 10⁻¹²), localizing to intron 12 (61,780,283 bp) of the RAD3-like helicase gene RTEL1 (ref 28) and mapping within a 65-kb region of LD at 20q13.33 (Fig. 2 and Table 1). Amplification of 20q13.33 is seen in ∼30% of gliomas with copy number change correlating with RTEL1 expression (Supplementary Fig. 2a). RTEL1 maintains genomic stability directly by suppressing homologous recombination²⁸. These data, coupled with the role of TERT in glioma, make RTEL1 an attractive candidate for the 20q13.33 association but do not preclude the association being through variation in other genes mapping to the region, such as TNFRSF6B, whose expression correlates with glioma grade²⁹.

The fifth-strongest signal was at rs498872 (OR = 1.18, 95% CI: 1.13–1.24; P = 1.07 × 10⁻⁸), which maps to the 5′-UTR of PHLDB1 within a 101-kb LD block on 11q23.3 (117982577 bp; Fig. 2 and Table 1). There was, however, evidence of between-study heterogeneity (P_het = 0.04). Under a random-effects model, the combined OR was 1.16 (95% CI: 1.05–1.28; P = 2.24 × 10⁻³). Most of this heterogeneity could be ascribed to the German series; excluding this, the OR was 1.22 (95% CI: 1.16–1.28; P = 2.56 × 10⁻¹⁰, P_het = 0.40). Although there is no direct evidence for a role of PHLDB1, 11q23.3 is commonly deleted in neuroblastoma³⁰.

We analyzed the association between SNP genotypes and risk of tumors of glial origin, GBM and astrocytic lineage. There was evidence that GBMs have a different risk profile, notably with respect to rs4295627, rs498872 and rs2736100 (Supplementary Table 3). However, because of study heterogeneity, we cannot exclude the possibility that the variants do not influence the risk of developing a specific histological form of glioma without central histopathological review.

We only found evidence of interactive effects between 8q24.21 SNPs (Supplementary Table 4). For all other pairwise combinations, each locus had an independent role in glioma (P_interaction > 0.10). Glioma risk increases with increasing numbers of variant alleles for the five loci (OR_per-allele = 1.31, 95% CI: 1.26–1.36; P = 1.39 × 10⁻⁷⁴; Fig. 3 and Supplementary Table 5). Individuals with eight or more risk alleles have an approximately fourfold increase in glioma risk compared to those with a median number of risk alleles. These ORs may be underestimates because the additive model assumes equal weighting across SNPs. We estimate that the five loci we have identified account for ∼7–14% of the excess familial risk of glioma (assuming naïve multiplicative or additive models and an overall familial relative risk of 2.1). However, we acknowledge that the present data provide only crude estimates of the effect on susceptibility attributable to variation at the loci, as the effect of the actual common causal variant responsible for the association will typically be larger.

Figure 3.

Cumulative effects of glioma risk alleles. (a) Distribution of risk alleles in glioma cases (green bars) and controls (blue bars). (b) Plot of the increasing ORs for glioma with increasing number of risk alleles. The ORs are relative to the median number of four risk alleles; vertical bars correspond to 95% confidence intervals. Horizontal line marks the null value (OR = 1). The distribution of risk alleles follows a normal distribution in both case and controls, with a shift toward a higher number of risk alleles in cases.

To investigate the basis of causality for each of the five associations, we interrogated HapMap to identify nonsynonymous SNPs (nsSNPs) correlated with the most strongly associated SNP signals. The strongest LD was between rs6010620 and nsSNP rs3848668 (D′ = 1.0, r² = 0.05). Although HapMap is not comprehensive, these data suggest the associations identified are probably mediated through LD with sequence changes that influence expression rather than protein sequence or through LD with low-frequency variants not catalogued. We searched for cis-acting regulatory effects of SNPs on nearby genes in lymphoblastoid cell lines and normal brain tissue using publicly available data. Although we did not identify any significant genotype–expression relationships (Supplementary Fig. 2b), this does not exclude subtle effects or the possibility that the variants exert effects through untested transactivation of genes.

Our findings show that common variants influence glioma risk and highlight the importance of variation in genes encoding components of the CDKN2A-CDK4 signaling pathway in glioma. Moreover, this pathway, elucidated through the extended interaction network of CDKN2A, incorporates TERT (through mutual interaction with HSP90) and other genes (including CCDC26) we have identified as risk factors.

Online Methods

Subjects

UK-GWA study

The UK study was based on 636 cases (401 male, 235 female; mean age 46 years; s.d. = 12) ascertained through the INTERPHONE Study³. Briefly, the INTERPHONE Study was an international multicenter case-control study of PBT coordinated by the International Agency for Research on Cancer (IARC), with material collected between September 2000 and February 2004. UK individuals with PBT were collected through neurosurgery, neuropathology, oncology and neurology centers in the Thames regions of Southeast England and the Northern UK including central Scotland, the West Midlands, West Yorkshire and the Trent area. Cases were individuals with glioma (International Classification of Diseases for oncology (ICD-O), 2nd ed., codes 9380-9384, 9390-9411, 9420-9451 and 9505; ICD10 code C71). Cases with previous brain tumors were excluded. To minimize population stratification, cases with self-reported non-western-European ancestry were excluded from the present study. Individuals from the 1958 Birth Cohort served as source of controls⁴.

US-GWA study

The US study was based on 1,247 cases (768 male, 479 female; mean age 47 years; s.d. = 13) ascertained through The University of Texas M.D. Anderson Cancer Center, Texas, between 1990 and 2008. Cases were individuals with glioma (ICD10 code C71; ICD-O codes 9380-9384, 9390-9411, 9420-9451, and 9505). Individuals from CGEMS^5,6 served as controls.

Replication series

The French Glioma Collection (FGC) is systematic series of 1,439 individuals with histologically proven glioma (WHO classification AI, AII, AIII, OII, OIII, OAII, OAIII, GBM-IV) ascertained through the Service de Neurologie Mazarin, Groupe Hospitalier Pitié-Salpêtrière Paris. The controls are taken from the SU.VI.MAX (SUpplementation en VItamines et Minraux-AntioXydants) study of 12,735 healthy subjects (aged 35 to 60 years) recruited from across France in 1994 (ref. 31). The Swedish Glioma Collection (SGCCS) comprises 215 glioma cases ascertained as part of the INTERPHONE Study³ conducted between 2000 and 2002 in Sweden, 134 cases from Umeå University Hospital, 122 from the Northern Sweden Health and Disease Study (NSHDS)³² and 203 from neurosurgery university clinics in Sweden. 383 controls for the INTERPHONE study were randomly selected from the population register, frequency-matched to brain tumor cases on age, sex and geographical region. In addition, 400 controls were ascertained from NSHDS that were age, sex and geographically representative of the cohort. The German Glioma Collection (GGC) comprises 564 individuals who underwent surgery for a glioma at the Department of Neurosurgery, University of Bonn Medical Center, between 1996 and 2008. All histological diagnoses were made at the Institute for Neuropathology/German Brain Tumor Reference Center, University of Bonn Medical Center. Control subjects were 576 healthy individuals (288 men, 288 women; mean age 31, range 18–45) of German origin recruited in 2004 by the Institute of Transfusion Medicine and Immunology, Mannheim.

Ethics

Collection of blood samples and clinico-pathological information from case and control subjects was undertaken with informed consent and relevant ethical review board approval in accordance with the tenets of the Declaration of Helsinki.

Genotyping

DNA was extracted from samples using conventional methodologies and quantified using PicoGreen (Invitrogen). Genotyping of both GWA studies was conducted by Illumina Service laboratory using the Illumina Infinium Human610-Quad BeadChips according to Illumina protocols. DNA samples with GenCall scores <0.25 at any locus were considered ‘no calls’. To ensure quality of genotyping, a series of duplicate samples was genotyped in the same batches.

Intensity data from arrays was imported into Illumina's BeadStudio clustering and calling software application. For the small subset of loci that were not clustered properly by the automated algorithm, the data were reviewed to identify loci that needed to be removed, manually edited or left unchanged. Clustered SNPs were evaluated using the metrics listed in the SNP Table of the BeadStudio software. These metrics are based on all samples for each locus and thus provide overall performance information for each locus. To identify loci potentially needing to be edited or removed, each quality metric column in the SNP table was sequentially sorted. Metrics used for identifying poorly or incorrectly clustered data included intensity, cluster separation, position of each cluster (AA, AB, BB), Hardy-Weinberg equilibrium, call frequency and variation of cluster width. The reproducibility of control samples on each plate as well as replicates were also used to identify misclustered loci. Although not all cluster plots were assessed, ∼10% of the lowest performing loci were examined. Of these 10%, ∼20% were edited or annotated (to indicate loci with nearby polymorphisms or hemizygous deletions) and ∼2% were excluded. Review of data was conducted by a second individual to determine if any metrics were missed or if further editing was required. Overall, this process provided substantially increased genotyping accuracy.

Subsequent genotyping of SNPs was conducted using either competitive allele-specific PCR KASPar chemistry (KBiosciences) or single-base primer extension chemistry MALDI-TOF MS detection (Sequenom). All primers and probes used are available on request. Genotyping quality control was further evaluated through inclusion of duplicate DNA samples in SNP assays. For all SNP assays, >99% concordant results were obtained. Samples having SNP call rates <90% were excluded from the analysis.

Statistical analysis

Statistical analyses were undertaken using R (v2.6), STATA (v8; State College, Texas, US) and PLINK (v1.05)³³ software. All reported P values are two-sided. Genotype data were used to search for duplicates and closely related individuals among all samples in each of the GWA studies. Identity-by-state (IBS) values were calculated for each pair of individuals, and for any pair with allele sharing of >80%, the sample generating the lowest call rate was removed from further analysis. To identify individuals who might have non-Western European ancestry, we merged our case and control data with the 60 western European (CEU), 60 Nigerian (YRI), 90 Japanese (JPT) and 90 Han Chinese (CHB) individuals from the HapMap Project. For each pair of individuals, we calculated genome-wide IBS distances on markers shared between HapMap and our SNP panel and used these as dissimilarity measures upon which to perform principal component analysis. The first two principal components for each individual were plotted and any individual not present in the main CEU cluster (that is, outside 5% from cluster centroids) was excluded from subsequent analyses.

In the UK-GWA study, genotyped samples were excluded from analyses due to non-CEU ancestry (n = 3). In the US-GWA study, genotyped samples were excluded from analyses for the following reasons: relatedness (n = 2) and non-CEU ancestry (n = 23).

The adequacy of the case-control matching and possibility of differential genotyping of cases and controls were formally evaluated using quantile-quantile plots of test statistics. Deviation of the genotype frequencies in the controls from those expected under Hardy-Weinberg equilibrium (HWE) was assessed by a χ² test or Fisher's exact test where an expected cell count was < 5.

The association between each SNP and risk of glioma was assessed by the Cochran-Armitage trend test. Odds ratios and associated 95% CIs were calculated by unconditional logistic regression. Relationships between multiple SNPs showing association with glioma risk in the same region were investigated using logistic regression analysis, and the impact of additional SNPs from the same region was assessed by a likelihood-ratio test.

The combined effect of each pair of risk locus was investigated by logistic regression modeling with evidence for interactive effects between SNPs assessed by a likelihood ratio test. The OR and trend test for increasing numbers of deleterious alleles was estimated by counting two for a homozygote and one for a heterozygote.

Meta-analysis was conducted using standard methods based on weighted average of study-specific estimates of the ORs, using inverse variance weights. Cochran's Q statistic to test for heterogeneity and the I² statistic to quantify the proportion of the total variation due to heterogeneity were calculated. The sibling relative risk attributable to a given SNP was calculated using the following formula:

$${\lambda }^{\ast }=\frac{p{(p{r}_2+q{r}_1)}^2+q{(p{r}_1+q)}^2}{{(p^2{r}_2+2pq{r}_1+q^2)}^2}$$ (1)

where p is the population frequency of the minor allele, q = 1 − p, and r₁ and r₂ are the relative risks (estimated as OR) for heterozygotes and rare homozygotes relative to common homozygotes. Assuming a multiplicative interaction the proportion of the familial risk attributable to a SNP was calculated as log(λ*)/log(λ₀), where λ₀ is the overall familial relative risk estimated from epidemiological studies, assumed to be 2.1 (ref. 2). A naïve estimation of the contribution of all of the loci identified to the excess familial risk of glioma under an additive model was calculated using the following formula:

$$\frac{{\text{OR}}_{\text{per allele}}-1}{2({\lambda }_0-1)}$$ (2)

Bioinformatics

LD metrics between SNPs reported in HapMap were based on Data Release 2/phaseIII Feb09 on NCBI B35 assembly, dbSNPb125, except for between rs2736098 and rs2736100 and rs2853676 which were only available in Data Release 23a/phaseII March08.

We used Haploview software (v3.2) to infer the LD structure of the genome in the regions containing loci associated with glioma risk. Prediction of the untyped SNP in the case-control datasets of GWA data was carried out using IMPUTE and SNPTEST on HapMap. In total, 894 HapMap SNPs were successfully imputed in the interval between 130,051,729 bp and 131,225,253 bp using available SNP genotype data from GWA scans (225 SNPs). We verified the integrity of imputed data, where possible, by crosschecking the concordance of imputed genotypes with that of available Illumina SNP genotype data. Directly genotyped SNPs are denoted in blue; those imputed are shown in green.

To examine whether there is a relationship between SNP genotype and mRNA expression in lymphocytes and normal human cortex, we used publicly available data. We analyzed 90 CEU-derived Epstein-Barr virus–transformed lymphoblastoid cell lines using Sentrix Human-6 Expression BeadChips (Illumina). Online recovery of data was carried out using WGAViewer v1.25 Software^34,35. We analyzed 193 normal human cortex cells using Illumina HumanRef-seq-8 Expression BeadChip arrays³⁶. Where SNPs genotyped in our study had not been typed, HapMap data were used to identify correlated SNPs in high LD (r² > 0.8). Differences in the distribution of levels of mRNA expression between SNP genotypes were compared using a Wilcoxon-type test for trend. Relationship between copy number change and mRNA expression at loci in glioma was investigated using data from The Cancer Genome Atlas (TCGA)¹⁸. The Bioconductor module CGHcall was used to assign copy number status.

We searched for interactions between proteins encoded by genes mapping to association signals using the program PINA³⁷.

URLs

Detailed information on the tagSNP panel, http://www.illumina.com/; WGAViewer, http://www.genome.duke.edu/centers/pg2/downloads/wgaviewer.php; 1958 Birth Cohort: http://www.cls.ioe.ac.uk/studies.asp?section=0001000200030012; PLINK, http://pngu.mgh.harvard.edu/∼purcell/plink/; Cancer Genetic Markers of Susceptibility (CGEMS), http://cgems.cancer.gov/; The Cancer Genomics Data Portal, http://cbio.mskcc.org/cancergenomics-dataportal; The Cancer Genome Atlas (TCGA) data portal, http://tcga-data.nci.nih.gov/tcga/homepage.htm; A Survey of Genetic Human Cortical Gene Expression, http://labs.med.miami.edu/myers/; IMPUTE, http://mathgen.stats.ox.ac.uk/impute/impute.html; SNPTEST, http://www.stats.ox.ac.uk/∼marchini/software/gwas/snptest.html; PINA (Protein Interaction Network Analysis platform), http://csbi.ltdk.helsinki.fi/pina/.