Genome-wide association study identifies three new melanoma susceptibility loci.

We report a genome-wide association study for melanoma that was conducted by the GenoMEL Consortium. Our discovery phase included 2,981 individuals with melanoma and 1,982 study-specific control individuals of European ancestry, as well as an additional 6,426 control subjects from French or British populations, all of whom were genotyped for 317,000 or 610,000 single-nucleotide polymorphisms (SNPs). Our analysis replicated previously known melanoma susceptibility loci. Seven new regions with at least one SNP with P < 10(-5) and further local imputed or genotyped support were selected for replication using two other genome-wide studies (from Australia and Texas, USA). Additional replication came from case-control series from the UK and The Netherlands. Variants at three of the seven loci replicated at P < 10(-3): an SNP in ATM (rs1801516, overall P = 3.4 × 10(-9)), an SNP in MX2 (rs45430, P = 2.9 × 10(-9)) and an SNP adjacent to CASP8 (rs13016963, P = 8.6 × 10(-10)). A fourth locus near CCND1 remains of potential interest, showing suggestive but inconclusive evidence of replication (rs1485993, overall P = 4.6 × 10(-7) under a fixed-effects model and P = 1.2 × 10(-3) under a random-effects model). These newly associated variants showed no association with nevus or pigmentation phenotypes in a large British case-control series.

Cutaneous melanoma is predominantly a disease of fair-skinned individuals. Risk factors include a family history[1], certain pigmentation phenotypes (notably the presence of fair skin, blue or green eyes, blond or red hair, sun sensitivity or an inability to tan[2-5]) and increased numbers of melanocytic nevi[6,7]. We previously reported Phase 1 of a genome-wide association (GWA) study of melanoma based on the Illumina 317k array[8]. This reinforced the importance of these genetically-determined melanoma-associated phenotypes, by showing the major common genetic determinants of risk in the populations considered were the MC1R locus (associated with red hair, freckling and sun sensitivity)[4,5,9,10], tyrosinase (TYR) gene variants which code for skin color[11] and a region near CDKN2A and MTAP which is associated with number of melanocytic nevi[8,12]. Furthermore we confirmed the importance of a haplotype spanning the agouti signaling protein locus (ASIP)[11,13] and a second locus determining nevus count variation at 22q13 identified by a GWA study of nevus count[12].

Both Phase 1 and Phase 2 of this study were carried out by the GenoMEL Consortium, a collaboration focusing on genetic susceptibility to melanoma. The study utilised samples collected by GenoMEL participants across populations of European ancestry living at different latitudes. In total, 14 GenoMEL groups contributed DNA samples from cases and controls of European (or Israeli) ethnicity (Supplementary Table 1). Phase 1 was based on 1,650 cases from Australian and European populations chosen to have a phenotype argued to “enrich” for genetic susceptibility (early onset, multiple primaries, or modest family history of melanoma). In Phase 2 a further 1,523 cases (1,211 of whom are genetically enriched: 532 with a family history, 277 with multiple primaries but no family history, and 402 with early disease onset but no multiple primaries or family history) and 1,112 controls were genotyped using the denser Illumina 610k array (see Supplementary Note). To optimise power for novel gene identification we combined the data from the two phases and performed an overall analysis. The Australian data used in Phase 1 were dropped from the combined Phase 1 + Phase 2 analysis as these samples are included in the Australian GWA study which formed one of the replication studies. After quality control was applied to SNPs and samples (see Supplementary Note), including Principal Components Analysis (PCA) to identify samples of non-European ethnicity (Supplementary Figure 1), the analysis utilised 2,804 cases (2,692 European and 112 Israeli) and 1,835 controls from GenoMEL studies and 5,783 controls from France and the UK Wellcome Trust Case-Control Consortium (WTCCC). A trend test, stratified by geographical region, was applied to each SNP (see Online Methods, also Figure 1). Little evidence was found of population stratification (λ=1.06, see Supplementary Figure 5).

Figure 1

Manhattan plot of results of Cochran-Armitage (CA) trend test stratified by geographic region, with -log10 p-values shown. The solid horizontal line indicates a p-value of 10−5. Markers within 50kb of a SNP associated with melanoma are marked in black for those identified in a previous GWA and replicated here and marked in red if first identified in the current study. The y-axis is truncated at p=10−15, although three SNPs in the MC1R region have stronger p-values, up to 2.7×10−27, as signified by the box and arrow.

Strong evidence was found for the previously identified loci, (Supplementary Table 2, Supplementary Figures 2 and 3)[8,11-18] and another pigmentation gene, SLC45A2, already reported to be associated with melanoma risk[15]. SLC45A2 is involved in melanosome maturation and pigmentation. The rs35390 SNP identified here is associated with melanoma[15] and with variation in hair color[15,20], in accordance with the observed pattern of known melanoma pigmentation risk factors[2-5].

We also confirmed a role for SNP rs401681 in the region of TERT and CLMPT1L, which has also previously been shown to modify melanoma risk[18] (Supplementary Table 2, Supplementary Figures 2 and 3)[8,11-16,21]. The confirmation of the SNP follows reported associations with risk of basal cell carcinomas, hematological malignancies and cancer of the bladder, cervix, lung, pancreas and prostate[18,22]. It was originally reported that the pattern of risk for melanoma was in the opposite direction to that for other cancers[18], and we confirm this observation.

Seven further regions showed evidence of association with melanoma susceptibility (Table 1 and Supplementary Table 3). Replication was sought from two other GWA studies for the SNPs with the strongest evidence, preferentially SNPs common to all arrays. In regions with no SNP common to all platforms, we followed up both our top genotyped SNPs and the most significant imputed SNPs which had been genotyped in the replication studies. Further, these SNPs were genotyped in a replication sample set from the UK and the Netherlands (1,579 cases and 2,036 controls in total, see Supplementary Note). Table 1 contains the evidence from both the hypothesis-generating and replication datasets. Of these seven regions, three (on chromosomes 2 (rs13016963), 11 (rs1801516) and 21) showed strong evidence of replication (p<10−3), three (on chromosomes 2 (rs10932444), 12 and 13) showed no evidence of replication and one (on chromosome 11 (rs1485993)) showed marginal evidence of replication. Three of the loci showed overall combined evidence of association at p<5×10−8, based on the fixed effects meta-analysis.

Table 1

Summary of results from this study for the 4 regions showing evidence of replication, listing each SNP under consideration, their position (in bp) and minor allele frequency (MAF); the per-allele OR (based on the minor allele) and p-value are given for this GWA study, for the meta-analysis of the replication data sets (from the Houston GWA study, the Australian GWA study and UK/Netherlands replication samples) and for the combined genome-wide and replication analyses. The Houston GWA study and the Australian study both used a different array to the current study for at least some samples, so some of their results presented here include imputed data. Further genotyping was conducted in the UK and Netherlands replication samples for SNPs with positive support from the GWA replication data. All meta-analyses are based on a fixed effects model with the exception of those for CCND1, marked with an asterisk, where random effects analysis was used because of the observed heterogeneity. Supplementary Table 3 is a fuller version of this table.

SNP	Chromosome	Position	Allele	MAF	GenoMEL Genome-wide		Replication samples (genotyped + imputed)		Genome-wide plus replication samples (genotyped + imputed)		Postulated gene
					OR	p	OR and 95% CI	P-value	OR and 95% CI	P-value
rs13016963	2	201852173	A	0.37	1.18	5.68 × 10−7	1.11 (1.06, 1.18)	9.2 × 10−5	1.14 (1.09, 1.19)	8.6 × 10−10	CASP8
rs1485993	11	69071595	A	0.37	1.19	4.15 × 10−7	1.07 (1.01, 1.13)	0.017	1.11 (1.04, 1.18)*	0.0012	CCND1
rs1801516	11	107680672	A	0.13	0.79	4.80 × 10−7	0.87 (0.78, 0.90)	3.4 × 10−4	0.84 (0.78, 0.90)	3.4 × 10−9	ATM
rs45430	21	41667951	G	0.39	0.85	5.60 × 10−7	0.91 (0.86, 0.96)	4.2 × 10−4	0.88 (0.85, 0.92)	2.9 × 10−9	MX2

Using random effects

The CASP8 region (chromosome 2) contains a number of SNPs showing evidence for association with melanoma risk; because of the lack of overlap in the SNPs across arrays, we report multiple SNPs either genotyped or imputed across platforms, all of which show evidence of association (Table 1, Supplementary Table 3, Figure 2). The strongest evidence for a single SNP is for rs700635 (p=2.4×10−9, OR=1.15 overall). All the SNPs are in the region of CASP8, which codes for a member of a family of proteases that play a critical role in the control of cell proliferation by inducing apoptotic cell death, making them candidate cancer susceptibility genes. A recent meta-analysis[23] of 3 polymorphisms in CASP8 found that individuals with one or more copies of the D302H variant have a decreased risk of multiple types of cancer. In this study, the D302H variant could be imputed, but showed only marginal evidence of association (p=0.05), suggesting this is not a variant for melanoma. The evidence for melanoma risk was consistent across populations (Figure 3).

Figure 2

Stratified CA trend tests for the three replicated regions on chromosomes 2, 11 and 21. The log10 p-values are from the CA trend test (stratified by geographical region) for genotyped and imputed SNPs, indicated on the left-hand vertical axis. SNPs genotyped for all samples are plotted as circles, SNPs imputed for all samples as crosses and SNPs genotyped for some samples and imputed for others (due to chip differences) as squares. The most significant genotyped SNP is colored purple (with its name above) and the degree of LD between that SNP and the others is indicated by color according to the key (red being the greatest degree of LD). The estimated recombination rate is given by the blue line and indicated on the right-hand vertical axis. The genes in the region and their positions are given underneath the graph. Plots produced using LocusZoom[19].

Figure 3

Forest plot of the per-allele OR for melanoma for SNPs in the 3 regions first identified in this study. Plots show the current evidence for effects by geography, in the genome-wide and replication samples, and by case type (family history, multiple primaries or early onset).

The ATM SNP rs1801516 (chromosome 11) (Table 1, Supplementary Table 3, Figure 2) is a missense mutation (D1853N, G > A) in a gene that codes for a protein which repairs double strand DNA breaks; association has been postulated between ATM and a number of cancer types[24]. For melanoma, the A allele is protective (p=3.4×10−9, OR=0.84 overall). Overall the evidence for melanoma is consistent across populations and case type, with no evidence of heterogeneity (Figure 3).

The third replicated region, around MX2 (chromosome 21), showed consistent effect sizes across the replication datasets (Table 1, Supplementary Table 3, Figure 2) and across populations (Figure 3). The SNP followed up in the replication study is rs45430 (p=2.9×10−9, OR=0.88 overall), which is intronic to MX2 and has not previously been associated with cancer susceptibility.

A fourth region, adjacent to CCND1 (chromosome 11), a proto-oncogene which is a key regulator of cell cycle progression, showed consistent effect sizes across all the replication sets (Table 1, Supplementary Table 3, Supplementary Figure 4), with best overall replication p-value of 0.011. However the replication sets produced a notably smaller OR (1.08) than the discovery set (1.19) (I2=0.507). This is potentially due to the well known Winner's Curse effect[25] that causes the initial discovery set to overestimate the OR, leading in turn to a discrepancy between the overall p-value based on fixed effects and random effects meta-analysis (p=1.7×10−7 and p=0.00046 respectively). Thus, while we have strong support from this region, the evidence cannot be considered conclusive (see Supplementary Note for further details). However, further support comes from the interim analysis of a recently completed melanoma GWA study in494 melanoma cases and 5,628 controls from the Nurses' Health Study and Health Professionals Follow-up Study (OR=1.18 for rs1485993, p=0.014, unpublished data). This gene therefore remains a strong candidate, being well known in melanoma carcinogenesis[26].

In Phase 1 of the study, all melanoma susceptibility loci identified were associated with either skin pigmentation or nevus count variation[8]. For the case-control samples from Leeds, UK, detailed nevus count and pigmentation information has been obtained for cases and controls[27]. Table 2 shows the association between nevus count, pigmentation and all SNPs associated with melanoma. (Note that not all SNPs show convincing evidence of melanoma association within the Leeds case-control samples, reflecting limited power). As expected, MC1R, SLC45A2, IRF4 and TYR are confirmed to be associated with pigmentation, while the rs4911442 SNP on chromosome 20 shows strong association with pigmentation, increasing the evidence that ASIP truly is the focus of this hit and implicating probable linkage disequilibrium (LD) with variants within an ASIP regulatory region. SNPs in the region of CDKN2A/MTAP and PLA2G6 are associated with nevus variation. The TERT/CLMPT1L SNP is also associated with nevus count variation, suggesting its effect on melanoma risk modification may be via this mechanism. We previously showed that IRF4 had a complex relationship with nevus count and melanoma risk[14], and there are suggestions for SNPs in the CASP8 region of a relationship between genotype and nevus count in controls; among cases the association is not apparent (Table 2). Finally, the SNPs in the ATM and MX2 regions show no association with either nevus count or pigmentation, suggesting alternative, unknown mechanisms, although these require evaluation in other populations (see Supplementary Note).

Table 2

Summary of results for nevus count/pigmentation/melanoma analyses from the Leeds case-control samples examining the 11 SNPs replicated for melanoma association in this or previous studies. Results are shown for (i) the proportion and significance of log nevus count variation explained by each SNP, adjusted for age and sex among cases and controls (adjusted for case-control status), (ii) the proportion and significance of case-control adjusted pigmentation variation score explained by each SNP, where the score is calculated from factor analysis of 6 correlated pigmentation phenotypes (see Online Methods), (iii) the association with melanoma risk (both as per-allele OR with 95%CI, and by genotype (compared to a baseline of the homozygote for the common allele)).

Chromosomal Region	Postulated gene	SNP	MAF	% of variation in log nevus count explained by SNP		% of variation in pigmentation explained by SNP		Per-allele OR (95% CI) for risk of melanoma	OR (95% CI) for risk of melanoma with one copy of Minor allele	OR (95% CI) for risk of melanoma with two copies of Minor allele
				R2	P	R2	P
2q33-q34	CASP8	rs13016963	0.33	0.21	0.083*	0.05	0.33	1.25 (1.07, 1.46)	1.26 (1.01, 1.56)	1.56 (1.11, 2.18)
5p15.33	TERT/CLMPT1L	rs401681	0.46	0.50	0.0070	0.13	0.11	1.08 (0.93, 1.25)	1.15 (0.90, 1.47)	1.15 (0.85, 1.55)
5p13.2	SLC45A2	rs16891982	0.03	0.02	0.62	1.33	1.9 × 10−6	0.72 (0.44, 1.18)	0.78 (0.47, 1.30)	NA
6p25-p23	IRF4	rs12203592	0.24	0.21	0.084	2.76	5.6 × 10−12	0.80 (0.67, 0.95)	0.72 (0.58, 0.91)	0.81 (0.49, 1.35)
9p21	CDKN2A/MTAP	rs7023329	0.49	0.29	0.047	0.02	0.55	0.86 (0.73, 1.00)	0.62 (0.47, 0.82)	0.73 (0.53, 1.01)
11q14-q21	TYR	rs1393350	0.27	0.00	0.95	1.07	2.0 × 10−5	1.34 (1.14, 1.58)	1.19 (0.96, 1.49)	2.12 (1.41, 3.19)
11q22-q23	ATM	rs1801516	0.14	0.07	0.33	0.00	0.95	0.88 (0.71, 1.09)	0.93 (0.73, 1.19)	0.59 (0.29, 1.21)
16q24.3	MC1R	rs258322	0.10	0.00	0.81	4.00	9.0 × 10−17	1.83 (1.44, 2.32)	1.71 (1.33, 2.22)	7.14 (1.70, 29.98)
20q11.2-q12	ASIP	rs4911442	0.13	0.07	0.34	0.93	8.2 × 10−5	1.35 (1.08, 1.68)	1.32 (1.03, 1.69)	2.06 (0.85, 5.00)
21q22.3	MX2	rs45430	0.38	0.00	0.80	0.05	0.32	0.90 (0.77, 1.05)	0.97 (0.77, 1.22)	0.77 (0.56, 1.07)
22q13.1	PLA2G6	rs6001027	0.37	0.39	0.018	0.12	0.16	0.78 (0.66, 0.91)	0.79 (0.63, 0.90)	0.60 (0.42, 0.84)
	TOTAL			2.33		9.83

p = 0.004 for controls only

Overall, we report 3 novel loci associated with melanoma risk, which achieve an overall significance level of 5×10−8 based on the fixed effects meta-analysis, and a potential fourth locus. The power to detect SNPs with effect sizes similar to those estimated from the replication studies is low, and we see many more SNPs in novel regions (from across the genome) reaching p-values between 10−4 and 10−5 than expected (68 with MAF>0.05 compared with an expected 46), suggesting that there may be many other genetic regions with a similar effect on melanoma risk (see Supplementary Note). Currently 11 loci have been identified (Table 2), with a suggestion that 5 of these regions act through the pigmentation phenotype and at least 3 through the nevus phenotype, reflecting the major phenotype-associated risk factors for melanoma. Interestingly, at least two of the newly identified loci appear to influence risk through a novel mechanism, opening up potential new directions for melanoma research.