Genome-wide association meta-analyses combining multiple risk phenotypes provides insights into the genetic architecture of cutaneous melanoma susceptibility

Abstract

Most genetic susceptibility to cutaneous melanoma remains to be discovered. Meta-analysis genome-wide association study (GWAS) of 36,760 melanoma cases (67% newly-genotyped) and 375,188 controls identified 54 significant loci with 68 independent SNPs. Analysis of risk estimates across geographical regions and host factors suggests the acral melanoma subtype is uniquely unrelated to pigmentation. Combining this meta-analysis with nevus count and hair color GWAS, and transcriptome association approaches, uncovered 31 potential secondary loci, for a total of 85 cutaneous melanoma susceptibility loci. These findings provide substantial insights into cutaneous melanoma genetic architecture, reinforcing the importance of nevogenesis, pigmentation, and telomere maintenance together with identifying potential new pathways for cutaneous melanoma pathogenesis.

Cutaneous melanoma is a deadly malignancy with increasing incidence and burden in fair-skinned populations worldwide¹. Increased risk for cutaneous melanoma is caused by high exposure to ultraviolet radiation², as well as host factors including family history^3,4, pigmentary phenotypes⁵, number of melanocytic nevi^6,7, longer telomeres^8,9, and immunosuppression¹⁰.

Identified melanoma genetic risk variants include rare, highly penetrant mutations in genes such as CDKN2A^11,12 and POT1^13,14, as well as more common variants (e.g., lower-penetrance variants in MC1R)^15,16. Genome-wide association studies (GWAS) of cutaneous melanoma susceptibility in populations of European ancestry have identified 21 genetic loci reaching genome-wide significance (P < 5 × 10⁻⁸)^17-24. Additional approaches, including family-based analyses of cutaneous melanoma^25,26, combining cutaneous melanoma and nevus count GWAS²⁷ and transcriptome-wide association studies (TWAS)²⁸ have identified further loci that, despite not containing SNPs reaching P < 5 × 10⁻⁸ in a cutaneous melanoma-only GWAS, most likely influence melanoma risk.

This meta-analysis of cutaneous melanoma susceptibility is more than three times the effective sample size of previous cutaneous melanoma GWAS, providing unprecedented power to identify cutaneous melanoma susceptibility variants and enhanced distinction of independent variants in known cutaneous melanoma susceptibility regions. We report here 68 independent cutaneous melanoma associated variants across 54 loci that confirm the importance of key functional pathways and highlight previously unknown cutaneous melanoma etiologic routes (Tables 1-2). Stratified analyses showed a lack of involvement of the pigmentation pathway for acral melanoma, in line with observational data²⁹. The combined analysis of cutaneous melanoma, nevus and hair color GWAS data, and use of expression data through TWAS, identified 31 secondary, potential loci.

Results

Study overview

We performed a GWAS meta-analysis of cutaneous melanoma susceptibility with 30,134 clinically-confirmed cutaneous melanoma cases (Online Methods), 6,626 self-reported cutaneous melanoma cases and 375,188 cutaneous melanoma -free controls from the United Kingdom, United States, Australia, Northern and Western Europe as well as the Mediterranean – a highly sun exposed population often under-represented in cutaneous melanoma studies (Supplementary Table 1). Of these, 24,756 cases (67%) and 358,734 controls (96%) had not been included in any previous melanoma GWAS.

Separately, we performed total (clinically confirmed cases + self-reported cases from 23andMe, Inc. and a subset of UK Biobank cases with only self-reported cutaneous melanoma status) and confirmed-only cutaneous melanoma meta-analyses to determine the power gained by including self-reported cutaneous melanoma cases. Risk loci were deemed genome-wide significant when variants had fixed effects meta-analysis P-values < 5 × 10⁻⁸ (P_meta); where variants exhibited notable heterogeneity (I² > 31%)³⁰ random effects P-values (P_{meta_r}) were also required to be < 5 × 10⁻⁸ (Online Methods). Q-Q plots (Supplementary Figure 1) and linkage disequilibrium score regression³¹ (LDSC; Online Methods) intercepts showed minimal inflation for individual studies (mostly < 1.04; Supplementary Table 1), indicating adequate control of population stratification.

Before including the self-report GWAS data, we used LDSC³¹ to verify their genetic correlation (Rg) with the confirmed-only GWAS meta-analysis (Supplementary Note; Supplementary Table 2). Based on the high Rg and similarity in h² estimates for self-report and clinically confirmed cutaneous melanoma cases (Supplementary Note), we performed an overall total cutaneous melanoma meta-analysis (h²_total= 0.11, 95% CI = 0.08–0.15). The lambda and LDSC intercept for the total cutaneous melanoma meta-analysis indicated that the majority of inflation is due to polygenic signal (λ = 1.165, intercept = 1.054, ratio = 0.17; Supplementary Table 2). A similar h²_total (12%) was estimated using genetic effect-size distribution inference from summary level data (GENESIS; Online Methods)³².

Conditional and joint analysis of the total cutaneous melanoma meta-analysis summary statistics using GCTA³³ identified a total of 54 loci meeting our requirements for genome-wide significance (Online Methods; Figure 1, Extended Data Figure 1-2). Results for loci previously reported by cutaneous melanoma GWAS reaching significance in the total meta-analysis are presented in Table 1. Results for loci not previously reported by a cutaneous melanoma GWAS are summarized in Table 2. In addition to the 54 lead variants, 14 independent variants with linkage disequilibrium (LD) r²_EUR < 0.05 with lead variants at or near 6 loci (TERT, AGR3, CDKN2A, OCA2, MC1R, and TP53) were identified (Supplementary Table 3). Individual regional association plots for the association signals have been provided as Supplementary Data 1. Conditional and joint analysis of summary data identified a further 9 variants at or near SLC45A2, IRF4, AGR3, CCND1, GPRC5A, FTO, and MC1R; however, these additional variants were not carried forward, having either P_meta > 5 × 10⁻⁸ in the single variant analysis or excess heterogeneity (I² > 31%) and P_{meta_r} < 5 × 10⁻⁸ (Supplementary Table 4). In addition, we used GENESIS (Online Methods), which enables a reformulation of the variance explained by associated SNPs to estimate a theoretical optimal area under the curve (AUC), rather than formally testing this using a training and prediction set³² to estimate the potential AUC. The estimated AUC was 66.8%, compared to ~64% in the 2015 cutaneous melanoma meta-analysis²³. This estimate does not include any host factors and would require benchmarking in a prospective study for validation.

Figure 1. Manhattan plot for the total cutaneous melanoma meta-analysis.

−log₁₀ of two-sided P-values for SNPs derived from a fixed-effects inverse variance weighted meta-analysis of logistic regression GWAS (Y-axis) plotted against SNP chromosome positions for the total meta-analysis (36,760 melanoma cases and 375,188 controls; for full details of analysis and covariates included see the Online Methods). The y-axis is limited to −log₁₀(1×10⁻²⁵) to truncate strong signals at loci such as MC1R and ASIP. The full plot is displayed in Extended Data Figure 2. To account for multiple testing, SNPs with a P-value less than 5×10⁻⁸ are deemed significant.

Previous cutaneous melanoma GWAS have identified 21 genome-wide significant loci^17-24. Family-based methods or the combination of cutaneous melanoma with nevus count have identified a further 12 loci including IRF4, MITF, HDAC4, and GPRC5A^25-27. The lead SNPs from many of these loci are associated with pigmentation, tanning response, nevus count, and telomere maintenance (Supplementary Table 5). Other SNPs are proximal to DNA repair genes. Some loci are associated with more than one trait (Tables 1-2). Our analysis confirms 19 of the 21 loci previously reaching genome-wide significance (Table 1; Supplementary Note). The total cutaneous melanoma meta-analysis also confirms the previously reported IRF4 and MITF associations^25-27,34,35, as well as 6 regions previously identified only by combining nevus count and cutaneous melanoma GWAS data²⁷ (Table 2; Supplementary Note). These results demonstrate the ability of cross-trait GWAS to identify disease loci. The remaining 27 loci have not previously been reported as cutaneous melanoma susceptibility loci (Table 2; Supplementary Table 3). The results for the pathologically confirmed-only cutaneous melanoma cases (N = 30,134; Supplementary Table 1) are reported in the Supplementary Note, Extended Data Figure 2-3, Supplementary Table 6. Our full meta-analysis identified 11 loci not found in the confirmed-only GWAS meta-analysis, demonstrating the advantage of including the 6,626 self-reported cutaneous melanoma cases and over 290,000 controls (Supplementary Table 1). Results for SNPs with a fixed or random P < 5 × 10⁻⁷, from the total meta-analysis are reported in Supplementary Table 7.

Melanoma associations by sex, age at diagnosis and subtype

We performed separate GWAS by sex, age at cutaneous melanoma diagnosis (≤40, 40–60, ≥60 years) and major cutaneous melanoma subtypes (superficial spreading (SS), lentigo maligna (LM), nodular melanoma (NM), and acral lentiginous (AL)) to identify variants associated with select subgroups (Supplementary Table 8). Our analysis identified no additional variants after adjustment for multiple testing (5 × 10⁻⁸/9), suggesting that if such variants exist they are undetectable at our current sample size.

We also performed polygenic risk score (PRS) analyses based on the lead independent genome-wide significant SNPs for nevus count (10 variants; Online Methods) and hair color (276 variants; Online Methods) to explore further whether either trait’s association with cutaneous melanoma differs across phenotypic subtypes (significance threshold = 0.05/28; Online Methods). We observed no significant differences in the distribution of the tested PRSs by sex or age at cutaneous melanoma diagnosis. We did, however, detect differences in the distribution of the hair color PRS for the acral lentiginous subtype compared to all non-acral subtypes (P = 2.1 × 10⁻⁴). Our analyses indicated that genetically-predicted pigmentation in acral lentiginous cases was no different to controls (P = 0.65, Extended Data Figure 4) and darker than in superficial spreading, lentigo maligna and nodular melanoma cases (P = 5.3 × 10⁻⁵, 0.01, 4.8 × 10⁻⁴, respectively). These findings provide strong genetic evidence that the pigmentation pathway is far less important for risk of acral lentiginous melanoma than for other subtypes of cutaneous melanoma. No significant differences were observed by subtype for the nevus count PRS.

Variant annotation with cutaneous melanoma risk phenotypes

To investigate possible biological pathways underlying cutaneous melanoma signals, variants independently associated with cutaneous melanoma in the total meta-analysis were evaluated in GWAS of telomere length, tanning response, pigmentation and nevus count (Online Methods, Table 1 and 2, Supplementary Tables 5,7-8). Using a Bonferroni-corrected threshold of phenotype P-value < 0.00074 (0.05/68 independent SNPs), 18 of the 35 novel loci are associated with tanning response or pigmentation (Table 2, Supplementary Table 5), further indicating the importance of pigmentation pathways in cutaneous melanoma susceptibility. Several new loci, including rs12473635 near DTNB and rs78378222 near TP53, are associated with nevus count, reinforcing the role of nevi in cutaneous melanoma susceptibility. Furthermore, four novel loci have previously been associated with telomere length (rs3950296/TERC, rs4731207/POT1, rs2967383/MPHOSPH6, and rs143190905/RTEL1³⁶) (Table 2, Supplementary Table 5) providing additional support for the role of telomere maintenance in cutaneous melanoma susceptibility following earlier findings that genetic determinants of telomere length are generally associated with melanoma risk^13,14,37. Other newly-discovered lead variants are not associated with these phenotypes, suggesting novel pathways.

Additional approaches to identify melanoma risk loci

To identify further loci influencing cutaneous melanoma risk and provide a more nuanced annotation of discovered cutaneous melanoma risk loci, we used a range of secondary approaches with correction for multiple testing (Online Methods). To explore the overlap between cutaneous melanoma loci and established risk factor phenotypes, we combined our total cutaneous melanoma GWAS meta-analysis with a nevus count GWAS meta-analysis (N = 65,777; Online Methods) and separately with a UKBB hair color GWAS (N = 352,662; Online Methods). For the total cutaneous melanoma GWAS meta-analysis and nevus count the Rg is 0.57 (SE = 0.11, P-value = 2.39 × 10⁻⁷), and for hair color scored from light hair to dark (Online Methods) the Rg is 0.290 (SE = 0.096, P-value = 0.0025). Pairwise GWAS (GWAS-PW)³⁸ was used to determine whether loci were associated with only one trait or pleiotropic with both cutaneous melanoma and either nevus count or hair color (Online Methods). Loci previously-reported through the combination of cutaneous melanoma and nevus GWAS²⁷ are now confirmed by our larger cutaneous melanoma GWAS meta-analysis (Table 2). Together these analyses identified secondary potential loci not associated at genome-wide significance levels in the total cutaneous melanoma GWAS meta-analysis. At the Bonferroni-corrected threshold of 1.25 × 10⁻⁸ (Online Methods), they included 8 loci jointly significant for cutaneous melanoma and nevus count, 17 for cutaneous melanoma and hair color, and 4 with cutaneous melanoma, nevus count and hair color (Table 3, Supplementary Table 9- 10).

In parallel, we examined data from a recently-established cell-type specific melanocyte cis-expression quantitative trait loci (eQTL) dataset²⁸ as well as tissue-based cis-eQTL datasets available through Genotype-Tissue Expression (GTEx)³⁹ resource to identify additional susceptibility loci using a transcriptome prediction mapping strategy (or transcriptome-wide association study; TWAS)^40,41. TWAS utilizing these expression datasets enabled gene-based testing for significant cis genetic correlations between imputed gene expression and cutaneous melanoma risk, aiding identification of additional susceptibility loci (Online Methods). While identification of significant genes by TWAS does not establish causation, it can indicate plausible gene candidates to be utilized in pathway analyses and investigated in future functional studies. This analysis built on a previous melanocyte TWAS that analyzed data from a prior cutaneous melanoma GWAS meta-analysis²⁸ and identified significant novel associations between cutaneous melanoma and imputed gene expression of five genes at four loci. Importantly, the CBWD1 locus on chromosome 9 was later identified as a genome-wide significant cutaneous melanoma+nevus count pleiotropic locus²⁷ (Table 3, Supplementary Table 9), and the other three loci (ZFP90 on chromosome 16, HEBP1 on chromosome 12, and MSC/RP11–383H13.1 on chromosome 8) are now at genome-wide significance with cutaneous melanoma in this larger GWAS meta-analysis (Table 2). This confirmation supports the TWAS approach for both identifying new loci and nominating potentially functional genes at GWAS-discovered loci (Tables 1-2).

To empirically identify the target tissues for cutaneous melanoma risk variants, we used partitioned LD score regression⁴² to determine the proportion of total cutaneous melanoma GWAS meta-analysis heritability that could be captured by genes expressed in melanocytes and in 50 GTEx tissue types. We found that partitioned cutaneous melanoma heritability was most enriched in genes specifically expressed in melanocytes (2.76-fold, P = 3.12 × 10⁻⁶ for top 4,000 genes; Extended Data Figure 5), followed by three other skin-related tissues (GTEx sun-exposed and non-sun-exposed skin, transformed skin fibroblasts). This enrichment was much stronger than the one based on the previously published melanoma GWAS²³. We then focused on these four tissues for discovery of new loci, applying Bonferroni correction for multiple comparisons based on the number of genes tested within each tissue set (Online Methods). TWAS using the melanocyte dataset (Supplementary Table 11, Supplementary Table 3) identified a total of 40 significant genes. Combining genes within 1 Mb of each other into discrete loci, 32 genes were located within 13 formally genome-wide significant cutaneous melanoma GWAS loci, and eight genes were identified within six novel loci. Considering the other skin-related tissues collectively (Supplementary Table 12, Supplementary Table 3), TWAS identified a single significant gene at one additional novel locus, as well as genes within 15 GWAS-significant loci. The TWAS using all GTEx tissues is reported in Supplementary Table 13.

In aggregate, these complementary approaches identified a total of 85 discrete loci (Figure 2; Supplementary Table 14): 54 formally significant at P < 5 × 10⁻⁸ in the total cutaneous melanoma meta-analysis (Table 1-2, Supplementary Table 3), and the remainder supported by one or more of the secondary analyses (Table 3-4, Supplementary Tables 7-10,14) and likely representing additional cutaneous melanoma risk loci, but requiring a larger sample size to reach genome-wide significance. In order to annotate cutaneous melanoma GWAS loci for candidate susceptibility genes for pathway analyses as well as future functional studies, we turned to eQTL colocalization analyses. These approaches identified multiple pathways that may play a role in developing melanoma and are described in the Supplementary Note.

Figure 2. Overlap of loci identified by primary and secondary analyses.

Loci identified in the total cutaneous melanoma meta-analysis (CM, green, Supplementary Table 3), the pleiotropic analysis with nevus count (CMnev, blue, Supplementary Table 9) and hair color (CMpig, red, Supplementary Table 10), melanocyte TWAS (TWASmel, yellow, Supplementary Table 10), and TWAS using the expression of three skin tissues (TWAS3skin, orange, Supplementary Table 12).

Discussion

We report the largest cutaneous melanoma GWAS meta-analysis to date with over three times the effective sample size of prior analyses (Supplementary Table 1). We identified 68 independent cutaneous melanoma -associated variants across 54 loci. TWAS analysis, eQTL colocalization and multi-marker genomic annotations, identified promising gene candidates at many of these risk loci. Joint pairwise GWAS with the cutaneous melanoma-related traits of nevus count and hair color, and TWAS identified a further 31 independent loci that, while not formally reaching genome-wide significance for cutaneous melanoma alone, represent potential additional risk loci. Our cutaneous melanoma meta-analysis also confirmed several loci previously identified only by TWAS²⁸, supporting the value of TWAS in identifying additional genes associated with cutaneous melanoma (Table 4). In total, our integrative analysis identified 85 loci associated with cutaneous melanoma susceptibility (Tables 1-4, Figure 2), constituting a substantial increase from the 21 loci previously identified by cutaneous melanoma susceptibility GWAS alone (Table 1), in addition to those found by family-based approaches or in combination with nevus GWAS data (Table 2).

Our analyses showed strong genetic correlation between self-reported and clinically-confirmed cases (Supplementary Table 2; Supplementary Note), and inclusion of self-reported cases enabled the identification of 11 additional cutaneous melanoma susceptibility loci (Supplementary Tables 3,6; Supplementary Note), indicating that self-reported cutaneous melanoma cases are a valuable and reliable resource for genomic cutaneous melanoma studies. Furthermore, we assessed cutaneous melanoma genetic susceptibility across several geographic regions, including the often-underrepresented Mediterranean population. Interestingly, we found little evidence for difference in cutaneous melanoma locus effect estimates by contributing GWAS (Supplementary Figure 2) or differences in effect size and allele frequency by geographic regions (Supplementary Figure 3), beyond minor variation in pigmentation genes (e.g., rs6059655 near ASIP and rs1805007 near MC1R). The stratified analysis based on cutaneous melanoma histological subtypes identified acral lentiginous melanomas as being uniquely unassociated with pigmentation loci, in line with observational data²⁹. In contrast, the stratified analyses based on age at diagnosis and gender found no evidence for differences in the distribution of nevus-related or pigmentation-related loci.

The discovery of new loci and genes augments our understanding of cutaneous melanoma risk and provides many new insights into cutaneous melanoma etiology. Many of the loci previously associated with nevus count²⁷ or pigmentation⁵⁷ are also associated with cutaneous melanoma (Table 2) confirming the close relationship between these traits. Specifically, of 10 loci previously significantly-associated in a joint analysis of cutaneous melanoma and Nevus, but not associated with cutaneous melanoma alone²⁷, 6 are now associated with cutaneous melanoma alone (Table 2), demonstrating the benefits of conducting joint analyses. The remaining 4 loci reach P < 5 × 10⁻⁸ in the joint cutaneous melanoma+Nevus analysis (Supplementary Table 9); 3 of which are significant at the Bonferroni corrected threshold of 1.25 × 10⁻⁸ (Table 3). In turn, we conducted further pleiotropic analyses and identified secondary loci associated with a combination of both these traits and cutaneous melanoma, but not significantly associated with cutaneous melanoma alone (Table 3). Loci found in such joint analyses are of value as they would likely be associated with cutaneous melanoma alone in a sufficiently large cutaneous melanoma GWAS meta-analysis. These joint analyses provide a direct biological interpretation that several GWAS risk loci may act through nevus development, in line with clinical evidence. Interestingly, following these expanded pleiotropic analyses, many loci were associated with neither nevus count or hair color, indicating that many risk variants act outside of these classic cutaneous melanoma risk phenotypes (Tables 1-2).

The discovery of many new loci, when added to the existing catalog of melanoma risk loci, augments our understanding of the genetic architecture of cutaneous melanoma, as discussed in the Supplementary Note. It is important to note that confirmation of the genes we have identified are causal for cutaneous melanoma, and the biological understanding of how variants at these loci influence cutaneous melanoma, remains to be functionally established. For example, melanocyte eQTL and TWAS analyses indicated PARP1 expression was associated with cutaneous melanoma risk SNPs at 1q42^28,58. While PARP1 is an established DNA repair gene, extensive functional characterization of the cutaneous melanoma risk locus over PARP1 demonstrated that its role in cutaneous melanoma appears to be through regulation of melanocyte proliferation, senescence, and transcriptional regulation of the key melanoma oncogene MITF ⁵⁸. Despite the need for follow-up functional studies, a preliminary, complex model of pathways potentially important for the development of melanoma is emerging through the candidate genes suggested by this and prior work, including pathways mediating protection against UV-induced DNA damage and DNA repair, telomere maintenance, immunity, melanocyte differentiation, and cell adhesion.

For example, we identified an association between multiple independent variants at the TP53 locus, rs78378222 and rs1641548, and cutaneous melanoma further reinforcing the potential importance of DNA repair and genome integrity for cutaneous melanoma susceptibility (see Supplementary Note). Rare germline mutations in TP53 lead to Li-Fraumeni syndrome⁵⁹ which is associated with early onset of cancer, including cutaneous melanoma⁶⁰. Notably, one of the common sequence variants we found to be associated with cutaneous melanoma has previously been shown to alter TP53 mRNA levels by disruption of TP53 polyadenylation. TP53 responds to cellular stresses to regulate target gene expression resulting in DNA repair, cell cycle arrest, apoptosis, and cellular senescence^61,62; variation resulting in loss of normal TP53 function could result in clonal expansion of cells that carry accumulated mutations, which may explain the association with both cutaneous melanoma and nevus count.

This study also adds to a growing body of evidence supporting a key role for telomere maintenance in cutaneous melanoma susceptibility^{8,9,13,14,37,51,63}, with cutaneous melanoma risk loci associated with telomere length or located near prominent telomere maintenance genes or loci, including POT1, TERC, RTEL1, MPHOSPH6, and OBFC1. Additional previously-identified GWAS loci are located near CCND1 (rs4354713), ATM (rs1801516), and PARP1 (rs2695237), all genes with established roles in telomere maintenance, DNA repair, and regulation of senescence^64,65.

The well-established role of immunity in melanoma biology has fueled a search for an association between variation within the HLA region and melanoma risk^66-68. While several studies have investigated associations between HLA alleles and cutaneous melanoma, these studies have largely been conducted on small, underpowered datasets and have not been consistently replicated^69-79. Here, we report identification of a genome-wide significant association between cutaneous melanoma susceptibility and rs28986343 at the HLA locus (see Supplementary Note). This additional evidence for a role for immunity adds to previous²⁸ and current TWAS and colocalization analyses suggesting association between rs408825 and expression of the innate immunity gene MX2. Additionally, many risk alleles for the autoimmune melanocyte-related disorder vitiligo^48,80 are protective for cutaneous melanoma with the lead SNPs either identical (rs1126809/TYR; rs6059655/ASIP), or in strong LD with cutaneous melanoma lead SNPs (rs251464 near PPARGC1B for vitiligo, rs32578 for melanoma, LD r²_EUR = 0.73; rs72928038 near BACH2 for vitiligo, rs6908626 for melanoma, r²_EUR = 0.95; rs1129038 near OCA2 for vitiligo, rs12913832 for melanoma, r²_EUR = 0.99). While the vitiligo and cutaneous melanoma associations share many similar loci, suggesting a role for immunity, we cannot rule out their action on cutaneous melanoma risk being through pigmentation or protection against UV damage. Taken as a whole, these data suggest further investigation into these potentially immune-related associations, and more broadly the role of immunity in melanoma risk.

New loci emerging from these analyses suggest a role of genes or networks regulating the development and differentiation of the melanocytic lineage. The cutaneous melanoma meta-analysis identified a locus near FOXD3, while the pleiotropic cutaneous melanoma+Nevus analysis and TWAS locus identified a novel locus significantly associated with allelic expression of NOTCH2 in melanocytes (Supplementary Note). FOXD3 participates as a part of a larger gene regulatory network governing the development of melanocytes from the neural crest, at least in part through transcriptional repression of one of the earliest markers of melanoblast development (and melanoma predisposition gene), MITF^81,82. NOTCH2, as well as NOTCH1, appear to play roles in both development of the melanocyte lineage as well as maintenance of melanocyte stem cells^53,83, and NOTCH signaling has been shown to lead to de-differentiation of melanocytes to multipotent neural crest stem-like cells⁸⁴. These two new candidate susceptibility genes join previously-identified loci also harboring genes involved in melanocyte fate. Whole-genome and targeted sequencing studies of melanoma-prone families led to the identification of a functional intermediate-penetrance missense mutation of MITF associated with both melanoma and nevus count (MITF p.E318K)^25,26, a variant that was rediscovered by this population-based meta-analysis (rs149617956, P = 5.17 × 10⁻²⁵, OR = 0.38). Additionally, a previously-identified melanoma and nevus risk locus⁸⁵ is located ~200 kb from SOX10, another key regulator of melanocyte development and differentiation and direct transcriptional activator of MITF. These genes, and others in this gene regulatory network, have likewise been variously implicated in the progression of melanoma^86-90.

The identification of a cutaneous melanoma risk locus for which risk genotype strongly correlates with higher melanocyte-specific expression of CDH1, encoding E-cadherin, suggests a potential role for cell-cell adhesion in melanoma risk (see Supplementary Note). E-cadherin plays a crucial role in cell-cell adhesion, epithelial-mesenchymal transition (EMT) and carcinoma progression. Germline mutations in this gene are associated with a variety of tumors including gastric⁹¹, breast⁹², and potentially colorectal cancer⁹³. In human skin, E-cadherin is typically expressed on the cell surface of both melanocytes and keratinocytes and is considered the major adhesion molecule between these two cell types^54,55. During melanoma progression, expression of E-cadherin is typically lost, with a concurrent switch to expression of N-cadherin, facilitating preferential association with fibroblasts and vascular endothelial cells⁵⁵. In contrast to loss of E-cadherin expression with melanoma progression, we find the cutaneous melanoma risk allele at this locus to be associated with higher expression of CDH1. Interestingly, melanocytes in non-lesional skin of vitiligo patients have been found to have loss of or discontinuously distributed E-cadherin expression. This loss of E-cadherin induces reduced adhesiveness to the basal layer under oxidative and mechanical stress, leading melanocytes to migrate passively to the exterior of the skin, and die by apoptosis⁹⁴. Thus, germline variation leading to higher melanocyte CDH1 could act as a protective mechanism, allowing cells damaged by oxidative stress to remain in the skin and survive without dying. A similar mechanism has been recently identified in breast cancer metastasis, where E-cadherin acts as a survival factor by limiting reactive oxygen-mediated apoptosis⁹⁵.

In summary, our large, international genetic meta-analysis showcases the utility of including self-reported cutaneous melanoma cases, complementary analytical approaches, and data from multiple sources to expand our understanding of cutaneous melanoma risk. While the biological mechanisms underlying many of the existing and novel cutaneous melanoma risk loci remain to be confirmed or discovered by post-GWAS functional studies and even larger GWAS, these data suggest potential pathways novel to melanoma susceptibility, and highlight nevus formation, pigmentation and telomere maintenance, the three pathways that appear to dominate the landscape of melanoma susceptibility.

Table 1.

Loci previously identified in cutaneous melanoma susceptibility GWAS. CHR, BP: hg19 positional information. rsID: dbSNP142 rs number. Publications. We also summarize Supplementary Table 3; Gene prioritizes the functional target if known, followed by melanocyte or skin tissue TWAS data, or finally the closest protein coding gene; ‘Multiple’ indicates three or more genes. GWS: We indicate with yes (Y) or no (N) whether this locus is genome-wide significant (P < 5 × 10⁻⁸) in the total meta-analysis. The effect allele (EA) and non-effect allele (NEA) are listed, as are the effect allele Frequency in the HRC reference panel¹⁰⁷; total fixed-effects inverse-variance weighted meta-analysis of logistic regression two-sided P-value (P_meta) and Odds Ratio (OR) are from an additive model and are reported per-allele with respect to the EA. Reported results are for the total meta-analysis (36,760 melanoma cases and 375,188 controls; for full details of analysis and covariates included see Online Methods). We also indicate whether this locus is associated with other traits: Nevi: Pleiotropically associated with cutaneous melanoma and nevus count (Online Methods; Supplementary Table 9); Hair: Pleiotropically associated with cutaneous melanoma and hair color (Online Methods; Supplementary Table 10). Tanning response (Tan) and Telomere length (Telo) indicates the lead SNP is associated with these traits when corrected for multiple testing (Online Methods, Supplementary Table 5).

CHR:BP	rsID	Pub	Gene	EA/ NEA	Freq	Pmeta	OR	Nevi	Hair	Tan	Telo
1:150,938,571	rs8444	108	Multiple	G/A	0.645	3.89 × 10−14	1.08	-	-	Y	-
1:226,603,635	rs2695237	27,108,109	PARP1	T/C	0.628	1.53 × 10−18	1.10	Y	-	-	-
2:38,298,139	rs1800440	23,27	CYP1B1	T/C	0.824	6.97 × 10−15	1.10	Y	-	Y	-
2:202,143,928	rs10931936a	20	CASP8	T/C	0.281	2.17 × 10−8	1.08	-	-	-	-
5:1,323,212	rs13178866a	20,110,111	TERT	C/T	0.554	2.59 × 10−18	0.87	-	Y	-	Y
5:33,951,693	rs16891982a	20,34,111	SLC45A2	C/G	0.122	1.96 × 10−28	0.51	-	Y	Y	-
6:21,163,919	rs6914598	23	CDKAL1	T/C	0.683	1.18 × 10−18	0.91	-	-	Y	-
7:17,134,708	rs117132860b	23,57	AGR3	G/A	0.981	3.83 × 10−21	0.71	Y	-	Y	-
9:21,803,880	rs871024a	18,27	MTAP, CDKN2A	C/A	0.477	2.72 × 10−23	1.18	Y	Y	-	-
9:109,054,417	rs10739220	23,27	TMEM38B	C/T	0.260	1.34 × 10−18	1.10	Y	Y	-	-
10:105,694,301	rs7902587	23,27	OBFC1	C/T	0.904	2.68 × 10−23	0.86	Y	-	-	Y
11:69x,380,898	rs4354713	20,23	CCND1	A/G	0.356	8.50 × 10−21	1.10	-	Y	-	-
11:89,017,961	rs1126809a	18	TYR	G/A	0.757	4.78 × 10−37	0.83	-	Y	Y	-
11:108,175,462	rs1801516	20	ATM	G/A	0.856	2.22 × 10−21	1.14	Y	-	-	-
15:28,365,618	rs12913832a	19,23	OCA2	A/G	0.335	4.85 × 10−12	0.88	-	Y	Y	-
16:89,986,117	rs1805007a	18	MC1R	C/T	0.937	5.86 × 10−52	0.57	Y	Y	Y	-
20:32,665,748	rs6059655a	17,18	ASIP	A/G	0.061	2.52 × 10−42	1.45	-	Y	Y	-
21:42,743,496	rs408825	20	MX2	C/T	0.413	1.03 × 10−32	0.89	-	-	Y	-
22:38,545,942	rs132941	18,27,35	MAFF	T/C	0.549	8.80 × 10−23	1.10	Y	-	Y	-

^aVariant meta-analysis results are heterogeneous (I² > 31%) and random effects estimates are presented.

^bWhile this locus overlaps the previously reported IRF4 or AGR3 locus, the lead variants are independent.

Table 2.

Novel loci not previously identified in cutaneous melanoma GWAS. CHR, BP: hg19 position. rsID: dbSNP142 rs number. Gene prioritizes the functional target if known, followed by melanocyte or skin tissue TWAS data, or finally the closest protein coding gene; multiple indicates three or more genes (Supplementary Table 3). The effect allele (EA) and non-effect allele (NEA) are listed, as are the effect allele Frequency in the HRC reference panel¹⁰⁷; total fixed-effects inverse-variance weighted meta-analysis of logistic regression two-sided P-values and Odds Ratio (OR) are with respect to the EA. Reported results are for the total meta-analysis (36,760 melanoma cases and 375,188 controls; for full details of analysis and covariates included see Online Methods). Nevi: Associated with cutaneous melanoma+nevus count (Online Methods; Supplementary Table 9); Hair: Associated with cutaneous melanoma+hair color (Online Methods; Supplementary Table 10). Tanning response (Tan) and Telomere length (Telo) indicate lead SNP is associated with these traits when corrected for multiple testing (Online Methods, Supplementary Table 5).

CHR:BP	rsID	Gene	EA/ NEA	Freq	Pmeta	OR	Nevi	Hair	Tan	Telo
1:63,727,542	rs670318	FOXD3	T/C	0.047	1.21 × 10−8	0.86	-	-	Y	-
1:154,994,978	rs76798800	ZBTB7B, ADAM15, GBA	G/T	0.753	3.86 × 10−15	0.92	Y	-	Y	-
1:205,181,062	rs2369633	DSTYK	T/C	0.083	1.24 × 10−8	1.10	-	-e-	Y	-
2:25,778,637	rs12473635	DTNB	T/C	0.776	5.17 × 10−9	0.93	Y	-	-	-
3:70,014,091	rs149617956a	MITF	G/A	0.998	9.00 × 10−14	0.39	-	Y	Y	-
3:169,493,283	rs3950296b	TERC	C/G	0.747	4.47 × 10−11	1.08	Y	-	-	Y
5:90,262,612	rs12523094c	GPR98	T/C	0.567	1.74 × 10−6c	1.07	-	Y	Y	-
5:149,211,868	rs32578b,d	PPARGC1B	G/A	0.658	6.58 × 10−17	1.09	Y	-	Y	-
6:1,145,265	rs12215602a	IRF4	G/A	0.721	7.91 × 10−9	0.94	Y	-	Y	-
6:22,719,379	rs72834823	HDGFL1	T/A	0.819	1.04 × 10−12	1.10	Y	-	Y	-
6:32,748,953	rs28986343	HLA-DQB2	C/T	0.952	1.61 × 10−8	1.15	-	-	-	-
6:91,005,743	rs6908626	BACH2	G/T	0.844	3.92 × 10−9	1.09	-	-	-	-
7:22,115,454	rs12539524	RAPGEF5	C/T	0.846	1.65 × 10−8	0.93	-	-	-	-
7:124,396,645	rs4731207	POT1	G/A	0.540	2.22 × 10−15	0.93	Y	-	-	Y
7:130,738,666	rs7778378	MKLN1	C/T	0.248	8.93 × 10−9	0.93	Y	Y	-	-
8:21,951,009	rs6994183	FAM160B2	A/T	0.866	4.84 × 10−9	0.92	-	-	-	-
8:72,864,240	rs13263376c	RP11-383H13.1, MSC	G/A	0.364	2.28 × 10−8c	0.93	Y	-	Y	-
9:12,587,153	rs10960710	TYRP1	G/T	0.393	3.08 × 10−12	0.93	-	Y	Y	-
9:110,711,586	rs1339759b	KLF4	C/G	0.666	5.61 × 10−19	1.10	Y	-	-	-
9:134,457,580	rs3780269	RAPGEF1	G/A	0.691	1.92 × 10−8	0.94	Y	-	-	-
11:16,041,305	rs7941496	SOX6	G/T	0.516	1.40 × 10−9	1.06	Y	-	Y	-
11:120,195,702	rs12290699	TMEM136	T/C	0.745	2.20 × 10−8	0.94	-	-	-	-
12:13,070,752	rs1056927b,c	Multiple	A/G	0.561	2.74 × 10−9b	0.93	Y	-	-	-
12:17,275,460	rs4237963	LMO3	T/A	0.207	1.27 × 10−9	0.93	-	-	-	-
12:96,378,807	rs10859996	HAL, RP11-256L6.3	C/T	0.635	2.09 × 10−10	1.07	-	-	-	-
12:116,580,291	rs113469387	MED13L	G/A	0.907	8.76 × 10−10	0.91	-	Y	Y	-
13:113,535,949	rs1278768	MCF2L	G/C	0.488	6.33 × 10−12	0.94	-	-	Y	-
15:33,277,710	rs117648907b	FNM1	C/T	0.983	7.29 × 10−12	0.80	Y	-	-	-
16:68,822,971	rs4420522	Multiple, CDH1	A/G	0.690	8.34 × 10−14	0.93	Y	Y	-	-
16:82,217,153	rs2967383	MPHOSPH6	G/T	0.267	2.24 × 10−9	1.06	-	-	-	Y
17:7,571,752	rs78378222	TP53	T/G	0.989	3.33 × 10−10	0.76	Y	-	-	-
19:3,540,539	rs12984831b	MFSD12	G/C	0.984	3.86 × 10−10	0.65	Y	-	Y	-
20:62,291,767	rs143190905	RETL1	G/T	0.907	6.54 × 10−13	1.15	-	-	-	Y
22:45,622,684	rs5766565	KIAA0930	A/G	0.647	1.44 × 10−9	1.06	Y	Y	Y	-
22:50,722,408	rs79966207	PLXNB2	T/C	0.849	8.68 × 10−9	0.92	-	Y	-	-

^aAssociated with cutaneous melanoma by non-GWAS based approaches - MITF^25,26, IRF4 ^27,34,35.

^bPreviously associated pleiotropically with cutaneous melanoma and nevus count²⁷.

^cVariant meta-analysis results are heterogeneous (I²>31%) and random effects estimates are presented. For rs12523094/GPR98 while the lead SNP selected in conditional mapping is heterogenous, other SNPs in LD pass this requirement (e.g., rs12173258, r²_EUR = 0.9, P_meta = 1.09 × 10⁻¹¹, I² = 29.6).

^dPreviously associated with tanning response⁵⁷.

^eJoint cutaneous melanoma+hair color P-value is greater than multiple testing corrected threshold of 1.25 × 10⁻⁸ (Supplementary Table 10).

Table 3.

Novel pleiotropic associations with cutaneous melanoma and nevus count or hair color. Reported cutaneous melanoma P-values are from the total fixed-effects inverse-variance weighted meta-analysis of logistic regression two-sided P-values from GWAS representing a total of 36,760 melanoma cases and 375,188 controls (Online Methods). Results for the lead variants from pleiotropic loci (lead SNP reaching P < 5 × 10⁻⁸ following a Stouffers sample size weighted meta-analysis of cutaneous melanoma P-values and either Nevus GWAS meta-analysis (N = 65,777) or Hair Color GWAS (N=352,662) and GWAS-PW Model 3 prior probability of association (PPA) > 0.5, Online Methods) distinct to those in the total cutaneous melanoma meta-analysis (Table 1, 2). CHR, BP: hg19 positional information. rsID: dbSNP142 rs number. Gene prioritizes genes that the variant is an eQTL for in GTEx skin datasets or otherwise is the closest protein coding gene; multiple indicates three or more genes. We report the total cutaneous melanoma meta-analysis P (CM P), and the CM+nevus or CM+hair color Stouffer’s meta-analysis fixed effect P-value. Full results can be found in Supplementary Tables 7 and 10.

CHR:BP	rsID	Gene	CM P	CM + Nevus P	CM + Hair P
1:24787947	rs195720	NIPAL3	7.97 × 10−6	-	2.24 × 10−12
1:78450517	rs34517439	DNAJB4	2.23 × 10−4	-	2.17 × 10−12
1:214673271	rs7533482	PTPN14	2.79 × 10−5	-	2.45 × 10−13
2:135430709	rs6745983	TMEM163	1.69 × 10−3	-	7.00 × 10−13
2:214065880	rs16849932	IKZF2	1.46 × 10−3	-	1.18 × 10−10
2:240065356	rs11677464a	HDAC4	4.00 × 10−5	1.10 × 10−9	-
4:37470753	rs11730662	KIAA1239	1.82 × 10−3	1.19 × 10−8	-
5:56011357	rs7714232	MAP3K1	6.99 × 10−4	-	3.32 × 10−22
6:7189567	rs75818295	RREB1	1.87 × 10−3	-	8.27 × 10−10
6:11637483	rs548304	ADTRP	2.67 × 10−5	-	1.46 × 10−10
6:15503696	rs10949304	DTNBP1	1.7 × 10−3	4.96 × 10−9	-
6:50790642	rs2857482	TFAP2B	3.59 × 10−5	3.44 × 10−10	-
6:151577739, 6:151577830	rs10434895, rs10434896b	AKAP12	8.17 × 10−8, 7.88 × 10−8	7.71 × 10−10	2.07 × 10−42
8:131138979	rs111595456	ASAP1	3.86 × 10−4	2.83 × 10−10	-
9:211762, 9:235201	rs520015, rs593179a,c	CBWD1	8.95 × 10−7, 3.78 × 10−6	4.13 × 10−12	1.10 × 10−43
10:5767177	rs76154345a	GDI2	4.43 × 10−6	7.80 × 10−11	-
10:111889779	rs11194997	MXI1	3.45 × 10−6	-	2.70 × 10−11
11:7543519	rs11041426	PPFIBP2	2.73 × 10−4	-	1.66 × 10−33
11:62203865	rs10897275	AHNAK	6.47 × 10−5	-	2.47 × 10−33
11:91616691	rs12225068	FAT3	3.80 × 10−5	-	6.48 × 10−10
13:76351286	rs474240	LMO7	2.53 × 10−4	-	9.28 × 10−9
13:114744546	rs75414584	RASA3	6.31 × 10−3	-	4.62 × 10−12
14:64390030	rs10873172a,d	SYNE2	6.29 × 10−8	5.95 × 10−13	6.47 × 10−27
14:69226931	rs11625064d	ZFP36L1	3.33 × 10−6	2.09 × 10−10	1.83 × 10−19
14:92795912	rs4904871	SLC24A4	2.06 × 10−4	-	2.15 × 10−278
14:103923475	rs2273699	MARK3	5.27 × 10−5	-	1.21 × 10−16
15:48400199	rs2675345	SLC24A5	4.92 × 10−3	-	1.09 × 10−9
16:54118132, 16:54131939	rs62034121, rs62034139a,e	FTO	1.16 × 10−9, 4.56 × 10−9	4.69 × 10−14	-
16:55322732	rs12930459a	IRX6	1.82 × 10−5	4.89 × 10−9	-

^aLocus previously reported as pleiotropically associated with cutaneous melanoma and nevus count, but not significant for cutaneous melanoma alone here.

^bLead SNP for Pigment (rs10434895) and nevus (rs10434895) are in LD r²_EUR = 1.0.

^cLead SNP for Pigment (rs520015) and nevus (rs593179) are in LD r²_EUR = 0.63.

^dSame lead SNP.

^eLead SNP for Pigment (rs62034121) and nevus (rs62034139) are in LD r²_EUR = 0.88.

Table 4.

Genes identified by TWAS outside of regions identified in the total cutaneous melanoma GWAS meta-analysis. For each gene with a Bonferroni-corrected P-value cutoff in melanocytes (P_TWAS < 3.22 × 10⁻⁶), or skin-related tissue types (P_TWAS < 5.28 × 10⁻⁷) that does not overlap with an existing cutaneous melanoma region we report the local peak cutaneous melanoma variant from the total confirmed plus self-report GWAS meta-analysis, and TWAS Z score. Full results for all genes with a P_TWAS < 1.48 × 10⁻⁵ can be found in Supplementary Tables 10,12. CBWD1 and C9orf66 are within 1 Mb of each other and are merged into a single locus. * RP11-676J12.7 was identified using sun-exposed skin expression data from GTEx (Supplementary Table 12), while all other genes were identified using melanocyte gene expression.

	TWAS		Locus Peak CM Variant
Gene	Z	P	rsID	CHR:BP	CM P
NIPAL3	4.84	1.28 × 10−6	rs2294524	1:24,770,594	2.74 × 10−7
RCAN3	4.83	1.33 × 10−6	rs2294524	1:24,770,594	2.74 × 10−7
NOTCH2	4.81	1.50 × 10−6	rs2793830	1:120,466,108	3.80 × 10−7
PTPN14	−4.84	1.30 × 10−6	rs6693492	1:214,685,978	2.68 × 10−5
CBWD1	−4.81	1.51 × 10−6	rs478882	9:205,964	1.64 × 10−6
C9orf66	5.05	4.48 × 10−7	rs478882	9:205,964	1.64 × 10−6
SYNE2	5.19	2.06 × 10−7	rs12881652	14:64,400,120	2.12 × 10−7
IRX6	−4.80	1.62 × 10−6	rs12919110	16:55,319,789	1.27 × 10−6
RP11-676J12.7*	−5.55	2.79 × 10−8	rs1703824	17:813,324	1.59 × 10−5

Online Methods

Quality control metrics, imputation and association analysis

Data cleaning was performed using Illumina GenomeStudio/BeadStudio (v2.0.4 San Diego, CA, USA) and PLINK (v1.90b5.4) ^96,97. Full details of the sample collections and genotyping arrays used for each GWAS are reported in the Supplementary Note. Prior to imputation any SNP with either minor allele frequency (MAF) < 0.01, Hardy-Weinberg Equilibrium (HWE) P-value < 5 x 10⁻⁴ in controls or < 5 x 10⁻¹⁰ in cases was removed. Similarly, any individual was removed who was missing > 3% of variants, had heterozygosity values either > 0.05 or < −0.05 or 3 sd from the mean, whose genetically-predicted sex did not match their recorded sex, or who was determined to be non-European based on principal component analysis (PCA). In addition, one of any pair of individuals estimated to be related with identity by descent (IBD) pihat > 0.15 was removed.

The Harvard, BNMS, and 23andMe GWAS were imputed to 1000 Genomes Project phase 1 v3; for all other sets (Supplementary Table 1) imputation was conducted using the Michigan Imputation Server with the Haplotype Reference Consortium panel (HRC version 1) and run using Minimac3 ⁹⁸. Following imputation, any imputed variant with imputation quality score r² < 0.5 or MAF < 0.0001 was rejected. As rare SNPs where one allele is missing in the case or control group can lead to very large (or infinite) OR estimates, variants with an OR < 1 × 10⁻⁴ (the minimum reported by PLINK) or > 1 × 10⁶ were also filtered. To handle variants with the same name (e.g., triplicate SNPs), variant IDs were converted to the format CHR:BP:A1A2 prior to meta-analysis.

Logistic regression under an additive model with ORs calculated on a per-allele basis was then conducted using PLINK (v1.90b5.4) ^96,97 with either geographic region (in GenoMEL Phase 1 and 2 data) or principal components as covariates to account for potential population stratification. Individual studies were checked for evidence of inflation by producing QQ plots (Supplementary Figure 1) and calculating the corresponding inflation factor λ and LDSC intercept (Supplementary Table 1).

Where individual studies have deviated from this protocol, details are included in the study description in the Supplementary Note. All reported tests are two-sided.

Meta-analysis and conditional-and-joint-analysis to identify independent loci

Meta-analyses of the GWAS were conducted in one stage using both inverse-variance weighted fixed effects and random effects meta-analysis ⁹⁹ as implemented in PLINK v1.90b5.4 ^96,97. Meta-analyses were conducted for confirmed only cases, and in the total set including self-report sets (23andMe, Inc. and a portion of UK Biobank).

Conditional and joint analysis of summary GWAS meta-analysis data was performed using Genome-wide Complex Trait Analysis (GCTA, v1.26.0) to identify independently associated variants ³³. To ensure we were only detecting completely independent SNPs the collinearity threshold (--cojo-collinear) was set to R² = 0.05. The threshold for genome-wide significance 5 × 10⁻⁸ and fixed effect meta-analysis p-values and log(OR) effect sizes were analysed.

Linkage-disequilibrium (LD) between SNPs for the conditional and joint analysis of summary data in GCTA (v1.26.0) reported in the manuscript was calculated using a reference population of 5,000 individuals selected randomly from the portion of the UK Biobank population determined to be European by PCA (LD_EUR). Variants were converted to best guess genotype (threshold 0.3). Best guess data were cleaned for missingness > 3%, HWE P < 1 × 10⁻⁶, MAF < 0.001

To limit the chance of false positive claims of novel SNP/loci, we further filtered the list of 77 conditionally independent variants (Supplementary Table 4) to those (i) genome-wide significant (P < 5 × 10⁻⁸) in single SNP and joint conditional analysis, and (ii) as recommended³⁰ where there was evidence of heterogeneity between studies (I² > 31%) the random effect P-value also needed to be < 5 × 10⁻⁸. Passing variants were further checked to ensure that MAFs and effect sizes were consistent across studies and that the result was not driven by a single study (Supplementary Figure 2-3). The 68 retained variants were combined into 54 loci using a concatenating 1 Mb window (Supplementary Table 3). Regional association plots for all 54 loci were interactively plotted by LDassoc (https://ldlink.nci.nih.gov/)¹⁰⁰ and included as Supplementary Data 1.

Multiple testing corrections

The primary aim of our study was to perform a GWAS meta-analysis of cutaneous melanoma risk. For this primary analysis our significance threshold was set at p < 5 × 10⁻⁸. Following this primary analysis, we conducted two classes of secondary analyses: 1) joint analysis of melanoma with a risk phenotype (Nevus or Pigmentation) and 2) TWAS.

To ensure robust adjustment for multiple testing, within the joint cutaneous melanoma-nevus and cutaneous melanoma-pigmentation GWAS analyses we Bonferroni-corrected for each of the two risk factor phenotypes (pigmentation and nevus count), as well as accounting for the two classes of secondary analysis (joint GWAS and TWAS). The resulting significance threshold was (5 × 10⁻⁸)/(2 × 2) = 1.25 × 10⁻⁸. Loci reaching this corrected threshold are indicated in bold in Supplementary Tables 7 and 10.

TWAS was performed on expression data from melanocytes, and then separately on the three skin tissues within GTEx (sun-exposed, not-sun-exposed, and fibroblasts) as these were the most enriched tissues in terms of enrichment for cutaneous melanoma heritability after melanocytes (Extended Data Figure 5) and are likely to be involved in cutaneous melanoma development.

For the melanocyte TWAS analysis, we Bonferroni corrected the significance threshold by the number of tested genes in melanocytes multiplied by the 2 classes of secondary tests and further for the 2 tissue sets; 0.05/(3878 genes × 2 classes × 2 tissue sets) = 3.22 × 10⁻⁶.

For the GTEx skin TWAS analysis we Bonferroni corrected for the total number of tested genes across the tissues multiplied by two classes of secondary tests and further for the 2 tissue sets; 0.05/( ( 8879 + 7458 + 7353 genes) × 2 classes × 2 tissue sets = 5.28 × 10⁻⁷.

The accuracy of p-value calculation for rare SNPs where case/control numbers are imbalanced

The non-normality of the test statistics may cause severely inflated P-values due to violation of asymptotic approximations, particularly for imbalanced case-control ratios. While we addressed this for extreme cases by filtering very rare SNPs (Online Methods), we also investigated whether this could be inflating the P-value of rare SNPs included in the meta-analysis by performing 5 × 10⁸ simulations. For each simulation, we first generated genotype data for 21 studies with the same sample size as in our meta-analysis (Supplementary Table 1) assuming Hardy Weinberg equilibrium for variants with MAF = 0.01.

We then performed association testing for each study and calculated the test statistics to derive an empirical P-value of 6.4 × 10⁻⁸ when using an asymptotic P-value of 5 × 10⁻⁸ as the threshold. While imbalanced case-control ratios had minimal impact on the calculation of asymptotic p-values for SNPs with MAF = 0.01, as the empirical P-value was slightly larger than genome-wide significance we further explored the results of our meta-analysis. Three of our 68 reported variants have a MAF less than 0.01: rs149617956 with MAF = 0.002, rs79356439 with MAF = 0.008 and rs3212371 with MAF = 0.003. All three variants had asymptotic p-values < 5 × 10⁻¹². We performed 5 × 10⁸ simulations for each of the variants using their MAF, and found no simulations had a nominal P-value < 5 × 10⁻¹². These simulations indicate that the actual p-values for these three SNPs are less than 1/(5 × 10⁸) = 2 ×10⁻⁹, and have reached genome-wide significance.

Joint analyses of cutaneous melanoma and nevus count and pigmentation

Nevus GWAS meta-analysis

Using beta meta-analysis weighted by SE as implemented in PLINK 1.90b5.4, we combined the recently published nevus meta-analysis (N = 52,506) ²⁷ which excluded samples with melanoma but may include a small portion of overlap with the controls used for some melanoma GWAS datasets; participants of the QSkin study with nevus count that are non-overlapping and unrelated (IBD pihat < 0.15) to the QSkin melanoma case control set (N = 12,930) and the final set of participants not previously included from the Brisbane Twin Nevus Morphology study (N = 341) ²⁷. The total sample size was 65,777.

Pigmentation GWAS

A GWAS for hair color was performed on 352,662 UK Biobank samples not included in the melanoma GWAS who self-reported having either blonde, light brown, dark brown or black hair (coded as 1, 2, 3 and 4). Hair color was then treated as a continuous variable and regressed on imputed genotype adjusting for principal comments using the same approach as for the melanoma GWAS.

Joint analyses

The melanoma results were then jointly analyzed first with nevus count and then with hair color. Two approaches were taken. Firstly the total confirmed plus self-report cutaneous melanoma GWAS meta-analysis results were combined with the separate nevus and pigmentation GWAS data using Stouffer’s method (P-value weighted by per SNP sample N) as implemented in METAL (version 2011-03-25)¹⁰¹. LD calculations were performed in PLINK using a reference panel of 10,000 white British UK Biobank individuals as implemented in the FUMA platform (v1.3.5)¹⁰² was used to identify independent SNPs with P < 5 × 10⁻⁸; independent SNPs within 1 Mb were considered to be single loci. Secondly, the melanoma and pigmentation/nevus GWAS results were analyzed using GWAS-PW (v0.21)³⁸, which estimates the posterior probability of four possible models for each genetic region: (i) association with cutaneous melanoma only, (ii) association with the second trait only, (iii) association with both traits (pleiotropic), (iv) association with both traits, but co-located and independent (v) no association with either trait. Given that nevus count and pigmentation are believed to act directly on melanoma risk, model (iv) seemed unrealistic so we only considered models (i), (ii), (iii) and (v). For nevus count, SNPs were assigned to blocks using the recommended boundaries for GWAS-PW (https://bitbucket.org/nygcresearch/ldetect-data). For cutaneous melanoma and hair color, 50 SNP windows were used for blocks as the default LD blocks contained multiple independent hair color loci. Following the approach taken by²⁷, any locus with a lead SNP reaching P < 1.25 × 10⁻⁸ for the combined cutaneous melanoma and nevus/hair color analysis and with a posterior probability > 0.5 that the locus is associated with both traits (model 3) to ensure that the association is not driven by a single trait was declared to be pleiotropically associated with both traits.

Analysis of pigmentation and nevi polygenic risk score across melanoma subtypes

For each subject in our study, we calculated two polygenic scores (PRS), using 276 genetic variants associated with pigmentation and 10 genetic variants associated with nevus count. Nevus count SNPs were derived from the same nevus GWAS meta-analysis used for the pleiotropic analysis (N = 65,597), with independent lead SNPs with P < 5 × 10⁻⁸ identified using LD calculations performed in PLINK using a reference panel of 10,000 white British UK Biobank individuals as implemented in the FUMA platform (v1.3.5)¹⁰², with the LD r² cut off for independence < 0.05. Pigmentation PRS SNPs were selected from the hair color GWAS used for the pleiotropic analysis (N= 352,662), with independent lead SNPs with P < 5 × 10⁻⁸ and LD calculations performed in PLINK using a reference panel of 10,000 white British UK Biobank individuals as implemented in the FUMA platform, with the LD r² cut off for independence < 0.025. PRS were calculated for each subject by applying the regression coefficient (from the GWAS of pigmentation or nevus count) to the genotype dosages. We then tested whether PRS distribution differed between males and females, across age groups, and histology subtypes. In total, we performed 27 comparisons and thus any comparison with p-value less than 0.05/27 (=0.00186) was declared as statistically significant.

GENESIS estimation of heritability and polygenic risk

We used GENESIS (https://github.com/yandorazhang/GENESIS)³² (Version 2019-06-01) to estimate the genetic architecture (number of causal SNPs and their effect size distribution) using the summary level statistics from the GWAS meta-analysis. Quantile-quantile plot comparing the p-values generated from this fitted distribution against the observed p-values suggested a three component Gaussian mixture model for the effect size distribution. Based on this estimated genetic architecture, we calculated the heritability at the observational scale and the number of SNPs reaching genome-wide significance for a given GWAS with known sample size. Similarly, GENESIS calculated the AUC for an additive polygenic risk prediction model built based on a discovery GWAS of known sample size.

UK Biobank melanoma risk phenotype GWAS

Four pigmentary GWAS were performed on UK Biobank participants not included in the melanoma GWAS (1) Ease of tanning with 367,229 UK Biobank samples who self-reported their ability to tan as either ‘Get very tanned’, ‘Get moderately tanned’, ‘Get mildly or occasionally tanned’ or ‘Never tan, only burn’ (coded as 1, 2, 3 and 4). Ease of tanning was treated as a continuous variable and regressed on imputed genotypes adjusting for principal components using the same approach as for the melanoma GWAS of UK Biobank data. (2) Skin color with 370,260 UK Biobank samples who self-reported having either ‘Very fair’, ‘Fair’, ‘Light olive’, ‘Dark olive’, ‘Brown’, or ‘Black’ skin color (coded as 1, 2, 3 and 4). Skin color was treated as a continuous variable and regressed on imputed genotype adjusting for principal components using the same approach as for the melanoma GWAS of UK Biobank data. (3) Number of childhood sunburns with 320,345 UK Biobank samples who self-reported their sunburn incidents pre-sixteen years old. The data were dichotomized into none and at least one pre-sixteen sunburn incident categories (coded as 1, 2). Number of childhood sunburns was treated as a binary variable and regressed using a logistic model on imputed genotype adjusting for principal comments using the same approach as for the melanoma GWAS of UK Biobank data. (4) Red hair with 120,925 UK Biobank samples who self-reported having either ‘red hair’ or other (coded as 1 or 2). Red hair was treated as a binary variable and regressed using a logistic model on imputed genotype adjusting for principal comments using the same approach as for the melanoma GWAS of UK Biobank data.

Linkage disequilibrium (LD) score regression

As LD score regression (LDSC) is sensitive to the quality of input SNPs, GWAS or meta-analysis variants were filtered to the list of high quality HapMap SNPs provided¹⁰³. Using LD Score regression v1.0.0 genomic inflation (Lambda), intercept and SNP-heritability (h²) were estimated. h² estimates were converted to the liability scale using the lifetime population prevalence for cutaneous melanoma in Australia (0.0588)¹⁰⁴.

LD score regression of tissue-specific genes

Cutaneous melanoma heritability enrichment for SNPs around tissue-specific genes was assessed by stratified LD score regression as described previously ^28,42 and implemented in the LDSC program v1.0.0 (https://github.com/bulik/ldsc). Briefly, RNA-seq data for all 50 GTEx (v7) tissue types and primary melanocyte were quantified as RPKM using RNA-SeQC (v1.18) ¹⁰⁵ and quantile normalized to reduce batch effect. Tissue-specific genes were defined by calculating the t-statistic of each gene for a given tissue, excluding all samples from the same tissue category. Tissue category assignment for GTEx tissue types was based on the previous publications ^28,106, and melanocytes were defined as “skin” category together with two types of skin and transformed skin fibroblasts from the GTEx. We selected the top 1,000, 2,000, and 4,000 tissue-specific genes from the t-statistic analysis, and added 100 Kb each to the transcription start site and transcription end site to define tissue-specific genes annotation. Stratified LD score regression was then applied on a joint SNP annotation to estimate the heritability enrichment against the total cutaneous melanoma GWAS data from the current study.

Colocalization of cutaneous melanoma GWAS and eQTLs

We performed colocalization analyses of cutaneous melanoma GWAS signals with eQTL signals from our melanocyte and 48 GTEx (v7) tissue eQTL datasets (note that 2 tissue types that were included for LDSC using expression data were not included here as well as in TWAS analyses due to lack of eQTL data from GTEx), using eQTL and GWAS CAusal Variants Identification in Associated Regions (eCAVIAR, v2.0, http://genetics.cs.ucla.edu/caviar/, https://github.com/fhormoz/caviar)⁴³. Consistent with the previous study, we used 50 SNPs upstream and downstream of each cutaneous melanoma GWAS lead SNP to extract both GWAS and eQTL summary statistics to be used as the input for eCAVIAR analysis. The LD matrix was calculated using the unphased 1000 Genomes reference set. For the CLPP score calculation, we allowed a maximum number of two causal SNPs in each locus. For a given cutaneous melanoma GWAS locus, an eGene with a CLPP score above 1% (0.01) was considered to display a positive co-localization. To avoid reporting spurious effects, we applied a conservative criterion and only reported variants displaying LD r² > 0.9 with the cutaneous melanoma GWAS lead SNP and eQTL P-value below a Bonferroni-corrected cutoff of each dataset (0.05/number of eGenes tested for each tissue dataset).

TWAS

We performed transcriptome-wide association studies (TWAS) for the cutaneous melanoma GWAS meta-analysis data using TWAS/FUSION (http://gusevlab.org/projects/fusion/) as previously described^28,41. TWAS was performed in three separate groups, using eQTL datasets from 1) melanocytes, 2) three skin tissues (sun-exposed, not-sun-exposed, and fibroblasts) within GTEx (V7), and 3) the rest of GTEx tissue types (a total of 45) by imputing the gene expression phenotypes for the total cutaneous melanoma GWAS meta-analysis data. The analysis parameters were set to allow for multiple prediction models, independent reference LD, additional feature statistics and cross-validation results⁴¹. The total cutaneous melanoma GWAS meta-analysis summary statistics were included with no significance thresholding. For GTEx data, we downloaded the precomputed expression reference weights for GTEx gene expression (v7) RNA-seq across 48 tissue types from the TWAS/FUSION website (http://gusevlab.org/projects/fusion/). We computed functional weights from the primary melanocyte RNA-seq data²⁸ one gene at a time. Genes that failed quality control during the heritability check (using minimum heritability P-value 0.01) were excluded from further analyses. We restricted the cis-locus to 500 Kb on either side of the gene boundary.

Data Availability

Genome-wide summary statistics for the confirmed meta-analysis have been made publicly available at dbGaP (phs001868.v1.p1), with the exclusion of genomic data from 23andMe. Results for SNPs with a fixed or random P < 5 × 10⁻⁷, from the total meta-analysis are reported in Supplementary Table 7. The total meta-analysis includes self-report cutaneous melanoma GWAS data from the UK Biobank and 23andMe. The raw genetic and phenotypic UK Biobank data used in this study, which were used under license, are available from: http://www.ukbiobank.ac.uk/. The genome-wide summary statistics from 23andMe, Inc. data were obtained under a data transfer agreement. Further information about obtaining access to the 23and Me, Inc. summary statistics are available from: https://research.23andme.com/collaborate/.

Extended Data

Extended Data Fig. 1. Quantile-Quantile plot of total CM meta-analysis.

Quantile-quantile plots of negative log₁₀ two sided P-value derived from a fixed-effects inverse-variance weighted meta-analysis of log(OR) effect-sizes derived from the logistic regression GWAS listed in Supplementary Table 1. All confirmed and self-report cases are included, with a total sample size of 36,760 melanoma cases and 375,188 controls.

Extended Data Fig. 2. Manhattan plots of melanoma risk loci from total and confirmed-only GWAS-meta-analyses.

Negative log₁₀ two sided P-value derived from a fixed-effects inverse-variance weighted meta-analysis of log(OR) effect-sizes derived from the logistic regression GWAS (y-axis) are plotted by their chromosome position. The confirmed-only analysis included 30,134 cases with histopathologically confirmed CM, and 81,415 controls. The total CM meta-analysis includes all confirmed and self-report cases, with a total sample size of 36,760 CM cases and 375,188 controls. Multiple-testing corrected genome-wide significance threshold was P<5×10⁻⁸. We display in order the total CM meta-analysis without limiting the y-axis; the pathologically confirmed CM cases only meta-analysis with the y-axis limited to 1×10⁻²⁵ and without a limit to more clearly display loci other than MC1R.

Extended Data Fig. 3. Quantile-Quantile plot of confirmed-only CM meta-analysis.

Quantile-quantile plots of negative log₁₀ two sided P-value derived from a fixed-effects inverse-variance weighted meta-analysis of log(OR) effect-sizes derived from the logistic regression GWAS listed in Supplementary Table 1. Only cases with histopathologically confirmed CM are included, with a total sample size of 30,134 melanoma cases and 81,415 controls.

Extended Data Fig. 4. Distribution of pigmentation polygenic risk scores across melanoma histological subtypes.

The figure shows whether PRS defined based on SNPs associated with hair colour differ across CM histological types (Online Methods; SSM: superficial spreading melanoma; NM: nodular melanoma; LM: lentigo melanoma; Acral: acral lentiginous melanoma). The higher the PRS the lighter the hair colour. When comparing subtype 1 vs. subtype 2, we report the effect size for the linear regression of PRS on subtype 1, including study and principal components as covariates to control for population stratification. The regression coefficient, 95% confidence interval, and statistical significance are shown. The positive beta indicates the PRS is higher in subtype 2 (e.g., non-acral melanomas). This analysis included 9828 SSM, 2137 NM, 900 LM, 353 acral melanoma cases and 44676 controls. Two-sided t-statistic was used for testing significance. P-values reported were not adjusted for multiple comparison.

Extended Data Fig. 5. LD score regression plots.

LD score regression was performed for the top 4000 (A) 2000 (B) and 1000 (C) tissue-specific genes from melanocyte and GTEx tissue types (v7 datasets), to assess the enrichment of melanoma heritability in these genomic regions using summary statistics from Total CM GWAS meta-analysis. The level of enrichment and P-values are shown, with an FDR = 0.05 cutoff marked as a dashed horizontal line (See Online Methods for statistical test). Tissue categories are color-coded, and a subset of top individual tissue types are shown on the plot. Tissue types from “Skin” category including melanocytes are highlighted in magenta.

Extended Data Fig. 6. Effect sizes for confirmed-only meta-analysis versus UKBB self-report set.

For each independent genome-wide significant (P<5×10⁻⁸) lead SNP from the confirmed only meta-analysis (30,134 melanoma cases and 81,415 controls), we plot on the Y-axis UK Biobank self-report GWAS (UKBB SR) log(OR) and standard error from a logistic regression GWAS (1,802 self-report CM cases and 7,208 controls) and on the X-axis we plot the log(OR) and standard error from a fixed-effects inverse-variance weighted meta-analysis of log(OR) effect-sizes derived from the logistic regression GWAS for confirmed melanoma cases listed in Supplementary Table 1. We also report the r² correlation from the linear regression of UKBB SR log(OR) on the confirmed met-analysis estimates, weighted by their standard error.