Functional characterization of the 9q34.13 locus identifies RAPGEF1 as a candidate gene modulating risk for melanoma and nevi via RAS activation

Genetic fine-mapping at the melanoma risk locus on chromosome band 9q34.13

We fine-mapped the melanoma risk locus on chromosome band 9q34.13 (a ±500 Kb window surrounding rs3780269) to identify potential causal variants. We applied multiple Bayesian fine-mapping methods including SuSiE (Sum of Single Effects) ^25,26, DAP-G ^27,28 (Deterministic Approximation of the Posteriors), and RSparsePro ²⁹, which is robust to mismatching LD between the GWAS and the LD-reference panels. We used default parameters for each method. The min_abs_corr parameter was set to 0 for SuSiE.

The summary data for 2020 melanoma GWAS by Landi and colleagues¹ were obtained from dbGaP (dbGaP: phs001868.v1.p1), with the exclusion of self-reported data from 23andMe, Inc. and UK Biobank. The full GWAS summary statistics for the 23andMe discovery dataset will be made available through 23andMe to qualified researchers under an agreement with 23andMe that protects the privacy of the 23andMe participants. Please visit https://research.23andme.com/collaborate/#dataset-access/ for more information and to apply to access the data. Summary data from the remaining self-reported cases are available from the corresponding authors of that manuscript (Matthew Law, matthew.law@qimrberghofer.edu.au; Mark Iles, m.m.iles@leeds.ac.uk; and Maria Teresa Landi, landim@mail.nih.gov).

Melanoma GWAS Z-score values for each variant were from the fixed-effect inverse variance weighted meta-analysis of the full set of confirmed and self-reported individuals diagnosed with melanoma and controls, as described previously by Landi and colleagues ¹. We used a pre-computed LD matrix of 5,000 UK Biobank individuals as described previously ¹. While we allowed for a maximum of five credible sets during fine-mapping, we only retained individual credible sets where the cumulative posterior inclusion probability was > 0.85, and retained all variants in the credible set where the individual variant posterior inclusion probability was > 0.005. To inclusively fine-map variants, we also independently selected any variants with log likelihood ratio (LLR 1:1000) relative to the GWAS lead variant rs3780269. Finally, to assess variants that were not successfully genotyped or imputed in the GWAS (including notably insertion/deletion variants that were not a part of the Haplotype Reference Consortium imputation panel), we included any such variants that had an LD r² > 0.8 (1000 Genomes Project, Phase 3, Version 5, EUR population) ³⁰ with rs3780269 as identified by the LDlinkR package ^31-33.

To further assess the possibility of other potential independent signals in the region, we conditioned on the lead GWAS variant (rs3780269) using the Genome-wide Complex Trait Analysis (GCTA, v1.94.1) ³⁴ COnditional and JOint association (COJO) module ³⁵, using default settings, a genomic window of ±1 Mb from the lead variant, and the UKBB LD matrix described above.

Annotation of fine-mapped variants

We assessed all fine-mapped variants at this locus for overlap with melanocytic regulatory elements (defined by chromatin states, or accessibility) using the BedtoolsR package (intersect function) ^36,37 and epigenomic datasets from melanocytes and melanoma cell line models. We classified genomic regions as enhancers or promoters using the Roadmap project chromHMM data from two primary melanocyte cultures ^17,20. Using this data, enhancers were defined using the following chromHMM states from all chromHMM tracks (primary: Enh, EnhG, EnhBiv; auxiliary: EnhG1, EnhG2, EnhA1, EnhA2, EnhWk, EnhBiv; and imputed: TxEnh5, TxEnh3, TxEnhW, EnhA1, EnhA2, EnhAF, EnhW1, EnhW2, EnhAc). Melanoma-specific enhancer annotations included 4_EnhA, 5_EnhM, 6_EnhW, 7_TxEnhM, 7_TxEnhW, and 9_TxWkEnhW, from a tumorigenic melanoma cell model ²¹. Promoter states were defined in melanocytes from imputed ChromHMM as PromU, PromD1, PromD2, TssA, PromP, PromBiv, and Tx_Reg; and in melanoma cells as 1_TssA, 2_PromWkD, and 3_TssWkP. We analyzed our previously generated ATAC-seq data from five independent melanocyte cultures ²², along with publicly available omni-ATAC-seq data from nine melanoma cell cultures ¹⁹. Any region annotated as enhancer or promoter or that overlapped ATAC-peak in either melanocyte or melanoma cells was considered cis-regulatory. Melanocyte ATAC-seq are publicly available (ArrayExpress accession: E-MTAB-15048).

Allele-specific transcriptional activity of these variants was previously assessed using Massively Parallel Reporter Assays (MPRA) in immortalized primary melanocytes and UACC903 melanoma cells ²³. All fine-mapped variants for the melanoma risk signal were included in the MPRA library design, however one failed data QC and was subsequently excluded from analysis. Allelic impacts on transcriptional activity were assessed using a linear regression model, accounting for strand and transfection batch as covariates. A scrambled 145-bp sequence was used as a negative control and as a baseline to assess relative enhancer activity of specific regions surrounding fine-mapped variants. Allele-specific differences for each variant were tested relative to scrambled controls. Multiple hypothesis testing was corrected using Benjamini-Hochberg false discovery rate. MPRA data are previously published ²³; raw sequencing can be accessed through NCBI Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) under the accession GEO: GSE210356.

We assessed the impact of fine-mapped variants on the protein-coding sequence Variant effect predictor tool (https://grch37.ensembl.org/Homo_sapiens/Tools/VEP) ³⁸ using GENCODE version 19 protein-coding transcripts.

Cell culture

Frozen aliquots of melanocytes isolated from foreskin of healthy newborn males, mainly of European descent, were obtained from the Specialized Programs of Research Excellence (SPORE) in Skin Cancer Specimen Resource Core at Yale University following an established protocol ³⁹. Melanocytes were grown in Dermal Cell Basal Medium (American Type Culture Collection/ATCC PCS-200-030) supplemented with Melanocyte Growth Kit (ATCC PCS-200-041) and 1% amphotericin B/penicillin/streptomycin (120-096-711, Quality Biological) at 37°C with 5% CO₂. Immortalized C283T ¹⁸ melanocytes as well as C283T cells stably expressing dCas9-KRAB ²² were generated as previously described and grown in identical conditions to primary melanocytes. All cells tested negative for mycoplasma contamination via MycoAlert PLUS Mycoplasma Detection Kit (LT07-710, Lonza). Use of these samples was reviewed and approved by the National Institutes of Health Office of Human Subjects Research Protections (OHSRP).

Assessing micro-RNA binding for variant overlapping RAPGEF1 3’-UTR region

To assess whether melanoma GWAS risk variant rs3739497, which resides in the RAPGEF1 3′ untranslated region (UTR), is involved in potential post-transcriptional regulation of RAPGEF1 through micro-RNAs (miRNAs), we assessed predicted miRNA binding sites near this variant. We queried several databases, including TargetScan ⁴⁰, miRTarBase ⁴¹, miRDB ^42,43, and mirCode ⁴⁴. Binding sites that overlapped or were located within ±50 nucleotides of the variant position were considered candidate variant-dependent miRNA sites. Finally, to explore post-transcriptional regulation via sequence elements other than miRNAs, we queried the 3'UTR variant region against the AREsite database ⁴⁵ that provides a genome-wide catalogue of AU-rich elements.

Small RNA sequencing data from primary melanocytes and melanomas

To assess small RNA levels in primary human melanocytes (n=106), total RNA, including microRNA (miRNA), was purified using Qiagen miRNeasy kit (#217004). Briefly, cells were lyzed in Trizol and heated at 65C followed by chloroform extraction. The aqueous phase was further purified using a miRNeasy column following manufacturer’s instruction. miRNA sequencing libraries were constructed using NEB Next Small RNA Library Prep kit (E7330s) and the sequencing was performed on HiSeq2500. All samples show the proportion of Q30 or higher bases of at least 93%. 28-79 million reads were achieved for each sample with overall alignment percentage between 67-85%. Sequencing reads were processed using the mirMaster pipeline ⁴⁶. Quality control stage processing steps included removal of 3′ adapter sequences (AGATCGGAAGAGCACACGTCTGAACTCCAGTCAC), trimming of low-quality bases, and retention of reads that exceeded the default minimum length thresholds. Reads were then aligned to the human genome (GRCh38) using Bowtie ⁴⁷, allowing up to five genomic alignments per read. miRNAs were quantified by aligning sequencing reads against the mature and precursor miRNA reference sequences from miRBase v22.1 ^48-53. Alignments permitted up to one mismatch. Expression values were normalized across samples using quantile normalization. For each miRNA, percentile rank of expression across the dataset was calculated based on median expression values.

RAPGEF1 mRNA expression levels were available from the bulk RNA-seq data we previously simultaneously generated from the same 106 melanocyte cultures ¹³. The correlation of RAPGEF1 mRNA expression levels with candidate miRNAs was measured using Spearman’s rank correlation.

Expression levels of micro-RNAs in TCGA melanoma samples ¹⁴ were accessed through FirebrowseR (https://github.com/mxxdxxx/FirebrowseR).

Quantitative Trait Locus (QTL) analysis of melanocytic cells

We assessed FDR significant QTL signals near the GWAS risk locus using previously-published melanocyte and melanoma datasets [9, 12]. Melanocyte expression QTL (eQTL), splice QTL (sQTL), and methylation QTL (meQTL) datasets were generated from 106 primary melanocyte cultures derived from individuals of European descent. Methylation probes from meQTL data were assigned to genes based on CpG location within 1.5 kb of the TSS, 5′-UTR, 1st exon, gene body, or 3′-UTR of a gene. Melanocyte genotyping, RNA-sequencing, and methylation data are previously published ^13,16 and accessible through NCBI database of Genotypes and Phenotypes (dbGaP) under accession dbGaP: phs001500.v2.p1.

Tumor eQTL data were derived from TCGA melanoma primary tumors as a part of the PancanQTLv2.0 eQTL study; only data from nominally significant SNP-gene pairs (P<0.01) were made available [79] (https://hanlaboratory.com/PancanQTLv2/).

GWAS colocalization using the melanocyte eQTL dataset was performed using the SuSiE + Coloc ⁵⁴ and SharePro tools ⁵⁵. SuSiE + Coloc performs fine-mapping of GWAS and QTL signals via SuSiE, followed by colocalization via the Coloc tool. SharePro leverages linkage disequilibrium (LD) to cluster correlated variants into effect groups, followed by colocalization. GWAS summary data, and the RAPGEF1 and UCK1 melanocyte eQTL summary data (± 500 Kb around the lead variant) were used as input for this analysis. Default parameters were used except the sigma parameter, which represents the prior probability of colocalization between traits. Since the region being assessed harbors multiple GWAS/QTL signals with complex LD, we provide results using multiple sigma values, ranging from most stringent (1 x 10⁻⁵) to less stringent (1 x 10⁻⁴, 1 x 10⁻³). However given that relaxing this value from the default of 1 x 10⁻⁵ increases the potential for false positive findings, we formally tested colocalization using this default value. Colocalization probability > 0.8 was considered strong evidence supporting colocalization, while a colocalization probability < 0.2 was considered evidence against colocalization ⁵⁵.

We also used these QTL datasets to perform transcriptome (melanocyte and TCGA melanoma tumors) and methylome-wide (melanocyte) association studies (TWAS/MWAS). To identify conditionally independent loci, we performed conditional TWAS analysis using the FUSION tool ⁵⁶.

Capture-HiC assay for linking risk variants to target genes

As described previously ²², we performed melanoma GWAS-region focused capture Hi-C assays in primary human melanocytes to facilitate variant-to-gene mapping across all 68 genome-wide significant signals for melanoma. Capture probes were designed by Arima Genomics (San Diego, CA) using 2x tiling with least stringent masking and XTHSBoosting, resulting in an Agilent SureSelect library (Santa Clara, CA) targeting all restriction fragments generated by DpnII and MboI (recognition sites: GATC, GANTC) within each association interval. We generated 15 capture Hi-C libraries from five independent primary melanocyte cultures (C56, C140, C205, C24, and C27), with three technical replicates prepared per culture. Barcoded libraries were pooled and sequenced on an Illumina NovaSeq platform across one SP and one S1 flow cell, yielding ~5.7 billion 150 bp paired-end reads in total. Sequencing reads were processed using the HiCUP pipeline ⁵⁷ with alignment to the hg19 reference genome via Bowtie2 ⁵⁸. Technical replicates from each melanocyte culture were combined prior to downstream analyses. Chromatin interactions were detected using the CHiCAGO pipeline (v1.16.0) ⁵⁹ at both individual restriction fragment-level and aggregated four-fragment resolution, as previously described ²². Interactions with CHiCAGO scores >5 were considered high-confidence visualized on the WashU ^60,61 and UCSC genome browsers ^62-66. To mitigate cross-linking variability and restriction fragment boundary effects, we expanded each fine-mapped variant into a ±500 bp window and evaluated interactions from both the variant-overlapping and adjacent fragments. Genes were nominated as candidate targets if a chromatin interaction was observed from a variant-harboring (or adjacent) restriction fragment to the target gene promoter. To map cis-regulatory targets of enhancer-promoter interactions, candidate genes initially nominated based solely on physical chromatin interactions with fine-mapped variants were further refined by evaluating whether these variants overlapped melanocytic regulatory regions. Melanocyte capture-HiC data are publicly available (ArrayExpress accession: E-MTAB-15048).

CRISPRi validation of regulation of RAPGEF1 and UCK1 transcription

To validate regulatory interactions from risk variants to target genes, we performed CRISPR interference (CRISPRi) assays in the immortalized human melanocyte cell line C283T stably expressing dCas9-KRAB ²². Briefly, C283T cells were infected with a lentiviral vector pLX_311-KRAB-dCas9 (gift from John Doench, William Hahn, and David Root; Addgene plasmid # 96918; http://n2t.net/addgene:96918; RRID: Addgene_96918) followed by monoclonal cell selection. We confirmed stable expression and functional activity of dCas9-KRAB in the clone used for CRISPRi.

Three different guide RNAs (gRNAs) were designed to target the genomic regions flanking each variant site. The gRNA sequences were located ±50bps from the variant (Table S1). We used two non-targeting gRNAs (NTC1, NTC2) as controls. gRNAs were ligated into the lentiviral vector pXPR-050 (gift from John Doench and David Root, Addgene plasmid #96925; RRID: Addgene_96925). C283T immortalized melanocytes ¹⁸ were infected with lentiviral particles encoding gRNA, and at 24h, 1.5 μg/mL of puromycin was added for selection. After two days of puromycin selection, puromycin was removed, and cells were harvested for RNA collection on the same day or one day after puromycin removal. Total RNA was isolated with RNeasy Mini Kit (Qiagen) and cDNA was generated with SuperScript IV VILO Master Mix (Thermo Scientific). Multiple infections were performed with total ten biological replicates for each gRNA against the 3 different variants. mRNA levels of either RAPGEF1 or UCK1 were measured by TaqMan assay (Thermo Scientific; Table S1) and normalized to GAPDH levels. qPCR triplicates (technical replicates) were averaged and subsequently considered as a single data point. Average 2^^(delta-Ct) values for NTC1 and NTC2 were used for statistical comparisons to other gRNAs, data in tables and graphs are represented as fold-change relative to the average of NTC1 and NTC2. The statistical analysis was performed using a paired two-tailed t-test comparing 2^^(delta-Ct) values of individual guides to the average 2^^(delta-Ct) of NTC1/NTC2.

Knock-down of RAPGEF1 and UCK1 expression by shRNA and qRT-PCR quantification of gene expression

Pre-designed shRNAs against human RAPGEF1 and UCK1 were purchased from Sigma (TRCN0000048129 and TRCN0000048130 for RAPGEF1, TRCN0000037763 and TRCN0000199254 for UCK1; Table S1) and expressed in lentiviral vector backbone (pLKO). After viral infection and puromycin selection, total RNA was isolated with the RNeasy Mini Kit (Qiagen) and cDNA was generated with SuperScript IV VILO Master Mix (Thermo Scientific). Transcript levels were measured using TaqMan qPCR assays purchased from Thermo Scientific. RAPGEF1 and UCK1 transcript levels were normalized to the levels of GAPDH, and PCR triplicates were averaged and considered as one data point.

Overexpression of RAPGEF1 in human melanocyte culture

RAPGEF1 was cloned from a cDNA construct (OHS5893-202500127, ORFeome Collab, Horizon Discovery) into pLX304 (a gift from David Root, Addgene plasmid # 25890; http://n2t.net/addgene:25890) or pCW57.1 (a gift from David Root, Addgene plasmid # 41393; http://n2t.net/addgene:41393) lentiviral vectors. This clone is the isoform most strongly expressed across neonatal melanocytes ¹³. pCW57.1 contains a doxycycline inducible promoter. Lentiviral vectors were co-transfected into HEK293 cells with the pSPAX2, pMD2-G, and pCAG4-RTR2 packaging vectors. Virus was collected two days after transfection and concentrated by Vivaspin Concentrators (76409-072, VWR). Cells were incubated with virus for 24 h, followed by selection with blasticidin (5-10ug/ml) or puromycin (1–2 μg/ml) for 2–4 days. For pCW57.1-RAPGEF1 infected cells, RAPGEF1 expression was induced by 0.2ug/ml doxycycline in melanocyte culture.

Gene-based CRISPR-Cas9 knockout screen

We assessed whether RAPGEF1 was prioritized in the gene set identified as critical for melanocyte cell growth and/or survival using data from a previously described gene-based pooled CRISPR-Cas9 cell proliferation knockout screen in an immortalized human melanocyte culture, C283T, as described by Xu and colleagues ¹⁸.

Cell proliferation assays

Cell proliferation was assayed using a BrdU flow kit (BD Pharmingen) according to the manufacturer's protocol. Briefly, cells were labeled with 10 μM BrdU for 1-3 hours before they were fixed, permeabilized, and subjected to DNase I treatment. Cells were then stained with FITC-conjugated antibody to BrdU and 7-AAD, followed by flow cytometry analysis using an AttuneNxT Cytometer (Thermo Scientific). For crystal violet staining, cells were seeded at equal numbers after infection and drug selection and were stained with crystal violet at different days after seeding. Crystal Violet was solubilized by methanol for quantification as absorbance at 560nm.

Immunoblotting

For immunoblot analysis, total cell lysates were generated with RIPA buffer (Thermo Scientific) and subjected to water bath sonication. Samples were resolved by 4–12% Bis-Tris ready gel (Invitrogen) electrophoresis. The primary antibodies used were mouse antibody to RAPGEF1 (MA5-26281, ThermoScientific), mouse antibody to RAS (05-516, Sigma), rabbit antibody to RAP1 (ab181858, Abcam), mouse antibody to phospho-ERK(9106S, Cell Signaling Technology), mouse antibody to Pan-AKT (2920s, Cell Signaling Technology), rabbit antibody to total ERK(4695s, Cell Signaling Technology), rabbit antibody to Phospho-AKT(4060s, Cell Signaling Technology), and mouse antibody to β-actin (A5316, Sigma).

Colony formation assays

Anchorage-independent growth was assayed with p'mel/BRAF p.Val600Glu cells (provided by H. Widlund, Dana-Farber Cancer Institute). Briefly, 5,000 cells were mixed in HAMF12 medium supplemented with 10% FBS containing 0.4% SeaKem LE agarose (Lonza) and plated on top of a bottom layer with 0.65% agarose prepared in the same HAMF12 medium in a six-well plate. After agarose is solidified, each well was covered with 0.5 ml of feeding medium, which was refreshed twice a week. Colonies were counted and imaged 4-6 weeks after seeding using Lionheart (Agilent) with total magnification of 4x. Colonies were counted from six wells for each condition, and significant differences between samples were assessed by unpaired two-tailed t-test assuming equal variance.

Zebrafish Melanoma Generation and Analysis

Zebrafish melanomas were generated by overexpressing human RAPGEF1 (cloned from a cDNA construct, OHS5893-202500127, ORFeome Collab, Horizon Discovery; same RAPGEF1 isoform expressed in human melanocyte experiments) or an empty control vector (MCS) in transformed melanocytes using the miniCoopR system ⁶⁷. This system involves injection of a Tol2-based miniCoopR vector into the Tg(mitfa:BRAF p.Val600Glu); p53(lf); mitfa(lf) zebrafish strain, where it rescues melanocyte development and simultaneously drives expression of a gene of interest under the mitfa promoter. To visualize melanocytes, we coinjected mitfa:Cas9, CRISPR gRNAs targeting tyrosinase (to block pigmentation), and mitfa:mCherry. As melanocyte rescue is mosaic in this system, we screened these cohorts for fish that had mitfa:mCherry melanocyte expression in either the RAPGEF1 cohort (n=35), or the MCS control cohort (n=48). These cohorts were comprised of two separate clutches per condition and were genotyped to confirm RAPGEF1 expression status. Starting at 13 weeks post-fertilization, we monitored fish for precursor lesions—termed “cancer precursor zones” (CPZ)—defined as patches of merged mitfa:mCherry-high melanocytes exceeding established size and fluorescence thresholds. Screening continued every few weeks until 46 weeks post-fertilization. Kaplan-Meier survival analysis and log-rank (Mantel-Cox) tests showed no significant differences in CPZ onset (p=0.1119) or tumor onset (p=0.211) between groups. Tumor anatomical distribution did not differ substantially between groups.

RAPGEF1 functions upstream of BRAF in the MAPK pathway, and since our primary zebrafish tumor model overexpresses oncogenic BRAF p.Val600Glu, this could obscure a role for RAPGEF1 or represent a different biological context to that seen in vitro. To assess this, we used the MAZERATI system (Modeling Approach in Zebrafish for Rapid Tumor Initiation) ⁶⁸ to generate zebrafish cohorts where rescued melanocytes expressed either BRAF p.Val600Glu/p53(lf) or RAPGEF1/p53(lf). In this system, mitfa:BRAF p.Val600Glu and p53(lf) mutations are not stably expressed in the zebrafish strain background, but are delivered via plasmid injection at the one-cell stage, allowing us to test whether RAPGEF1 alone can drive tumor initiation in the absence of an activating BRAF mutation. While the BRAF p.Val600Glu/p53(lf) cohort developed tumors at the expected rate, RAPGEF1/p53(lf)– fish failed to develop precursor lesions or tumors.

RAP1 and RAS activation assay

RAP1 activation was measured using the RAP1 activation assay kit from Abcam (ab212011) and RAS activation was measured using GST-fusion Raf-1 RBD beads (Sigma, #14863) or RAF-RBD Protein GST Beads (RF02-A, Cytoskeleton). Cells were grown in supplement-free ATCC medium for 24 hours before treated with 50ng/ml human EGF protein (R&D, Cat# 236-EG-200) at the indicated time. Whole cell extract was taken to be pulled down by the Raf-RBD beads for RAS activation assay or by the RalGDA RBD beads from the Abcam kit for measuring RAP1 activation. Beads were then lysed in SDS loading buffer, resolved by 4–12% Bis-Tris ready gel (Invitrogen) electrophoresis, and blot with RAS (Sigma) or RAP1 (Abcam) antibody. Only the active GTP-bound form of RAS or RAP1 can be pulled down by the beads and the intensity of RAS or RAP1 activation was determined by comparing the pulled-down signal with that from the whole cell extract.

Tumor cohort descriptions

TCGA Melanoma Cohort (Table S2): We assessed the extended dataset from the Cancer Genome Atlas (TCGA) skin cutaneous melanoma study ¹⁴, comprising 439 individual samples with matching mutation and expression profiling data accessed via the cBioPortal database ^69-71, Skin Cutaneous Melanoma (TCGA, PanCancer Atlas) dataset (https://www.cbioportal.org/study/summary?id=skcm_tcga_pan_can_atlas_2018). This extended dataset was included as part of the PanCancer Atlas analyses ^72,73. Mutation profiling was performed using whole-exome sequencing, and expression profiling was performed using RNA-sequencing. Four tumor purity measures were available ⁷⁴. Only a small subset of samples in this dataset are primary tumors (n=76), whereas the majority represented metastatic tumors. In our analyses, we included only samples labeled as metastases (n=363); samples annotated as primary tumors were excluded. Among the remaining samples, 21 individuals received ipilimumab, 2 individuals received pembrolizumab, 1 individual received both ipilimumab and pembrolizumab, and 1 individual received nivolumab; mutation and gene expression profiles from the samples from these individuals were generated from tumor samples collected prior to immunotherapy treatment. The vast majority of other samples were collected prior to treatment (treatments across these samples were varied), however a few (n=13) were collected following some form of treatment. Age at diagnosis, sex, tumor site-based staging, and all four estimates of tumor purity were used as covariates.

Assessing RAS-MAPK alterations across multiple melanoma cohorts

We assessed RAPGEF1 expression association with activating mutations in the RAS-MAPK signaling pathway in melanoma tumors. We analyzed cutaneous melanoma tumors across six cohorts: TCGA n=363 ¹⁴, Leeds n=503 ^75-77, Lund n=146 ⁸⁰, Liu n=84 ⁸³, Van Allen n=38 ⁸⁴, Riaz n=39 ⁸⁶ (Tables S2-8). Briefly, the Leeds, Lund, and TCGA cohorts included very few samples treated with immunotherapy (TCGA n=23 ⁸⁷; Leeds n=12 ⁷⁵; and Lund n=4 ⁸⁰), while the other three cohorts consisted of individuals treated with immunotherapy following tumor collection. Our sample inclusion criteria required melanoma samples that had both mutation and gene expression data available across the cohorts.

Tumor driver mutation definitions

The driver annotations for tumor alterations were available on the cBioportal database for TCGA melanoma, as well as the Liu and Van Allen cohorts, which utilizes information from OncoKB ^88,89 and Cancer HotSpots ^90,91 databases for annotating an alteration as driver. Where available, gene fusion events classified as drivers were included in the mutant subgroup classifications. For the remaining three other cohorts, the alterations were defined as driver mutations for oncogenes via manual look up on the OncoKB and Cancer HotSpots databases. For tumor suppressor genes such as NF1 across these cohorts, we used the ANNOVAR ⁹² tool for annotating variant predicted effect across 11 prediction algorithms, SIFT ⁹³, PolyPhen2 ⁹⁴, Mutation Assessor ⁹⁵, SVM radial, LR predict, MetaSVM and MetaLR ⁹⁶, PROVEAN ⁹⁷, PrimateAI ⁹⁸, AlphaMissense ⁹⁹, and CLNSIG ¹⁰⁰. An alteration was considered as a potential driver if predicted pathogenic or likely pathogenic by the majority of these tools. This integrative approach combined expert curated driver annotations with computational predictions to provide comprehensive information for classification of gene alterations.

Assessing RAPGEF1 expression association with RAS-pathway

We tested several models classifying RAS-pathway wild-type and mutant tumors, focusing initially on the most commonly-mutated oncogenic driver genes and subsequently adding in less-commonly altered genes. Specifically, within the Leeds sample set (n = 503 total), only specific mutations in BRAF (affecting Val600∣Lys601) and NRAS (affecting Gln61∣Gly12∣Gly13) were assessed in a subset of the tumors (n=234), while for the remaining a full set of putative MAPK driver mutations ¹⁰¹ were assessable from sequencing data, so our initial model (Model 1) defined the mutant subgroup as those samples with these specific hotspot mutations leading to BRAF (Val600∣Lys601) and NRAS (Gln61∣Gly12∣Gly13). Additional models were as follows: Model 2 mutant group, any driver mutation in BRAF or NRAS; Model 3 mutant group, any driver mutation in BRAF, NRAS, KRAS, or HRAS; Model 4 mutant group, any driver mutation in BRAF, NRAS, KRAS, HRAS, or NF1 (“triple wild-type); and Model 5, any putative driver mutation in the 44 genes defined from the TCGA Pan-Cancer analysis ¹⁰¹ (Table S9). Outside of the Leeds dataset, all cohorts had full mutation information for BRAF, NRAS, KRAS, HRAS, and NF1. As described above, Leeds and Lund cohorts had targeted sequencing of select cancer-related genes and were not fully assessed for all genes included in Model 5. In this case for Model 5, we classified wild-type and mutant subgroups based on the available gene sequencing data from each dataset. Methods for annotating individual mutations as drivers are described in the above paragraph. We assessed an association of RAPGEF1 levels with RAS-MAPK wild-type and mutant subgroups using a linear regression model, adjusting for age, sex, tumor purity, and tumor stage when available.

Survival analysis

We conducted survival analyses across multiple melanoma cohorts using the Cox proportional hazards models to assess the association between RAPGEF1 expression and clinical outcomes in tumors wild-type for BRAF (Val600∣Lys601) and NRAS (Gln61∣Gly12∣Gly13) hotspot mutations. For each cohort, RAPGEF1 expression was dichotomized into high- and low-expression groups based on the median expression value of the full dataset. Associations between RAPGEF1 expression and survival were then evaluated within the wild-type tumor subgroup using the Cox models. To ensure comparability across datasets, all survival times were uniformly converted to days. We performed melanoma-specific survival in the TCGA, Leeds, and Lund cohorts, and overall survival in the Liu, Van Allen, and Riaz cohorts as they lacked melanoma-specific death data. Because stage I tumors were profiled exclusively in the Leeds cohort and are associated with a high survival rate (93-97%) ⁷⁸, we excluded these tumors from the survival analysis to ensure comparability between the Leeds cohort and other melanoma cohorts, which predominantly included more advanced-stage tumors. For these analyses, we adjusted for covariates specific to each cohort as follows: TCGA: Age, Tumor site; Leeds: Age; Lund: Age; Liu: Cosmic aging; Van Allen: Age. Across all cohorts, the WT tumor subgroup included only a small number of samples, and when the number of events is low relative to the number of covariates, the Cox model can encounter monotone likelihood issues, resulting in extremely large or unstable coefficient estimates. To address this, we applied Firth’s correction.

Meta-analysis of tumor data

We performed meta-analysis of the RAPGEF1 expression association with RAS-MAPK subgroups, as well as the association of RAPGEF1 expression with survival across the analyzed melanoma cohorts as follows: Let $z_i$ be the z-score for cohort i. Let $n_i^{+}$ be the number of subjects with the mutation and $n_i^{-}$ be the number of subjects of wild type, then the effective sample size $m_i=\frac{n_i^{+}{n}_i^{-}}{n_i^{+}+n_i^{-}}$. We choose weights $w_i=\sqrt{m_i}∕{\sum }_{i=1}^5{m}_i$ and perform meta-analysis as $Z={\sum }_{i=1}^5{w}_i{z}_i$. One can show that $Z\sim N(0,1)$ under $H_0$ and has optimal statistical power assuming equal effect sizes (after normalization).

Fine-mapping of the 9q34.13 melanoma risk signal

In order to functionally characterize the melanoma risk locus on chromosome band 9q34.13, which is also associated with nevus count, but not hair color, tanning ability, or telomere length ¹, we first assessed whether the locus harbors a single or multiple risk signals.

Conditional GWAS analysis using the lead variant at this locus (rs3780269, P_GWAS=1.92 x 10⁻⁸, rs3780269-G OR=0.94) revealed no evidence of a secondary signal within +/− 1Mb of the main GWAS association signal (Table S10-11, Figure S1A). Multiple Bayesian fine-mapping methods (SuSiE, DAP-G, and RSparsePro) ^25-29 similarly converged on a single and similar credible causal set of variants (CCVs; Table S12).

To facilitate variant-to-gene mapping, we defined a credible causal set by taking the union of CCVs identified using Bayesian fine-mapping approaches. Within this set, the GWAS lead variant (rs3780269) was consistently assigned the highest posterior probability of being causal across methods (DAP-G PIP = 0.52, RSparsePro PIP = 0.58, SuSiE PIP = 0.55). We also included as CCVs variants identified by fine-mapping directly from GWAS summary statistics by assessing log-likelihood ratio (LLR) relative to the lead variant. Finally, to account for additional variants not genotyped or imputed in the GWAS, we performed LD-based fine-mapping (r² > 0.8 with the lead variant). These methods collectively identified 14 potential causal variants (Figure 1A, Table S12).

None of these variants mapped to protein-coding sequence or canonical splice junctions (Table S13), nor was there a significant or colocalizing splice QTL (sQTL) for rs3780269 (see Supplemental Notes). One fine-mapped variant, rs3739497 (P_GWAS = 1.9 × 10⁻⁶; OR = 0.95; risk allele = A; protective allele = G; r² to GWAS lead = 0.88), is located in binding site for two microRNAs (miRNAs; miR-124-3p and miR-506-3p) in the 3′ untranslated region (UTR) of RAPGEF1, however we found no evidence that levels of RAPGEF1 were correlated with either in melanocytes, making it unlikely this variant impacts either miRNA binding or RAPGEF1 levels (see Supplemental Notes).

The 9q34.13 melanoma risk region physically interacts with and regulates RAPGEF1 and UCK1

Several melanoma credible set variants overlap active enhancer or accessible chromatin regions in human melanocytes or melanoma cell models (Figure 1B, Table S14), while the lead variant is located within a region broadly annotated as an enhancer across multiple cell types (ENCODE cis-regulatory element EH38E2732514)¹⁰², suggesting a cis-regulatory function. Consistent with this annotation, previous massively parallel reporter assays (MPRA) of fine-mapped risk variants ²³ showed the regions surrounding several had significantly different transcriptional activity compared to scrambled DNA sequences in melanocytes and melanoma cells (scramble vs alt comparisons in Table S15), indicating enhancer and/or repressive function (Figure S2).

Capture-HiC-based variant-to-gene mapping ²² of the full region of association in primary melanocytes identified chromatin interactions between enhancers harboring fine-mapped variants (Table S14) and the promoters of RAPGEF1 and UCK1 (Figure 1A; Table S16), as well as to NTNG2 (which is lowly-expressed in melanocytes and melanoma). CRISPR-inhibition (CRISPRi) of regions surrounding select variants interacting with RAPGEF1 and UCK1 promoters (Figure 1B) in immortalized melanocytes confirmed a cis-regulatory effect of the region surrounding the GWAS lead variant (rs3780269) for both RAPGEF1 (2/3 guide RNAs; 0.85–0.88-fold relative to non-targeting controls, P = 0.0031 and P = 0.0018 respectively; paired Student’s t-test) and UCK1 (2/3 gRNAs 0.78–0.82-fold, P = 0.0025 and P = 0.0021 respectively; paired Student’s t-test), as well as rs3818580 for UCK1 (2/3 guides; P = 0.025 and P = 0.031) (Figure 1C). However, melanocyte and melanoma MPRA data for credible set variants ²³ did not find variants with significant allelic differences in transcriptional activity (ref versus alt columns in Table S15), so while elements in this region regulate both genes, these data do not implicate a specific variant.

RAPGEF1 and UCK1 are cis-regulatory targets at the 9q34.13 melanoma risk locus in melanocytic cells

Transcriptome wide association studies (TWAS) using eQTL data from human primary melanocytes (n=106) ¹³ as well as primary melanoma tumors from TCGA (n=103) ^14,15,103 collectively implicate two genes in the region, with melanoma risk positively associated with predicted melanocyte and melanoma expression of RAPGEF1 (P_melanocyte = 4.8 x 10⁻⁵; P_melanoma = 4.7 x 10⁻⁵) and negatively associated with melanocyte UCK1 expression (P_melanocyte = 1.8 x 10⁻⁵; Table S17; UCK1 was not included as a predictable gene from the melanoma dataset). Consistent with TWAS, which considers the collective effects of multiple sequence variants, the risk allele of rs3780269 itself is individually associated with higher RAPGEF1 expression in both melanocytes and melanomas (P = 1.4 × 10⁻¹¹ and P = 5 × 10⁻⁵, respectively; Figure S3A, Table S18, Table S19, Table S20) and negatively but not significantly associated with UCK1 (P = 0.22; Figure S3A, Table S18).

In contrast to melanoma GWAS fine-mapping results which suggest a single signal, colocalization and fine-mapping of RAPGEF1 and UCK1 eQTL signals point to potentially distinct cis-regulatory variants for RAPGEF1 and UCK1, each in partial LD with rs3780269. Melanocyte eQTL results show different lead SNPs for RAPGEF1 (rs13295485, P_eQTL = 2.8 x 10⁻²⁰, P_GWAS = 4.8 x 10⁻⁵) and UCK1 (rs7867616, P_eQTL = 2.8 x 10⁻²⁰; P_GWAS = 3.3 x 10⁻⁵; Tables S21, S22), each in partial LD with rs3780269 (1000G EUR r² = 0.50, D’ = 0.74 for rs13295485; r² = 0.17, D’ = 0.66 for rs7867616) and each other (r² = 0.12, D' = 0.5). Analyses of GWAS and eQTL signals using multiple Bayesian approaches (SuSiE + COLOC ⁵⁴ and SharePro ⁵⁵) do not provide support for colocalization of these eQTLs with the melanoma GWAS (Tables S23, S24; Figure S3B, C). Fine-mapping the RAPGEF1 eQTL in melanocytes reveals multiple sets of CCVs collectively containing over 100 variants and spanning the entire gene and promoter (Figure 1A, Table S25); critically several overlap the set of melanoma CCVs (Table S12). The single fine-mapped UCK1 eQTL credible set variant overlaps annotated UCK1 promoter (rs7867616, Figure 1A, Table S26) but not the melanoma credible set. Consistent with the eQTL data, we note a strikingly similar pattern in melanocyte methylation QTL (meQTL) data (see Supplemental Notes), where a methylome-wide association study identified significant CpG methylation probes in both genes and individual meQTL signals mirror eQTL signals for the respective genes. These data collectively reflect a complex cis-regulatory structure underlying the 9q34.13 association signal.

Contributions of RAPGEF1 and UCK1 expression to the melanoma association signal

To further reinforce RAPGEF1 and or UCK1 as strong causal candidates regulated by a complex pattern of cis-regulatory signals, we performed additional conditional GWAS analyses. Conditioning on predicted melanocyte RAPGEF1 and UCK1 expression each partially attenuated the GWAS signal for lead SNP rs3780269 (P_GWAS = 1.9 x 10⁻⁸, RAPGEF1 P_conditional = 1.0 x 10⁻⁴; UCK1 P_conditional = 2.1 x 10⁻⁵; Table S27, 28). When both genes were included in a joint model, both are significant (Table S29) and the signal for rs3780269 is dramatically attenuated (P_conditional = 0.002; Table S30). We observe similar results when instead conditioning on respective lead eQTL SNPs (RAPGEF1 P_conditional = 4.9 x 10⁻³, UCK1 P_conditional = 6.33 x 10⁻⁵; Table S31, 32; Figure S1B, 1C), suggesting a contribution of cis-regulatory variants for both genes to the melanoma risk signal. We note that conditioning only on the fine-mapped RAPGEF1 eQTL SNP with the highest PIP identified by DAP-G itself resulted in complete attenuation of the GWAS signal (P_conditional = 0.29, Table S33, Figure S1D), perhaps suggesting a critical role for RAPGEF1. Taken together, these results point to multiple causal variants in partial LD to rs3780269 respectively influencing two plausible effector genes, RAPGEF1 and UCK1, with both genes having opposing effects correlated with melanoma susceptibility: decreased UCK1 expression and increased RAPGEF1 expression are linked to risk.

RAPGEF1 and UCK1 both are critical for melanocyte growth and overexpression of RAPGEF1 stimulates melanocyte growth

Given the opposing effects for RAPGEF1 and UCK1 on melanoma risk, we assessed whether their expression impacts melanocyte growth and/or survival. We previously included RAPGEF1, but not UCK1, in a small pooled CRISPR proliferation screen in human melanocytes, noting the set of guides targeting RAPGEF1 as the third most depleted amongst all genes screened (P = 4.97 x 10⁻⁶) ¹⁸ suggesting a strong positive role in melanocyte proliferation and or viability. Here, we validated this finding, silencing RAPGEF1 using two independent shRNAs in immortalized human melanocytes and assessing cell proliferation by crystal violet staining and BrdU incorporation assays. Knockdown of RAPGEF1 using both guides reduced melanocyte number as assessed by crystal violet assay (Figure 2A, 2B). Consistent with these data, BrdU flow cytometry revealed a significant decrease in DNA synthesis after targeting with both shRNAs, indicating a block in cell cycle progression upon RAPGEF1 knockdown (Figure 2C). Parallel knockdown of UCK1 similarly resulted in reduced proliferation in immortalized melanocytes, as evidenced by both assays (Figures 2D-F), mirroring the effects observed with RAPGEF1 suppression. Critically, while a positive role for RAPGEF1 in melanocyte proliferation is consistent with observation that increased melanocyte RAPGEF1 is associated with melanoma risk, the essential role of UCK1 in melanocyte proliferation contrasts with eQTL and TWAS findings which show that increased UCK1 expression is protective against melanoma. While we cannot exclude a potential role for UCK1 in melanoma susceptibility at the 9q34.13 locus via means other than influencing melanocyte proliferation, the strong concordance between the essential function of RAPGEF1 in melanocyte proliferation and its elevated expression linked to increased melanoma risk led us to prioritize RAPGEF1 for further investigation.

We then assessed whether increased RAPGEF1 levels contributed to increased proliferation in immortalized melanocytes using a doxycycline induced promoter construct. Upon doxycycline treatment, cells showed significantly enhanced growth by crystal violet staining compared to untreated controls (Figure 2G, H), demonstrating that increased RAPGEF1 expression stimulates melanocyte growth and could contribute to melanoma development. In summary, RAPGEF1 expression influences melanocyte proliferation, with elevated levels driving increased cell growth that could in turn contribute to nevus and or melanoma development.

RAPGEF1 overexpression triggers colony formation of human melanocytes

Having established that RAPGEF1 overexpression enhances melanocyte growth, we next investigated whether RAPGEF1 functions as a potential oncogene capable of promoting colony formation, which is a hallmark of carcinogenesis and may serve as an in vitro marker for malignant transformation. While RAPGEF1 has diverse roles in cancer depending on the cellular context, tumor type, and stage, its function as a GEF for the small GTPase protein RAP1 is central to its potential oncogenic activity. Mechanistically, RAP1 can initiate and maintain MAPK/PI3K signaling, which can trigger malignant transformation of cells. To further evaluate whether RAPGEF1 expression can stimulate colony formation by affecting melanocyte growth, we examined the potential effect of RAPGEF1 on anchorage-independent growth of TERT-immortalized human melanocytes (p’mel cells) by soft-agar colony formation assay. While p’mel cells themselves do not efficiently form large colonies in soft-agar, stably expressing BRAF p.Val600Glu in p’mel together with other oncogenes lead to colony formation ^104,105. NRAS and its constitutively active mutants have been shown to promote malignant cellular transformation in multiple cells of different origins ^106,107 while the direct role of RAP1 in cellular transformation has not been well-reported. Here, we observed enhanced colony formation (>70 μm) in p’mel cells stably expressing both BRAF p.Val600Glu and constitutively activated NRAS, specifically NRAS p.Gly12Val or NRAS p.Gln61Arg (P < 0.0001, P < 0.0001, respectively; Figure 3A, B) relative to those expressing BRAF p.Val600Glu alone. Not surprisingly, enhanced colony formation was also observed in p’mel cells stably expressing both BRAF p.Val600Glu and a synthetic constitutively activated RAP1A p.Gly12Val mutant (P < 0.0001). By introducing RAPGEF1 overexpression in p’mel cells, we observed that RAPGEF1 similarly cooperated with BRAF p.Val600Glu to enhance colony formation (P < 0.0001). The NRAS p.Gln61Arg mutant induced a stronger colony formation effect than the NRAS p.Gly12Val mutant (P < 0.0001), consistent with NRAS mutations at codon 12 (such as those encoding p.Gly12Val) generally exhibiting intermediate oncogenic potency compared to the stronger activating mutations at codon 61 ¹⁰⁸. The number of colonies formed upon RAPGEF1 overexpression was lower than those induced by NRAS p.Gln61Arg (P < 0.0001) and RAP1A p.Gly12Val (P = 0.0002), but on a more similar level to wild-type RAP1A (P = 0.10) or NRAS p.Gly12Val (P = 0.32; Figure 3A, B). We note that we did not observe such cooperation in a BRAF p.Val600Glu-expressing zebrafish model (Tg(mitfa:BRAF p.Val600Glu); p53^−/−) ⁶⁷, where overexpression of human RAPGEF1 did not accelerate formation of cancer precursor zones (CPZs; P = 0.11) or tumors (P = 0.21), nor was RAPGEF1 overexpression sufficient to drive tumor formation in fish that did not overexpress BRAF p.Val600Glu. Overall, these data suggest that RAPGEF1 possesses an intermediate level of oncogenic potential, contributing to colony formation in melanocytes but insufficient to drive tumorigenesis in the absence of a strong oncogenic mutation in vivo.

RAPGEF1 overexpression in serum-starved melanocytes potentiates activation of both RAS and RAP1

Since melanocyte growth and colony formation effects were noted for RAPGEF1 overexpression, we hypothesized that its overexpression promotes oncogenic activity via activation of its known target RAP1. To test this, we overexpressed RAPGEF1 in melanocytic cells using a tetracycline-inducible system, and RAP1 activation was assessed by pulldown with RalGDS RBD beads, which selectively bind to the GTP-bound active form of RAP1. Under standard melanocyte culture conditions, RAP1 activation remained minimal regardless of RAPGEF1 overexpression, suggesting that cell growth supplement factors may suppress or mask RAPGEF1 related effects (data not shown). However, when cells were grown in supplement-free medium for 24 hours and then stimulated with human Epidermal Growth Factor (hEGF), a weak, transient increase in RAP1 activation was detected in cells not treated with doxycycline (dox), while dox-treated RAPGEF1 overexpressing cells showed enhanced activation, with a marked increase in active RAP1 at 15 minutes following hEGF treatment (Figure 4A). These results are consistent with previous studies showing that RAPGEF1 acts as a GEF for RAP1 in mammalian cells.

Given that the RAS signaling pathway is a key regulator of melanoma development, RAS alterations are common in melanoma, and that nearly every nevus harbors a mutation in BRAF or RAS, we asked whether RAPGEF1 overexpression affects RAS activation in melanocytes as we observed for RAP1. Compared to controls, RAPGEF1 overexpression led to RAS activation even in the absence of hEGF, with stronger RAS activation observed in RAPGEF1 overexpressing cells following hEGF stimulation (Figure 4B). We also assessed downstream signaling activity following RAS activation, noting increases in AKT activation correlating with levels of activated RAS (Figure 4B). Notably, although RAPGEF1 overexpression enhanced both RAP1 and RAS activation in response to hEGF, the kinetics of activation between RAS and RAP1 is clearly different, with strong RAS activation from 2 to 15 minutes in comparison to noticeable active RAP1 only at 15 minutes post hEGF treatment. These data raise the possibility that the activation of RAS and RAP1 by RAPGEF1 may not happen concurrently, and may be subjected to spatio-temporal regulation ¹⁰⁹. These data support a model in which dysregulated RAPGEF1 expression amplifies not only RAP1, but also RAS signaling in melanocytes, driving abnormal melanocyte proliferation and survival and contributing to melanogenesis and tumor development.

Elevated RAPGEF1 levels are associated with tumors lacking activating mutations in RAS-MAPK signaling

Hyperactivation of the RAS-MAPK signaling is a hallmark of melanoma ^110,111. Most melanomas acquire activating mutations in BRAF or NRAS, or loss-of-function mutations in NF1, which result in constitutive RAS-MAPK pathway activation ^112,113. In tumors lacking these canonical drivers, rare mutations in other RAS-MAPK pathway members have been reported to sustain RAS-MAPK signaling ^14,114-116. In melanocytes, RAPGEF1 overexpression induced RAS activation in the absence of hEGF treatment. Based on this, we hypothesized that RAPGEF1 expression may act as an alternative contributor to driving RAS-MAPK activation in melanoma tumors without known or strongly-activating RAS-MAPK activating alterations. To test this, we analyzed RAPGEF1 expression in six publicly available melanoma cohorts that included both mutation and gene expression profiling data (TCGA melanoma n=366 ¹⁴, Leeds n=503 ^75-77, Lund n=146 ⁸⁰, Liu n=84 ⁸³, Van Allen n=38 ⁸⁴, Riaz n=39 ⁸⁶; Tables S2-8). Since mutational data for specific strongly-activating BRAF (codons Val600∣Lys601) and NRAS (codons Gln61∣Gly12∣Gly13) alterations were available across the broadest range of tumors in all datasets, and because a subset of the Leeds cohort was assessed only for these mutations, we initially focused on defining the RAS-MAPK wild-type and mutant groups based on these specific alterations. After adjusting for clinical and technical covariates (age, sex, tumor stage, and tumor purity), we consistently observed significantly elevated RAPGEF1 expression in wild-type tumors compared to the mutant subgroup across five of six cohorts (TCGA P = 3.4 x 10⁻⁴; Leeds P = 0.01; Lund P = 0.007; Liu P = 0.004; Van Allen P = 0.02; regression adjusting for age, sex, tumor purity, and tumor stage when available), with the direction of effect in the non-significant Riaz cohort consistent with the other five (P = 0.2) (Table S34; Figure 5A). Meta-analysis across these datasets showed a highly significant difference (P_meta = 6.6 x 10⁻¹⁰). We also expanded to consider models cumulatively adding additional BRAF and NRAS driver mutations (Model 2), HRAS and KRAS driver mutations (Model 3), loss-of-function NF1 mutations (Model 4), and driver alterations in other RAS-MAPK pathway members using definitions from the TCGA Pan-Cancer analysis ¹⁰¹ (Model 5; models listed in Table S9). RAPGEF1 expression remained consistently higher in the wild-type subgroups in meta-analyses of each model, with the strongest association in models considering driver mutations specifically in BRAF and NRAS (Model 2; P_model2 = 4.54 x 10⁻¹⁰; P_model3 = 1.09 x 10⁻⁸, P_model4 = 2.73 x 10⁻⁶, P_model5 = 1.96 x 10⁻³; regression adjusting for age, sex, tumor purity, and tumor stage when available; Table S34). Taken together, the consistent enrichment of RAPGEF1 expression in tumors lacking RAS-MAPK activating alterations, coupled with its ability to induce RAS, RAP, and MAPK activation in melanocytes, supports a model in which RAPGEF1 contributes to RAS-MAPK pathway activation in the absence of strongly-activating canonical driver mutations.

RAPGEF1 expression is associated with prognosis in tumors lacking RAS-MAPK activating alterations

We next evaluated whether RAPGEF1 expression has prognostic significance in melanoma tumors lacking RAS-MAPK activating alterations. Since BRAF (affecting Val600∣Lys601) and NRAS (affecting Gln61∣Gly12∣Gly13) hotspot variants were profiled across most cohorts; we limited survival analysis to tumors that were wild-type for these alterations. To minimize potential confounding due to treatment-specific effects, we analyzed datasets comprised primarily of individuals not treated with immunotherapy separately from those focused on immunotherapy-treatment. The Leeds, Lund, and TCGA cohorts predominantly included individuals who had not received immunotherapy, whereas the Liu, Van Allen, and Riaz cohorts comprised immunotherapy-treated individuals. RAPGEF1 expression was dichotomized into high and low groups based on a cohort-wide median cut-off for each cohort. In the analysis, non-immunotherapy-treated WT tumors from TCGA (n = 92), Leeds (n = 70), and Lund (n = 26) cohorts, higher RAPGEF1 expression subgroup was associated with poorer prognosis compared to RAPGEF1 low expression group (TCGA HR = 1.5, Leeds HR = 1.7, Lund HR = 2.8; Table S35, Figure 5B). Though these associations are below significance threshold within individual cohorts, a meta-analysis of these cohorts is significant (P=0.049) for RAPGEF1 overexpression predicting shorter survival (HR = 1.6; Table S35; Figure 5B). In contrast, for cohorts of individuals with WT tumors treated with immunotherapy (Liu = 26, Van Allen = 16, Riaz = 14), the RAPGEF1 high expression group was associated with improved prognosis (Liu HR = 0.4, Van Allen HR = 0.2, Riaz HR = 1.03) and was significant upon meta-analysis of these cohorts confirming a significant association between higher RAPGEF1 expression and longer survival (HR = 0.4, P = 0.01; Table S35, Figure 5B). The opposing effects of RAPGEF1 expression association between relatively non-treated and treatment cohorts is noteworthy, however given the low sample size across most cohorts, these preliminary findings should be validated in larger, independent cohorts to confirm and extend these observations.