The Genetic Evolution of Treatment-Resistant Cutaneous, Acral, and Uveal Melanomas

Tumor evolution begins with a single cell initiating a clonal expansion that can eventually lead to metastasis and, in some cases, therapeutic resistance. Melanoma exemplifies this evolutionary paradigm, originating with a normal melanocyte that transforms into a malignancy via molecular alterations (1). Melanoma subtypes differ in anatomic origin, pathogenesis (including canonical driver gene mutations) and therapeutic response (including immune checkpoint inhibitors and mutation-targeted therapies). Metastasis represents a lethal stage of melanoma across these subtypes. Although the treatment options for patients with metastatic melanoma have greatly improved because of recent therapeutic advances, patient outcomes differ widely due to treatment resistance associated with heterogeneity of driver gene mutations and treatment resistance mechanisms among histologic subtypes (2).

Acquired resistance occurs when a melanoma evolves in response to treatment. For example, some tumors develop somatic mutations (e.g., in NRAS or MEK) or amplifications that drive altered signaling in response to RAF kinase inhibitors (3, 4). In these patients, initial or mixed responses may be seen followed by progression months later. In contrast, intrinsic resistance occurs when genetic mutations that confer a high degree of fitness and hence drug tolerance are present in tumor cells prior to administration of therapy. In these cases, melanomas manifest early progression despite treatment (5). Mechanisms of melanoma intrinsic resistance are diverse and include somatic mutations in driver genes that current therapies do not directly target, such as NF1 and GNAQ mutations that mediate resistance to RAF inhibition, defects in immune signaling or antigen presentation, and a number of other biological factors (6–11). What is currently unknown is the extent to which acquired and intrinsic resistance mechanisms coexist in the same patient during metastatic progression, if they are shared across multiple metastases versus arise independently at each site, and if there are organ site–specific differences in acquired resistance. These issues represent an unmet clinical need to maximize the utility of immunotherapy and mutation-targeted therapies in patients with melanoma, among whom only approximately 50% derive long-term benefit from existing therapies. Furthermore, these therapies are more active against tumors of cutaneous versus acral or uveal origin, largely due to divergent genomic profiles among these distinct melanoma subtypes.

Studies have leveraged multilesional sequencing of primary tumors and metastases in cutaneous, uveal, and acral melanomas (12–22). However, in those studies, the number of sequenced metastases was generally <5 per patient, and the extent of genetic heterogeneity varied among the studies. Therefore, in a single study, we leveraged extended multilesional sampling of disseminated metastatic disease in 7 patients with metastatic melanoma coming to autopsy, representing cutaneous, acral, and uveal subtypes. Such an approach permits near “sampling to completion” to comprehensively define the genomic evolution of a patient's disease (23), thereby more accurately accounting for genetic heterogeneity of metastases across time and anatomic space.

This study was approved by the Institutional Review Boards of Johns Hopkins Medicine (Baltimore, MD) and Memorial Sloan Kettering Cancer Center (New York, NY), and was conducted in accordance with the Declaration of Helsinki. Informed written consent was obtained from each subject or each subject's guardian. Tumor samples and matched normal tissues were collected from 7 patients with pathologic confirmation of melanoma. Across these 7 patients, the therapeutic interventions received included surgery, immunotherapies, radiation, chemotherapies, and mutation-targeted therapies (Fig. 1; Supplementary Table S1). The timeline from primary diagnosis to death for patient MEL6 was 13 months, while the remaining 6 patients ranged 2–5 years. Tumor specimens collected premortem, including the primary melanoma lesion and metastases (formalin-fixed, paraffin-embedded; FFPE), were retrieved from pathology archives. Autopsies were conducted and tissues were collected within 2–44 hours after death. Following collection, each tumor and normal tissue was immediately frozen in liquid nitrogen and stored at −80°C, while a matching aliquot of each tissue was also preserved in 10% buffered formalin. Each frozen tissue sample was embedded in Tissue-Tek optimum cutting temperature compound and sectioned at 5–10 μm on a Leica Cryostat to create a hematoxylin and eosin–stained slide for review. Each FFPE sample was similarly sectioned using a Leica Microtome. Upon pathologic review, each tumor sample was macrodissected from sectioned slides to remove nonneoplastic components. The neoplastic component of each tumor underwent genomic DNA extraction using a Qiagen DNAeasy Blood and Tissue Kit and protocol, while the FFPE samples were extracted using a QIAamp DNA FFPE Tissue Kit and protocol.

DNA quantification, library preparation, whole-exome sequencing (WES), and bioinformatic analysis were performed by the Integrated Genomics Operation and the Bioinformatics Core at Memorial Sloan Kettering Cancer Center (New York, NY). After genomic DNA quantification using Qubit, sequencing libraries were created with the Agilent SureSelect XT Human All Exon V4 and WES was performed on a Illumina HiSeq 2500 in a 100 bp paired-end run using the Illumina HiSeq SBS Kit v4, to a target coverage of 150× for each tumor and 70× for each patient matched normal. The average depth of coverage achieved was 209× and 131× for the tumor and matched normal samples, respectively. Sequencing data were aligned to the hg19 reference human genome using Burrows-Wheeler Aligner (24), and read deduplication, recalibration, and sequence realignment were completed using the Picard toolkit and GATK (25).

For all 7 patients, after analysis with MuTect (ref. 26; v1.1.7) for single-nucleotide variants (SNV) and HaplotypeCaller (25) for small insertions and deletions (INDEL), we evaluated the putative somatic mutations to identify true positives while limiting false negatives, false positives, and sequencing artifacts. Each mutation required at least a 5% variant allele frequency with 20× coverage in at least one tumor sample, and less than 2% of the reads (or two reads total) of the matched normal sample with a coverage of 10×. We also removed passenger mutations found in multiple patients at >2% variant allele frequency or in ExAC at >0.04% (27). We restricted our analysis to missense, nonsense, frameshift INDELs, in frame INDELs, nonstop, silent, translation start site, and splice site mutations that affected <5 bp. In combination with manual review, across the 7 patients' tumors, this resulted in a total 4,680 unique somatic SNVs and INDELs.

Allele-specific somatic copy-number alterations (SCNA) were detected with FACETS (28). Allelic losses were defined as total copy number (tcn) = 0 and lower copy number (lcn) = 0, and gains were defined as tcn ≥ 10. For each tumor sample, we calculated the fraction of the genome altered (FGA) by somatic copy-number alterations (Supplementary Fig. S1). For the three cutaneous melanomas, the median FGA was 84%, although several tumor samples in MEL3 exhibited lower than 50%. For the two uveal melanomas, the median FGA was 64%. The two acral melanomas, MEL6 and MEL7, were categorically different from one another: MEL6 had a median FGA of 41%, whereas MEL7 had a median FGA of 100%. Supplementary Table S3 has the SCNAs we detected across the seven cases.

The FGA can also serve as a sequencing-based metric for tumor purity calculated by FACETS. From this analysis, three tumor samples indicated relatively low purities (less than 25% tumor cells): MEL2-M3, MEL3-M13, and MEL7-M4. These three tumors also exhibited the lowest SNV and INDEL burdens among the samples from each respective patient. As an orthogonal metric, we used Treeomics to estimate tumor purity for each sample using the SNV and INDEL data (29). These results agreed that the three tumor samples exhibited relatively low purities that were not due to differences in sequencing coverage, formalin fixation, or treatment status of the samples (Supplementary Fig. S1; Supplementary Table S1). Histologic review also showed low cellularity and viability across the samples. Given the concordance between the two metrics, and that low purity samples complicate phylogenetic inference, these three samples were removed from further analysis. For the remaining 107 tumor samples, the median purity estimates were 70% and 67% for the FACETS and Treeomics metrics, respectively.

For each tumor sample, whole-genome duplication (WGD) was assessed (Supplementary Fig. S1). WGD was called if the major copy number was greater than or equal to ≥ 2 and overall ploidy was ≥ 2.5, with > 50% of the autosomal genome affected.

All SNVs with at least 5% variant allele frequency in the respective sample and a C>T transition at a dipyrimidine site (YC>T) were considered consistent with UV-induced DNA damage. In addition, mutation signatures were analyzed using Palimpsest (30).

We used Treeomics v. 1.7.4 to infer the phylogeny for each patient's tumor based on the SNVs and INDELs (29). In cases with more than 10 tumor samples, a consensus phylogeny was inferred by consolidating Treeomics phylogenies derived from sample subsets. We used Treeomics to quantify the Jaccard similarity coefficients for each intrapatient pair of tumor samples and to infer phylogenies after removing somatic mutations consistent with UV induced DNA damage [C>T transition at a dipyrimidine site (YC>T)].

Each SNV and INDEL underwent driver gene mutation (31) analysis to determine the statistical likelihood of the specific mutation either activating an oncogene or inactivating a tumor suppressor gene (TSG). A mutation was considered a driver if the mutation was (i) nonsilent, (ii) CHASMplus assigned a TSG q-value or a gwCHASMplus q-value of ≤ 0.1, and (iii) the gene was defined as a driver by two or more previously published melanoma whole genomic or exomic datasets (Supplementary Table S4). Hotspot mutations in TERT were reviewed manually. Mutations were also analyzed for relevance to therapeutic resistance, including targeted therapies and immunotherapies (Supplementary Table S4). The mutations in Supplementary Table S5 underwent analysis by the Variant Effect Scoring Tool (VEST; ref. 32). SCNAs were annotated manually based on the same published melanoma datasets that were used for the SNVs and INDELs.

HLA typing was performed using a sequencing-based method to validate each patient's loci. Given the HLA status of each patient, each missense SNV was translated into a 17-mer peptide sequence that centered on the mutated amino acid. Each 17-mer was used to create 9-mer peptides using a sliding window approach to quantify putative MHC class I binding affinities via NetMHC v4.0. Neopeptides with a MT.Score of < 2, as defined by patient-matched specific HLAs, underwent further analysis.

Sequence data have been deposited at the European Genomephenome Archive (EGA), which is hosted by the European Bioinformatics Institute and the Centre for Genomic Regulation, under accession number EGAS00001003582. Further information about EGA can be found at https://ega-archive.org and "The European Genome-phenome Archive of human data consented for biomedical research" (http://www.nature.com/ng/journal/v47/n7/full/ng.3312.html).

We analyzed 7 patients who were diagnosed with cutaneous, uveal, or acral melanoma, and developed metastatic disease to multiple anatomic sites. Patient MEL1 was the subject of our previous analysis (33), while patients MEL2-MEL7 were newly analyzed in this study. Figure 1 shows the clinical timeline for each patient (Supplementary Table S1). Patient MEL1′s cutaneous metastases exhibited a mixed response to immunotherapy at the time of death, characterized by distinct gene expression profiles in responding versus progressing lesions (33). Metastases from patients MEL2-MEL7 were progressing at the time of death, in some cases demonstrating resistance to multiple therapeutic interventions (e.g., patient MEL6). Patient MEL4 (uveal melanoma) received local therapy to the primary tumor site but did not receive systemic treatment for metastatic disease. Upon death, each patient underwent a research autopsy to sample multiple metastatic sites (110 total tumor samples, range 11–26 per patient). We also recovered premortem tumor specimens from the pathology archives when available, including the primary tumors from 6 of 7 patients, and longitudinal biopsies of metastatic lesions. After performing WES of DNA collected from each tumor, we identified SNVs and INDELs (Fig. 2; Supplementary Table S2). We also analyzed the percentage of the genome altered by SCNAs, as well as WGD in each case (Supplementary Fig. S1; Supplementary Table S3). In cases MEL1, MEL2, MEL5, and MEL7, we detected WGD in nearly all tumor samples within each case, indicating that WGD in these cases was a relatively early event in the evolution of these tumors. Interestingly, in MEL3 (cutaneous) and MEL4 (uveal), we observed heterogeneity for WGD, indicating that WGD occurred relatively late in tumor evolution in these two cases. We did not detect any evidence for WGD in the acral melanoma, MEL6.

Recent sequencing studies have demonstrated that cutaneous, uveal, and acral melanomas acquire distinct driver gene mutations during tumor evolution (1, 2, 34–37). The driver gene mutations we observed reflected the respective melanoma subtype of each case (Supplementary Tables S4 and S5). Specifically, the cutaneous melanomas harbored mutations in ARID2, BRCA1, CDKN2A, EZH2, KIT, MDM2, NF1, NRAS, RAC1, SMARCA4, and TERT; the uveal melanomas had mutations in BAP1, CDKN2A, GNA11, and GNAQ; and the acral melanomas had mutations affecting BRAF, CDKN2A, and NRAS. With the exception of a single SNV in CDKN2A in patient MEL3, the driver gene mutations were found to be present in all metastases within each case (Fig. 2). This implies that these common melanoma driver mutations occurred prior to systemic dissemination.

Cutaneous melanoma is one of most frequently mutated human cancers (34, 35, 37). UV light is known to cause a specific mutation signature (C>T at dypyrimidine sites) that is common in cutaneous melanomas (34, 35), but not in uveal or acral melanomas. The importance of UV damage in other melanoma subtypes has been a subject of scientific debate, with studies supporting (17) or refuting (38) a significant role. The cutaneous melanomas in our study exhibited a higher tumor mutational burden (TMB; Fig. 2, SNV median = 1,033, range 665–1,233) compared with uveal and acral melanomas (SNV median = 21, range 15–39). Upon analyzing these SNVs for mutational signatures, the three cutaneous melanomas exhibited an abundant UV-damage signature, whereas the uveal and acral melanomas did not (Fig. 2; Supplementary Figs. S2–S4; Supplementary Table S6). We also observed a high proportion of C>T mutations at dipyrimidine sites as compared with nondipyrimidine sites in the cutaneous melanomas (Supplementary Fig. S2). In the uveal and acral melanomas, we observed somatic SNVs that correlate with a UV signature, albeit at a lower proportion compared with the cutaneous melanomas. We found that the proportions of C>T mutations in the uveal and acral melanomas were similar between dipyrimidine and nondypyrimidine sites, as were the percentages of other substitutions at these sites (Supplementary Fig. S2; ref. 38). Overall, this indicated that uveal and acral melanomas did not exhibit a pronounced UV signature to the same extent as found in cutaneous melanomas. This was confirmed in a second mutation signature analysis using Palimpsest (30, 39), in which we observed that six of 12 mutation signatures were predominantly C>T transitions (Supplementary Figs. S3 and S4; Supplementary Table S6). Interestingly, a seventh signature represented a putative indirect effect of UV exposure (see Palimp_10; ref. 39). We identified one predominant mutation signature in each of the uveal and acral melanomas (Supplementary Fig. S3; Supplementary Table S6), each of which was dissimilar to the aging signatures previously described in these tumor types (37, 40), suggesting that the mutational processes of uveal and acral melanomas require further elucidation.

Phylogenies based on the somatic SNVs and INDELs (Fig. 2) indicated that the cumulative lengths of cutaneous melanoma trunks and branches were significantly longer than those of the acral and uveal melanomas. In part, this was due to an approximate 40-fold higher mutational burden in cutaneous melanomas (P = 0.02, t test), predominantly due to longer trunks rather than longer branches or leaves. We further quantified this observation using the Jaccard similarity coefficient (29), and found that the cutaneous melanomas had significantly higher Jaccard similarity coefficients among all tumors in an individual patient, than the uveal or acral melanomas (P < 0.00001, t test; Supplementary Fig. S5; Supplementary Table S7). This difference was not due to differences in tumor purities as assessed by the SNVs and INDELs (P = 0.93, t test; Supplementary Table S1). Overall, these results suggest that tumors from an individual patient with cutaneous melanoma exhibited higher genetic similarity than those found in patients with uveal or acral melanoma, although there were some exceptions (e.g., sample MEL1-M4). These findings indicate that in addition to having a higher TMB, cutaneous melanomas also exhibit different phylogenetic structures, typified by longer trunk lengths (i.e., a greater frequency of common mutations among all lesions), compared with uveal or acral melanomas. Thus, intrinsic treatment resistance may be more likely to occur in uveal and acral melanomas, which display more genetic diversity among different tumor lesions in an individual patient, compared with cutaneous melanomas. To determine the extent to which UV signature mutations specifically contribute to the differences in TMB or phylogenetic relationships, we recomputed the phylogenies after removal of all mutations corresponding to this signature from all cases (Supplementary Figs. S5 and S6). Trunks of the cutaneous phylogenies remained approximately 4× longer than those of the uveal or acral phylogenies, which suggests two features of cutaneous melanoma: first, UV mutations alone do not fully account for the large differences in mutational burden among these subtypes, and second, the TMB contributed by UV damage in cutaneous melanoma mostly occurs prior to metastatic dissemination.

We next explored somatic alterations that may have contributed to treatment resistance, given that genomic mechanisms have been implicated in intrinsic or acquired resistance to mutation-targeted therapies and immune checkpoint blockade in melanoma. Six patients received immunotherapy with checkpoint blocking antibodies (anti-PD-1, anti-CTLA-4), including 1 patient (MEL2) who also received IL2 and IL15; 2 of these patients also received BRAF and/or MEK inhibitors (cobimetinib, dabrafenib, and trametinib; Fig. 1). Thus, we analyzed each patient's SNVs, INDELs, and SCNAs for somatic mutations in genes curated from >70 melanoma studies (Supplementary Tables S4 and S5). One patient with ocular melanoma (MEL4) did not receive any systemic therapy and was not included in this analysis.

In 4 of 6 patients who received immunotherapy, we identified multiple somatic mutations that may have contributed to treatment resistance: an MDM2 gain and two point mutations in JAK2 in MEL1, a loss of heterozygosity in B2M and a SNV in EZH2 in MEL2, a B2M SNV in MEL3, and a PAK1 gain in MEL7 (Supplementary Table S5; refs. 8, 10, 11, 41, 42). These somatic mutations were detected on the trunk of each patient's respective phylogeny, except for the JAK2 point mutations in MEL1 and the B2M SNV in MEL3 (Fig. 2). In MEL1, one JAK2 mutation (L1001*) was detected in a subset of the cutaneous metastases that were progressing or regressing under immunotherapy (Supplementary Table S1), while a second frameshift mutation in JAK2 (E268Gfs*5) occurred in one regressing metastasis (MEL1-M15). Both JAK2 mutations were predicted to have a high pathogenicity impact by VEST (E268Gfs*5 P = 0.012, L1001* P = 0.002; ref. 32). Thus, both mutations could have resulted in a loss of function of JAK2. The two JAK2 mutations may have thus represented emerging resistant subclones that did not yet manifest as fully resistant to immunotherapy. The B2M mutation in MEL3 occurred in two lung metastases, MEL3-M17 and MEL3-M19. We did not identify any somatic mutation associated with immunotherapy resistance in patients MEL5 or MEL6. In summary, we identified 1–3 putative resistance mechanisms to immunotherapy in the cutaneous melanomas MEL1, MEL2, and MEL3, and in acral melanoma MEL7.

Two patients, MEL5 (uveal) and MEL6 (acral), received therapies specific to their tumor driver mutations, in addition to receiving immunotherapy. Patient MEL5 received cobimetinib, and patient MEL6 received combination therapy with dabrafenib + trametinib. Whereas we did not identify any known mechanisms of intrinsic resistance in MEL5, we identified a single therapeutically relevant mutation, a SCNA deleting CDKN2A on the trunk of the phylogeny, that may have contributed to resistance of combined BRAF and MEK therapeutic targeting in MEL6 (43).

To further explore genetic mechanisms potentially related to endogenous antimelanoma immunity and immunotherapeutic response, we analyzed the somatic SNVs and INDELs for putative neoantigens (Fig. 3; Supplementary Table S8). Each of the seven cases exhibited multiple putative tumor neoantigens predicted to bind to autologous MHC class I molecules (HLA-A, HLA-B, HLA-C) for potential immune recognition (Fig. 3), although these mutations may not have been expressed in RNA and peptides may not have existed on the cell surface, or may have been unrecognizable by immune cells within each patient. Expectedly, the number of neoantigens in the cutaneous melanomas was higher than in uveal and acral melanomas, reflecting the higher mutational burdens found in the former. Interestingly, regardless of melanoma subtype, approximately half of the somatic mutations in each melanoma generated at least one predicted neoantigen (Fig. 3; range 44%–59%). Likewise, the proportion of mutations that yielded single versus multiple neoantigens did not widely vary across cases (Fig. 3). This demonstrates that, despite widely divergent mutational burdens, these seven melanomas exhibited similar proportions of somatic mutations that may have generated neoantigens recognizable by the immune system.

We also investigated the evolutionary patterns of neoantigens. We found that the proportion of neoantigens on the phylogenetic trunk varied among the cases; however, these proportions simply recapitulated the overall mutational pattern (Fig. 3). Thus, each case acquired neoantigens that were shared by all samples as well as neoantigens that were private to certain melanoma lesions. For example, in patient MEL4, the GNA11 p.Q209L mutation was a driver mutation that was shared by all tumors and also generated three putative neopeptides, one of which was predicted to bind to multiple HLA molecules in this patient (Supplementary Table S8). This GNA11 mutation exemplifies the possibility that driver gene mutations may create neoantigens, thus providing targets for immunotherapy that will be present across a patient population (44). Patient MEL4 also acquired a private mutation in the OGN gene in a liver metastasis (MEL4-M2; Fig. 3), a small deletion that was predicted to generate 52 unique neopeptides with binding affinities ranging from weak to strong (Supplementary Table S8). This illustrates that a private mutation, although restricted to a single metastasis, may nonetheless generate multiple neopeptides that could be targeted by the immune system. As another example of this principle, we observed private mutations generating neopeptides among metastases from patient MEL6 (Fig. 3).

Previous multilesion studies of melanoma genomic evolution have most often focused on a single subtype—cutaneous (12–14, 22), uveal (15, 16), or acral (17, 18)—although some studies have included both cutaneous and acral melanomas (19, 20). Birkeland and colleagues sequenced cutaneous, uveal, and acral melanomas (21); however, only the cutaneous melanomas underwent multilesion analysis. To our knowledge, our study is unique in utilizing WES to survey an extensive set of tumors representing these three melanoma subtypes. This is important because the dataset we analyzed here was normalized for sequencing approach, mutational calling, and evolutionary analysis, thus providing an opportunity to compare evolutionary patterns across patients with melanoma.

With improvements in the utility of immune checkpoint blockade and mutation-targeted therapies against metastatic melanoma comes the need to better understand the mechanisms of treatment resistance. In this study, we uncover several aspects of the genetic evolution of such resistant tumors across three main melanoma subtypes.

In 5 of 6 patients in this study who received systemic immune-based and mutation-targeted therapies (cobimetinib, dabrafenib, and trametinib), we found 1–3 genetic alterations that may represent functionally nonredundant resistance mechanisms coexisting in the same patient. We also found that more than half of these resistance mechanisms (5/8) were shared by all metastases within a given patient. Notable exceptions that may represent clonal or organ-site specific differences among metastases were found in JAK2 in MEL1 and B2M in MEL3. Several studies in melanoma have described putative molecular markers of therapeutic resistance, including concordant driver genes (6), discordant driver genes (45), heterogeneously acquired resistance (43, 46), and transcriptional changes (33, 47). Our results indicate that these melanomas harbored genetic changes that may have contributed to therapeutic resistance, many of which were shared by all tumor samples within an individual patient (12, 14, 16, 21, 36). This indicates that future interventions to circumvent intrinsic resistance hold the potential to successfully target most (if not all) metastases within a patient. We did not identify any somatic mutations associated with immunotherapy resistance in patients MEL5 or MEL6. One explanation for this could be that the immunotherapy resistance in these cases was due to nongenetic factors (33), or that unknown genetic factors may confer immunotherapy resistance that are specific to uveal or acral melanomas.

Our results showed that the evolution of predicted MHC I–restricted neoantigens largely recapitulated the broader evolutionary patterns of somatic mutations. However, we found multiple examples of somatic mutations that generated an exceptional number of putative neopeptides. Such mutations were likely passenger gene mutations, a finding that has potential clinical implications. For example, MEL4 was diagnosed with a uveal melanoma, a tumor type that typically has a relatively low mutational burden. However, the single OGN mutation in liver metastasis MEL4-M2 was predicted to generate multiple neoantigens that may have served as targets for the immune system. This example illustrates that a single passenger gene mutation may generate multiple targets for immune cells, even in tumors with a low mutational burden. This finding is important because it provides optimism for treating low TMB metastases, as private mutations isolated to a single lesion may nonetheless provide additional molecular targets for patients receiving immunotherapies. Future studies will be needed to elucidate the extent to which such private mutations result in beneficial response to immunotherapy.

Melanomas arise from melanocytes that inhabit tissues at select organ sites throughout the body and are thus subject to diverse microenvironments and mutagens, resulting in several genetically and phenotypically distinct melanoma subtypes (2). For example, DNA damage from UV radiation represents a major mutagenic process in cutaneous melanoma, but not in uveal or acral melanomas [although UV radiation has been implicated in uveal melanomas arising from the iris (40)]. We found that the high mutational burden in the cutaneous melanomas was associated with a high genetic similarity among multiple tumor lesions in an individual patient. In contrast, the lower mutational burden in the uveal and acral melanomas was associated with a lower genetic similarity among tumor deposits. This finding aligns with the notion that UV-induced mutagenesis occurs prior to metastatic clone dissemination (48); however, the relatively high genetic similarities of cutaneous melanoma metastases did not depend solely on the UV mutational signature. This is consistent with three possible scenarios: (i) the three main melanoma subtypes (cutaneous, uveal, acral) have distinct cells of origin; (ii) the differentiation state of the origin cell differs among subtypes, possibly affecting mutation rates and DNA repair; and/or (iii) the origin cells across the subtypes experience diverse microenvironments.

The phylogenetic analyses described herein also have implications for the relative timing of metastatic dissemination in these cases. In all seven cases, the genomes of multiple metastatic clones were highly similar or identical despite inhabiting distant organs, implying that there was not sufficient time for DNA damage to accumulate among these clones during dissemination. This is surprising because several premortem tumors were nonetheless found to be highly similar to metastases collected years later (e.g., patients MEL1 and MEL2; Fig. 1). A relatively low degree of genetic heterogeneity has also been observed in pancreatic and lung cancers, in contrast to other tumors types with high genetic heterogeneity (e.g., kidney and endometrial carcinomas), although further investigation is needed to define the extent to which underlying tumor biology and clinical interventions influence the genetic heterogeneity observed (49).

Our study has several limitations. Although we analyzed a large number of specimens, they were derived from a limited number of cases (7 patients) with melanomas that were progressing at the time of death, and it is unclear how our results would generalize to a larger patient population. Our results suggest that the relative infrequency of acquired resistance to immunotherapy among patients with cutaneous melanoma may be due to a relatively high genetic similarity among multiple tumors within an individual patient, as exhibited by the cutaneous melanomas here. A second limitation of our study was that with the exception of MEL1, we did not analyze patients whose melanomas responded to therapy. A third limitation was that we only focused on somatic mutations, and did not analyze other types of molecular variation, or tumor microenvironmental and immune cell factors that are posited to contribute to melanoma pathogenesis and treatment resistance (50). Regardless of the molecular mechanism, our results quantify the extent that putative innate and acquired resistance mechanisms occur throughout the genome during melanoma progression and metastasis.

A.P. Makohon-Moore reports grants from NCI during the conduct of the study. E.J. Lipson reports grants from Melanoma Research Alliance, Bloomberg-Kimmel Institute for Cancer Immunotherapy, Barney Family Foundation, Moving for Melanoma of Delaware, and Laverna Hahn Charitable Trust during the conduct of the study, as well as grants and personal fees from Bristol-Myers Squibb, Merck, and Regeneron and personal fees from Array BioPharma, EMD Serono, MacroGenics, Novartis, and Sanofi Genzyme outside the submitted work. J.E. Hooper reports grants from NIH/NCI during the conduct of the study and other from Springer Publishing outside the submitted work. V. Makarov reports a patent for EP3090066A2, Determinants of cancer response to immunotherapy issued and with royalties paid. N. Riaz reports grants from BMS and Pfizer, and grants and personal fees from REPARE Therapeutics outside the submitted work. M.A. Postow reports grants and personal fees from BMS, Merck, and Array BioPharma; personal fees from Eisai; and grants from Novartis, RGenix, Infinity, and AstraZeneca outside the submitted work. D.B. Solit reports personal fees from Pfizer, Loxo Oncology, Lilly Oncology, BridgeBio, Illumina, Vividion Therapeutics, and Scorpion Therapeutics outside the submitted work. T.A. Chan reports other from Gritstone Oncology; grants from AstraZeneca, BMS, and Pfizer; and grants and personal fees from Illumina outside the submitted work. B.S. Taylor reports grants and personal fees from Genentech, Inc. and personal fees from Boehringer Ingelheim and Loxo Oncology at Lilly outside the submitted work. S.L. Topalian reports grants from Melanoma Research Alliance, Bloomberg-Kimmel Institute for Cancer Immunotherapy, and NCI R01 CA142779 and other from Barney Family Foundation, Moving for Melanoma of Delaware, and Laverna Hahn Charitable Trust during the conduct of the study, as well as other from Aduro Biotech, Dracen Pharmaceuticals, Jounce Therapeutics, Tizona LLC, and Trieza Therapeutics; personal fees from Amgen, AstraZeneca, Bayer, Dynavax Technologies, Immunocore, Immunomic Therapeutics, Janssen Pharmaceuticals, and Merck; grants from Bristol-Myers Squibb and Compugen; personal fees and other from DNAtrix, Dragonfly Therapeutics, Ervaxx Ltd., Five Prime Therapeutics, RAPT, and WindMIL; and grants and other from Potentza Therapeutics outside the submitted work; in addition, S.L. Topalian has a patent for Aduro Biotech licensed, Arbor Pharmaceuticals licensed, Bristol-Myers Squibb with royalties paid, Immunomic Therapeutics with royalties paid, NexImmune licensed, and WindMIL licensed. C.A. Iacobuzio-Donahue reports other from Bristol Myers Squibb outside the submitted work. No disclosures were reported by the other authors.

A.P. Makohon-Moore: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. E.J. Lipson: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. J.E. Hooper: Resources, data curation, investigation, methodology, writing-original draft, project administration. A. Zucker: Resources, data curation, investigation, methodology, writing-original draft, project administration. J. Hong: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. C.M. Bielski: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. A. Hayashi: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. C. Tokheim: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. P. Baez: Resources, data curation, investigation, methodology, writing-original draft, project administration. R. Kappagantula: Resources, data curation, investigation, methodology, writing-original draft, project administration. Z. Kohutek: Resources, data curation, investigation, methodology, writing-original draft, project administration. V. Makarov: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. N. Riaz: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. M.A. Postow: Resources, data curation, investigation, methodology, writing-original draft, project administration. P.B. Chapman: Resources, data curation, investigation, methodology, writing-original draft, project administration. R. Karchin: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. N.D. Socci: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. D.B. Solit: Resources, data curation, formal analysis, investigation, methodology, writing-original draft, project administration. T.A. Chan: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. B.S. Taylor: Data curation, software, formal analysis, investigation, methodology, writing-original draft, project administration. S.L. Topalian: Conceptualization, resources, supervision, funding acquisition, visualization, methodology, writing-original draft, project administration, writing-review and editing. C.A. Iacobuzio-Donahue: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing.

The authors thank the patients and families who generously participated in this study. They thank Richard White for critical comments on the article, and Annamalai Kumar (MSKCC) for assisting with HLA genotyping. This study was supported by the Melanoma Research Alliance (to E.J. Lipson, C.A. Iacobuzio-Donahue, and S.L. Topalian), the Bloomberg-Kimmel Institute for Cancer Immunotherapy (to E.J. Lipson and S.L. Topalian), the Barney Family Foundation (to E.J. Lipson and S.L. Topalian), Moving for Melanoma of Delaware (to E.J. Lipson and S.L. Topalian), the Laverna Hahn Charitable Trust (to E.J. Lipson and S.L. Topalian), the NCI (R01 CA142779, to S.L. Topalian), and the MSKCC TROT fellowship (T32 CA160001-06 and K99 CA229979 to A.P. Makohon-Moore). The Legacy Gift Rapid Autopsy Program at Johns Hopkins is supported by an NIH Cancer Clinical Core Support grant P30 CA006973 and a grant from the Sol Goldman Pancreatic Research Center. This research was also funded in part through the NIH/NCI Cancer Center Support Grant P30 CA008748. The authors also acknowledge funding sources, including Pershing Square Sohn Cancer Research grant (to T.A. Chan), the PaineWebber Chair (to T.A. Chan), NIH R01 CA205426 (to T.A. Chan), NIH R35 CA232097 (to T.A. Chan), and the STARR Cancer Consortium (to T.A. Chan).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.