Somatic inactivating PTPRJ mutations and dysregulated pathways identified in canine malignant melanoma by integrated comparative genomic analysis

Canine malignant melanoma, a significant cause of mortality in domestic dogs, is a powerful comparative model for human melanoma, but little is known about its genetic etiology. We mapped the genomic landscape of canine melanoma through multi-platform analysis of 37 tumors (31 mucosal, 3 acral, 2 cutaneous, and 1 uveal) and 17 matching constitutional samples including long- and short-insert whole genome sequencing, RNA sequencing, array comparative genomic hybridization, single nucleotide polymorphism array, and targeted Sanger sequencing analyses. We identified novel predominantly truncating mutations in the putative tumor suppressor gene PTPRJ in 19% of cases. No BRAF mutations were detected, but activating RAS mutations (24% of cases) occurred in conserved hotspots in all cutaneous and acral and 13% of mucosal subtypes. MDM2 amplifications (24%) and TP53 mutations (19%) were mutually exclusive. Additional low-frequency recurrent alterations were observed amidst low point mutation rates, an absence of ultraviolet light mutational signatures, and an abundance of copy number and structural alterations. Mutations that modulate cell proliferation and cell cycle control were common and highlight therapeutic axes such as MEK and MDM2 inhibition. This mutational landscape resembles that seen in BRAF wild-type and sun-shielded human melanoma subtypes. Overall, these data inform biological comparisons between canine and human melanoma while suggesting actionable targets in both species.


Introduction
Human melanoma is of increasing clinical concern. It is one of a few cancers with rising incidence, while five-year survival for patients with metastatic disease has until recently remained low (15-20%) due to a dearth of curative systemic therapies [1]. Discovery of frequent activating BRAF mutations in melanoma and treatment with selective inhibitors of this mutant kinase has led to dramatic responses in the setting of metastatic disease [2][3][4]. However, not all BRAF-mutant melanomas respond to targeted therapy and responses that do occur are often brief and followed by the emergence of drug-resistant disease [5]. Moreover, targeted treatment options in melanoma subtypes without activating BRAF mutations are limited. New treatment paradigms such as immunotherapy, drug combinations, and alternative dosing strategies may circumvent resistance and broaden the scope of precision medicine in melanoma [6][7][8][9], but rapid preclinical study of such regimens requires access to robust models that recapitulate complex tumor features such as intratumoral genomic heterogeneity and tumorhost interactions. Meanwhile, few animal models exist for uncommon molecular or histological melanoma subtypes such as BRAF wild-type (BRAFwt) or mucosal melanoma.
Naturally-occurring canine cancers are increasingly recognized as meeting a need for complex cancer models that develop gradually amidst interactions with the host stroma and immune system [10][11][12][13][14][15][16]. Spontaneous canine malignant melanomas, which are almost universally BRAFwt and for which the mucosal subtype is the most prevalent clinically significant form, may fill a specific gap in models of BRAFwt and rare histological melanoma subtypes [11]. Human mucosal melanoma is an aggressive histological subtype that is predominantly BRAF, RAS, and NF1 wild type (Triple Wild Type or TWT) with occasional mutations in KIT or NRAS. It carries a five-year survival rate between 12.3% and 35.3% [17][18][19][20][21][22][23][24][25][26]. Study of this subtype is limited by its low prevalence, accounting for only 1-2% of human melanomas in the United States with as few as 1,500 cases per year [27]. On the other hand, canine malignant melanoma accounts for up to 100,000 yearly cancer diagnoses in the United States, occurring most commonly in the oral mucosa, but also arising in cutaneous and acral epithelium [28][29][30][31].
Canine malignant melanoma is highly prevalent, closely mirrors human melanoma clinically and pathologically, and is extremely aggressive, with median survival for oral cases being a mere 200 days [32][33][34][35][36]. However, little is known about its genetic etiology. It is predominantly BRAFwt with frequent copy number alterations of regions of canine chromosomes (CFA) 13,17,22, and 30, alongside frequent MYC amplifications and deletions of CDKN2A. Targeted sequencing studies, though limited, have shown that it infrequently bears alterations in other known drivers of human melanoma [32,[36][37][38][39][40][41][42]. It has been shown that CFA30 aberrations are characteristic of canine oral melanoma and complex copy number profiles on this chromosome homologous to the same profiles on human chromosome (HSA) 15 in human mucosal melanoma are suggestive of rearrangements that may drive this melanoma subtype [41]. Despite the very low prevalence of BRAF mutations, immunohistochemistry (IHC) has shown that the mitogen-activated protein kinase (MAPK) and/or phosphoinositide 3-kinase (PI3K) pathways are activated in 52-77% of cases [32,[36][37][38][39][40]. These data hint at underlying mutations driving these pathways that could guide future biological exploration and therapeutic development in the canine and human diseases.
We therefore set out to map the genomic landscape of canine melanoma using a combination of massively parallel whole genome sequencing (WGS), array-based platforms and targeted sequencing to identify somatic changes driving these cancers. Here we report the identification of recurrent inactivating mutations in the candidate tumor suppressor gene PTPRJ in addition to frequent RAS mutations, and mutually-exclusive MDM2 and TP53 alterations. We thereby define the genomic landscape of these cancers and identify similarities between melanoma subtypes across species while highlighting subtype-specific aberrations that may be used to guide future research.

Patterns of mutation identified by whole genome analysis of canine melanoma
We undertook comprehensive analysis of acquired genetic alterations in a discovery cohort of seven melanomas and matched germlines from six dogs (two tumors were derived from one dog) using WGS for detection of subtle sequence alterations alongside long-insert WGS (LI-WGS, see Materials and Methods) [43] for sensitive detection of structural variants. We then performed copy number and targeted gene analyses in an additional 27 canine malignant melanoma tumors and three canine malignant melanoma cell lines (Table 1). Snap-frozen tumors (all primary tumors except one acral metastasis) and matching whole blood were collected through an IACUC-approved protocol at the Van Andel Research Institute (VARI) from dogs undergoing surgery at 21 specialty veterinary clinics located in 10 states (see Materials and Methods). Diagnosis of melanoma was confirmed by two independent board-certified veterinary pathologists (an on-site pathologist and BD) in addition to staining for three melanocytic differentiation markers where tissue was available (in 26 samples as indicated in S1 Table) [36, 44]. Diverse breeds are represented in this cohort with enrichment for Cocker Spaniels and Golden Retrievers (five dogs of each breed), an equal ratio of male and female dogs and a median age at resection of 11 years. Clinicopathologic characteristics for this cohort are described in S1 Table and S1 Fig. For WGS and LI-WGS respectively a median of 38/11-fold sequence coverage and 209/ 155-fold physical coverage was achieved (S2 Table). Read alignment was performed using the canine reference genome CanFam 3.1 and stringent criteria were used to call somatic sequence variants intersecting Seurat v2.6, Strelka v1.0.13, and Mutect v1.1.4 (Materials and Methods). A total of 31,053 somatic single nucleotide variants (SNVs) and small insertions and deletions (indels) were found with a median of 4,223 genome-wide SNVs (range 1,880-6,342) and 316 indels (range 88-655) and a median mutation rate of 2.03 mutations per callable haploid megabase (range 0.97-3.14, Table 2). The genome-wide SNV spectrum showed C:G>T:A transitions to be most prevalent, at a median of 27.09% of total SNVs followed by T:A>C:G transitions (median of 21.19%) and C:G>A:T transversions (median 15.74%, S2 Fig). Despite the prevalence of C:G>T:A transitions, most occurred in CpG dinucleotides and were not enriched at dipyrimidines (median 22.5%). Therefore, a canonical UV signature was not present in any of these cases (S2 Fig) [45, 46]. We additionally looked for TERT promoter mutations, which have been reported in 71% of human cutaneous melanomas and are associated with UV damage [47], but no somatic mutations were found within one kilobase of the TERT transcription start site. The most common mutation overall was C:G>T:A in GCG trinucleotides (median 3.29%) followed by C>T in ACG (median 2.6%) and C>A in TCT (median 2.5%) (S2 Fig). No evidence of localized hypermutation (kataegis) was identified in these tumors [48]. Comparative genomic analysis of canine malignant melanoma

Somatic coding mutations identified in canine melanoma
Tumors assessed by whole-genome analysis displayed an abundance of somatic structural variants (SVs) and copy number variants (CNVs), with a modest burden of SNVs in coding regions (Fig 1A and 1B). The landscape of somatic mutations in the full cohort of 37 tumors based on multi-platform analysis is shown in Fig 1C. Circos plots depicting somatic alterations in each tumor in the discovery cohort are shown in S3 Fig. Of the genome-wide SNVs described above, a median of 26 nonsynonymous (ns) single-base substitutions and indels occurred within coding regions (nsSNVs, range 14-42) with a median nonsynonymous: synonymous mutation ratio of 2.3 (range 1.9-3.9) (Fig 1B). We additionally performed RNA sequencing in this cohort, aligning with STAR2. 4 [49], calling SNVs with HaplotypeCaller (GATK 3.3.0), and utilizing IGV to manually validate expressed sequence variants (Materials and Methods). Ninety-seven percent of nsSNVs (all but five) identified by WGS and genotyped on more than one sequencing platform were confirmed in at least one additional platform (S3 Table). A number of mutations in orthologs of human cancer genes were present in a single tumor each. These genes include : ATF6, EPAS1, FAT2, FAT4, FOXA3, FOXO1, GAB2, HRAS, KIT,  KRAS, MMP21, NRAS, PBX1, and XPO1. Although no recurrent SNVs were seen in the discovery cohort, three genes were mutated in two cases: FAT4, LRFN2, and PTPRJ. Of these, only PTRPJ was validated in multiple platforms in both cases. Both cases containing somatic PTPRJ mutations were mucosal (ND10-166 and ND10-376) and both putatively contained two hits.
To determine the prevalence of mutations in a panel of genes whose orthologs are known to play a role in human melanomagenesis, as well as the PTPRJ gene mutated in two cases, we performed targeted Sanger sequencing of all protein-coding regions of BAP1, BRAF, CDK4, GNA11, GNAQ, KIT, KRAS, MDM2, MITF, NF1, NRAS, PTEN, PTPRJ, and TP53 in the expanded cohort. BRAF, CDK4, GNAQ, MDM2, MITF, and NF1 were all found to be universally wild-type whereas putative pathogenic mutations were discovered in BAP1, GNA11, KIT, KRAS, NRAS, PTEN, PTPRJ, and TP53 as described below and in S4 Table.

Somatic copy number and structural variants identified in canine melanoma
Somatic CNVs in the discovery cohort were identified by analysis of short-insert whole genome sequencing (SI-WGS) using established methods (Materials and Methods). A median of 27 focal CNVs (range 4-68), two focal amplifications with a log 2 ratio ! 2 (range 0-61), and eight focal deletions with a log 2 ratio 0.2 (range 3-41) were identified in the discovery cohort ( Table 2 and S5 Table) comprising 0%-10% of the genome (  Table). Nine of 10 GISTIC regions contained genes and included gains in orthologs of the human cancer genes MDM2 and CDK4. Additional cancer driver alterations (homozygous deletions of tumor suppressor genes or focal amplifications of oncogenes) included CDKN2A homozygous deletion (3%) and KIT focal amplification (8%) (S7 Table). Somatic SVs including translocations, inversions, and duplications, were identified in the discovery cohort, based on calls from Delly v0.7.6 [52] in LI-WGS (Materials and Methods). Between 9 and 65 predicted SVs were identified in each tumor (median 34) and were predominantly inversions ( Table 2 and S8 Table). No recurrent rearrangements were present. Notable alterations in human cancer gene orthologs impacted by SVs in single cases include an ARH-GEF12 inversion, a BIRC3 inversion, a CLPTM1L-TERT translocation, a DDIT3 inversion, a MYO5A translocation, and a TCF12 inversion. However, two regions of CFA10 and 30 were found to contain somatic SVs in two or more tumors. CFA10 rearrangements occurred in five of seven cases, four of which bore alterations in the region spanning 10-12 Mb (also a significant GISTIC region from CNV analysis). CFA30 SVs were also present in three tumors with alterations occurring within a region spanning 15-24 Mb (also encompassing a GISTIC region) in each case. Complex chromosomal rearrangements reminiscent of chromothripsis were observed in four tumors (ND09-345, ND10-370, ND10-361, and ND10-441), with chained or clustered breakpoints localized to a subset of chromosomes in regions that also contained copy-number oscillations [53] (S3 Fig).

BRAF, RAS, NF1, and KIT mutations
Approximately 90% of human cutaneous melanomas are driven in part by BRAF, RAS, NF1, and KIT mutations that confer constitutive mitogenic signaling through the MAPK pathway [24,45,54]. However, these alterations are far less common in human mucosal and acral melanomas [20,22,23,[55][56][57]. No somatic alterations in BRAF were identified within any platform in our canine melanoma cohort. However, RAS family members, whose protein products are predicted to share 100% sequence identity with their human orthologs, were the most commonly mutated genes in aggregate, occurring in 24% of cases in human-conserved hotspots (Figs 1C and 2A). NRAS codon 61 (Q61R/H/K) and KRAS codon 12 (G12C) mutations Comparative genomic analysis of canine malignant melanoma occurred each in four cases while a single case bore an HRAS Q61R mutation (nine total RAS mutations). All three acral and two cutaneous cases bore NRAS or KRAS mutations, while only 4/31 (13%) of mucosal cases bore an NRAS, KRAS, or HRAS mutation. Although NF1 copy number losses occurred in six cases, no homozygous deletions or truncating mutations were identified (S7 Table). KIT mutations were present in one cutaneous and two mucosal tumors Comparative genomic analysis of canine malignant melanoma (S3 and S4 Tables). In the cutaneous case, the mutation results in a glutamine (Q) to arginine (R) change in codon 396, notably a site of variation between canine and human orthologs, a change that is not predicted to be damaging by PROVEAN, and may constitute a germline SNP, but germline DNA was not available in this case [58]. KIT mutations in the mucosal cases included an in-frame deletion of amino acids 560-562, a likely damaging mutation in a commonly mutated region of the human ortholog, as well as an aspartic acid (D) to valine (V) change in codon 815 corresponding to the most common hotspot D816V mutations occurring in the kinase domain of KIT in human cancers (S5 Fig) [59]. Copy number gains encompassing KIT were also present in 10 samples (eight mucosal, one acral, and one cutaneous-Jones, 17CM98, ND10-104, ND10-158, ND10-365, ND10-370, ND10-376, ND10-361, ND10-363, and ND10-441), although no focal amplifications were identified (S7 Table).

PTPRJ mutations
The most commonly mutated gene in this cohort was the putative tumor suppressor gene PTPRJ, not previously shown to have frequent inactivating point mutations in cancer (Figs 1C and 2C). PTPRJ (also known as density-enhanced phosphatase 1 (DEP-1) or CD148) is a protein tyrosine phosphatase receptor originally discovered by virtue of its overexpression in dense cultures of human lung fibroblasts [60]. It has since been shown to be frequently involved in allelic loss or loss of heterozygosity (LOH) in human cancers and mouse models [61, 62] and to potentially play a role in oncogenesis in diverse cancer types, but somatic homozygous deletions or truncating mutations have yet to be described in cancer from any species and its tumor suppressor status remains controversial [63][64][65][66][67][68][69][70][71]. Canine and human orthologs share 70% sequence identity with a highly conserved C terminus containing the protein tyrosine phosphatase catalytic domain that is nearly 100% identical between species (S6 Fig). Sequencing of PTPRJ across all 37 tumors revealed nine mutations in seven cases (all mucosal), comprising 19% of all tumors and 23% of mucosal cases. Six frameshifts or stop gains were discovered in addition to two splice site mutations, a C-terminal 10-amino acid deletion, and a single predicted damaging missense mutation. Two cases-ND10-190 and ND10-376 -contained two mutations each, consistent with putative bi-allelic inactivation of a tumor suppressor gene. Further, LOH was evident by analysis of adjacent SNPs in WGS data in case ND10-166 bearing the M110fs mutation (S9 Table). Consistent with this finding, the PTPRJ frameshift in the ND10-166 tumor occurred at an allele ratio of 61% in DNA and 100% in RNA. Finally, PTPRJ transcript was observed in RNAseq data from the two PTPRJ-mutant tumors profiled by WGS and RNAseq (270.21 Fragments Per Kilobase of transcript per Million mapped reads (FPKMs) in ND10-166 and 92.37 FPKMS in ND10-376) as shown in S7 Fig. ND10-376, containing two somatic PTPRJ mutations (a frameshift and a splice site mutation) and 92.37 FPKMs, bore the lowest transcript abundance among all seven profiled tumors. Median PTPRJ FPKMs for the five PTPRJ wild-type tumors was 171.76 (range 92.37-512.25).

MDM2 amplifications and TP53 mutations
Inactivation of the p53 network is a critical step in tumorigenesis in nearly all cancers [72]. Both truncating TP53 mutations and amplifications of MDM2, a negative regulator of p53, are key routes to p53 inactivation [73]. Although TP53 mutations and MDM2 amplifications in human melanoma are less common [23][24][25]45, 54, 56], 16/37 (43%) of the cases in our cohort of canine melanoma bore focal amplifications of MDM2 or truncating TP53 mutations ( Fig  1C). A recurrent focal amplification on CFA10 was identified by whole genome analysis in three of seven tumors in the discovery cohort with extended SNP array analysis in the prevalence cohort revealing an additional eight tumors bearing these amplifications (minimal region 10.9-11.8 Mb) (Figs 1C and 2C). In total, 11/38 cases (29%) bore this amplification involving seven genes, with MDM2 being the likely amplification target ( Fig 2B). All such amplifications occurred in mucosal melanomas (11/31, 35%). CDK4, a cancer gene 10 Mb proximal to MDM2 in both human and canine genomes and often the target of bipartite amplification alongside MDM2 [74,75], was co-amplified in three of these cases. Identification of focal MDM2 or CDK4 amplification in the WGS-and RNA-sequenced cohort coincided with high transcript abundance for these genes relative to their wild-type counterparts (S7 Fig). MDM2 FPKMs were 357.48 and 331.21 for the amplified cases (ND09-345 and ND10-370) relative to a median of 67.02 (range 37.57-82.26) for wild-type cases. CDK4 FPKMs were 2,730.13 for the amplified case (ND09-345) versus a median of 201.24 (range 69.87-471.2 for the wild-type cases). Additionally, twenty tumors were additionally assessed for MDM2 expression by IHC (S10 Table and  We additionally discovered seven tumors with mutations in TP53 whose protein product shares 80% identity with its human ortholog (S9 Fig). Three of these mutations were truncating-a homozygous T90X in ND10-252, heterozygous K151fs in ND11-201, and a heterozygous Q306X in ND10-564 ( Fig 2D and S4 Table). Of the three missense mutations, R145C and R270H were predicted to be damaging. R145C occurred in two tumors and R270H in a single tumor, with both mutations confirmed somatic through analysis of matched germline DNA. Codon 270 in canine TP53 is homologous to codon 282 in human TP53, the fifth most common hotspot for mutations in human cancer [59]. The missense G290R variant is a likely SNP. It occurs in a tumor for which matched germline DNA is unavailable and it is predicted to be neutral, although it has not been previously described [76][77][78]. In keeping with findings in other cancers, no sequence mutations were present in MDM2 and MDM2 amplifications were mutually exclusive with TP53 mutations. Further, TP53 and MDM2 alterations were mutually exclusive with RAS mutations in all but one case (ND10-748, Fig 1).

Pathway dysregulation in canine melanoma
Common genomic subtypes of human cutaneous melanoma (BRAF, RAS (N/H/K), and NF1 in 90% of cases) that engage oncogenic signaling through the MAPK pathway are less common in human non-cutaneous melanoma and in canine malignant melanoma (24% of cases here, Fig 1C). Therefore, to undertake unbiased identification of pathways contributing to canine melanomagenesis, we performed pathway analysis using WGS data from the discovery cohort. We generated a list of all genes bearing nonsynonymous mutations, lying within chromosomal breakpoints or significant CNV regions from GISTIC (n = 1047) in order to determine enrichment of these mutated genes within specific KEGG and Reactome pathways (Materials and Methods) [79][80][81]. Network analysis of the affected genes identified 97 pathways with significant Benjamini-Hochberg corrected P-values (S11 Table). The most significantly enriched pathways were Insulin Receptor Substrate (IRS)-mediated signaling, and IRS-related events, for which 23% (19 genes) of the pathway members are mutated in this cohort. Notably, these pathways converge on MAPK and PI3K mitogenic signaling and contain core pathway members such as FGFs, EIF4G1, HRAS, KRAS, NRAS, and RPTOR. Indeed the majority of the enriched pathways contain members of MAPK, PI3K, or growth factor receptor signaling (e.g. PI3K cascade P = 0.002, mTOR signaling P = 0.008, signaling by Rho GTPases P = 0.012, VEGF signaling P = 0.017, RAF activation P = 0.017, melanoma signaling P = 0.021, RAS signaling P = 0.031, and MEK activation P = 0.036) and, in many cases, intersections with MDM2 signaling.

Discussion
Melanoma is a clinically significant disease in dogs, the study of which holds untapped potential for developing clinical approaches to improve the lives of pet dogs while also informing human melanoma biology and treatment. Few treatment options are available for locally advanced or metastatic canine melanoma in part because the molecular etiology is still largely unknown. Similarly, limited molecular understanding of rare sun-shielded and BRAFwt subtypes of human melanoma has constrained clinical innovation. In order to identify the molecular alterations underlying canine melanoma, we undertook a comprehensive multi-platform genomic investigation. Our integrated analysis confirms that although these tumors are driven by mutational landscapes distinct from those in human cutaneous melanoma, they share important similarities with BRAFwt and rare histological subtypes of human melanoma. These data not only guide biological and therapeutic studies in canine melanoma, but they also lend further support for the use of the naturally occurring canine model in comparative studies of human cancers.
This study builds on knowledge of the cytogenetic landscape of canine melanoma [41] to provide a comprehensive view of numbers and types of somatic coding mutations in this cancer. Given the dearth of genomic data for canine melanoma, we focused overall on collecting primary tumors from diverse breeds. While this study was not sufficiently powered to draw conclusions regarding breed associations with somatic mutations, it is nonetheless important to consider potential associations between breed and somatic mutational landscapes, particularly because such associations have been shown to occur in other canine cancers such as lymphoma [82]. Several breeds have been suggested to be at increased risk for malignant melanoma, particularly breeds with heavily pigmented skin or oral mucosa such as Cocker Spaniels, Schnauzers, Scottish Terriers, Poodles, Chow Chows, and Golden Retrievers [83]. Our WGS discovery cohort primarily consisted of Cocker Spaniels (four Cocker Spaniels, one English Cocker Spaniel, and one Labrador), a breed reported to be at higher risk of oral melanoma, but our extended cohort then included targeted sequencing of 13 melanoma hallmark genes (as well as PTPRJ, which was the only additional recurrently mutated gene in the WGS cohort) and copy number assessment from SNP arrays across 20 total breeds. Given that our WGS cohort was predominantly Cocker Spaniel, it is possible that other recurrent, breed-specific somatic SNVs in non-melanoma-hallmark genes could exist that were not captured here. Thus, future expanded study of breed-specific cohorts will be critical for further understanding the role of germline variation in shaping somatic cancer landscapes across species. It will also be important to further define subtype differences in expanded cohorts of canine acral and cutaneous tumors as well as benign and precursor lesions.
Overall, the genomic landscapes of human melanoma vary by anatomic site and degree of sun exposure [22,26,57]. Cutaneous sun-exposed melanoma is characterized both by high point mutation frequencies linked to UV damage [45] and also only modest burdens of structural variation. In contrast, sun-shielded and non-cutaneous melanomas harbor a low point mutation, but high structural mutation burden. Here, we establish that the canine malignant melanoma genome landscape resembles that reported in human sun-shielded melanoma. Canine melanoma of all subtypes in our discovery cohort is likely sun-shielded, including cutaneous tumors which occur in densely hair-bearing skin, although cropping or shaving during summer months may in some cases increase UV exposure. In keeping with this status, WGS in these two canine cutaneous malignant melanoma cases provides a deep view of their genome-wide mutation burden revealing low point mutation frequencies (median 2.03 somatic mutations per Mb) similar to that seen in human acral and mucosal melanoma WGS data from Hayward et al. 2017 (Fig 3A) [26]. Although we only profiled two such cases and larger cohorts are needed, a low point mutation burden relative to human sun-exposed melanoma has potential bearing on expected responses to immunotherapy such as anti-CTLA4 and anti-PD1 checkpoint blockade. Numerous studies have shown a clear positive correlation between mutation burden, abundance of neoantigens, and clinical benefit in human melanoma and other cancers [84,85]. Nonetheless, other molecular determinants of response to immunotherapy exist beyond simply mutation burden and the activity of such agents in canine malignant melanoma remains to be determined. Notably, CNV and SV burden from our WGS in canine malignant melanoma was markedly lower than all subtypes as described in Hayward et al. (Fig 3B and 3C) [26].
WGS additionally provides a deep view of genome-wide mutation signatures. High point mutation burden in sun-exposed cutaneous melanoma is understood to result from UVinduced over-representation of C>T transitions occurring in dipyrimidines versus non-dipyrimidines. UV-induced C>T mutations occurring in dipyrimidines comprise a low proportion of total SNVs in our cohort (25%), reflective of human sun-shielded cutaneous, mucosal and acral melanoma, in contrast to 85-90% of C>Ts occurring in dipyrimidines in human sunexposed melanoma (Fig 3C) [24,26,45,55,56,86]. This lends support for a non-UV etiology of canine melanoma.
The genome-wide SNV spectrum further revealed that C>T transitions in CpGs were the most common sequence alterations (S2 Fig). These mutations correlate with age in human cancers and are due to spontaneous deamination of 5-methylcytosine [46]. Enrichment for these mutations in canine melanoma is not surprising given that the largest risk factor for cancer in humans and dogs is biological (not chronological) age [87][88][89][90][91][92] and that the mean age of these dogs at the time of surgical resection was 13 years (range: [10][11][12][13][14][15][16]. Relative to the average number of human somatic mutations, these data provide further evidence that not only cancer incidence, but also mutational burden increases with biological, rather than chronological, age [93]. Commonly observed mutational patterns in human melanoma such as kataegis were not observed, although four tumors exhibited clustered or chained translocations suggestive of breakage-fusion-bridge events due to telomere crisis or of chromothripsis, in which one or a few chromosomes undergo punctuated shattering and reassembly events [53]. Such events have been linked to poor outcome in human melanoma [94] and may be enriched in tumors with p53 dysfunction or those that lack means to extend telomeres [95,96]. Notably, we show here that MDM2 and mutually exclusive TP53 alterations are common in canine melanoma. Similarly, inactivating p53 mutations have been found in human mucosal and acral melanoma, suggesting p53 pathway dysregulation may be crucial in non-UV induced melanoma development. Further, UV-induced TERT promoter mutations are common in human cutaneous melanoma, and, although they are rare in sun-shielded subtypes, these subtypes have been shown to bear enrichment for other types of mutation that drive TERT overexpression such as SVs and CNVs [57]. The cutaneous tumors in this cohort do not bear somatic TERT promoter mutations or other known genetic lesions that would enable telomere extension. Thus, telomere crisis and the survival of structurally aberrant genomes may play a significant role in the molecular etiology of canine melanoma. Comparative genomic analysis of canine malignant melanoma Our comprehensive analysis of canine melanoma reveals that most canine melanomas bear a low coding mutation burden and are also less structurally complex than human melanoma. Two WGS approaches coupled with array-based platforms have enabled deep interrogation of these changes, complementing recent cytogenetic analyses of this tumor type [41]. Significant copy number gains on CFA10 and 30 that have been reported as a defining signature of these lesions are recapitulated in this dataset (S6 Table). Our multi-platform approach was also able to further elucidate complex chromosomal rearrangements present in these regions. Both regions are involved in multiple intra-and inter-chromosomal structural events across this cohort (S8 Table). Additionally, focal amplification of the CFA10 10-12MB region encompasses MDM2, a gene which is known to drive human cancers and is currently being explored as a drug target in TP53 wild type tumors [97]. CNVs associated with canine melanoma also include gain of CFA13 and loss of CFA22. While not statistically significant via GISTIC in this cohort, both events are present in individual samples. Overall, extensive copy number and structural variation suggest high levels of large-scale chromosome instability, i.e. gain and loss of whole chromosomes or chromosome arms. Intriguingly, mutually exclusive focal amplification of MDM2 or inactivating mutation in TP53 have been shown to be enriched in BRAF-, NRAS-, and NF1-wild-type human melanoma, although human TP53-mutant melanomas tend to also display higher mutation burden and presence of C>T transitions [98]. Taken together the high degree of structural complexity, the lack of TERT mutations (barring one putative translocation) or telomere-lengthening mechanisms, and the frequency of MDM2/ TP53 mutations all suggest that chromosome instability plays a key role in canine melanomagenesis.
In the discovery cohort, putatively pathogenic somatic mutations in orthologs of human cancer genes were present in a single tumor each including ATF6, EPAS1, FAT2, FAT4, FOXA3, FOXO1, GAB2, HRAS, KIT, KRAS, MMP21, NRAS, PBX1, and XPO1 (S3 Table). Of the 14 melanoma hallmark genes evaluated in the prevalence cohort (including PTPRJ), an additional 24 putatively pathogenic somatic mutations were identified in seven genes-NRAS, TP53, PTPRJ, KIT, KRAS, GNA11, and BAP1 (S4 Table). Overall, across discovery and prevalence analyses, RAS gene family members were the genes most commonly bearing somatic SNVs, occurring in 24% of cases (Figs 1C and 2A), followed by TP53 and PTPRJ mutations each in 19% of cases, KIT in 8% and PTEN in 5%. Combined, these mutations most commonly impact proliferative and cell cycle/apoptosis pathways in patterns that display both key similarities and differences with human melanoma subtypes [26] (Fig 3D). For example, despite an absence of BRAF and lower abundance of RAS and KIT mutations in canine versus human mucosal melanoma, these tumors display likely MAPK activating events in 35% of cases. Further, canine mucosal melanoma displays a higher burden of cell cycle and apoptotic events (51%) than all subtypes from the Hayward comparator human melanoma cohort assessed here due largely to enrichment for mutually exclusive MDM2 and TP53 mutations in canine mucosal melanoma. However, these mutations are common in human cutaneous melanoma (ranging from 36% of cases in the Hayward comparator cohort to 62% in the TCGA cutaneous melanoma study [98]). Some of these differences are likely due to the still small sizes of noncutaneous human and canine melanoma cohorts and the need for greater resolution, particularly across different anatomic sites. For example, all but two of the canine mucosal cases described in this study originate from the oral cavity, whereas the eight human mucosal comparators are largely vulvar (3) or from nasal cavity (2). Thus, an ongoing need exists for broader profiling of these melanoma subtypes by anatomic site. Overall, however, these findings nonetheless suggest that both MAPK pathway inhibition (e.g. via MEK inhibitors) and p53 pathway inhibition (e.g. via MDM2 inhibitors) are important therapeutic axes for development in canine melanoma just as they are in human [38].
The oncogenic MAPK pathway is critically important in many cancers given its central role in conveying extracellular signals to the nucleus in order to regulate cancer hallmarks including proliferation, invasion, metastasis, and angiogenesis. The majority of human cutaneous melanomas are driven in part by constitutive activation of the MAPK pathway through mutation of genes such as BRAF, NRAS, NF1, KIT, GNAQ, and GNA11, often in a mutually exclusive pattern [99]. The high frequency of these mutations has motivated the TCGA classification of these tumors according to MAPK mutation status: BRAF (~50% of cases), RAS (~30%), NF1 (~15%), and TWT (~10%) [98]. These genomic categories are correlated with clinical, pathological, molecular, and biological features of melanoma and thus may comprise distinct subtypes. However, less common histological subtypes of melanoma such as mucosal, acral, and uveal melanoma bear unique mutation spectra that are not uniformly centered on canonical activating mutations in the MAPK pathway. Correspondingly, it has been shown that BRAF mutations are exceedingly rare in predominantly oral canine malignant melanoma and, to date, few alterations in other MAPK members have been discovered. These findings were recapitulated in our cohort, which showed no canonical BRAF or NF1 mutations. Nonetheless, MAPK and/or PI3K signaling have been shown to be activated in nearly all cases [100]. Additional mutations impacting the MAPK and PI3K pathways include three KIT mutations, two PTEN mutations, and one GNA11 mutation. In total, 35% of mucosal and 43% of all canine melanomas bear an alteration impacting the MAPK pathway (Figs 1C and 3D). Prior to our studies described here, the mutations underlying such activation have remained largely unknown.
Here we show a complete absence of somatic BRAF mutations (SNVs, CNVs, or translocations encompassing the BRAF locus) in canine malignant melanoma in keeping with prior studies [32, 37, 41, 101]. We also did not uncover truncating SNVs in or homozygous deletions of NF1. A higher proportion of our cohort bear RAS mutations than the 6-13% previously described [32,100], although prior studies have focused almost exclusively on NRAS exons one and two. All three major RAS family members are highly conserved (100% protein identity) between canine and human. In humans, of these family members, malignant melanomas predominantly bear NRAS mutations with only very rare KRAS and HRAS mutations. In our cohort, we found four NRAS codon 61 alterations (11%), four KRAS G12C mutations and one HRAS Q61R mutation. Further, four of these RAS alterations (two NRAS, one KRAS, and one HRAS mutation) occur in mucosal tumors, a frequency of 13% in this subtype. However, in our cohort all three acral tumors and both cutaneous tumors had detectable RAS alterations (three KRAS and two NRAS mutations). This unusual pattern of RAS mutation in canine melanoma may reflect important differences in biological, tissue, and species specificities of RAS family members.
These data point to the genomic lesions underlying MAPK and PI3K activation in a substantial proportion of canine melanomas, and to subtle genetic differences in disease subtypes within and across species. Most striking is the discovery of a putative novel tumor suppressor gene, PTPRJ, a receptor-type protein tyrosine phosphatase, which has been genetically and functionally implicated in cancer [61, 62], but for which clear genetic mechanisms of inactivation have yet to establish its definitive role as a canonical tumor suppressor gene. PTPRJ consists of an extracellular domain with eight fibronectin III motifs, a transmembrane domain, and an intracellular catalytic domain. It was originally cloned from HeLa cells and characterized by overexpression and hyper-activation in dense cultures of fibroblasts, by regulation of contact inhibition, and by its role in regulation of cancer cell proliferation and invasion [60, [102][103][104][105][106][107]. Early genetic studies of quantitative trait loci for mouse cancer susceptibility with homologous regions in human cancers pointed to recurrent PTPRJ deletions, LOH, and missense mutations in small cohorts of colorectal (49%), lung (50%), and breast (78%) carcinomas in addition to a correlation between PTPRJ LOH and colorectal cancer progression [61,62]. Additional sequencing studies in larger cohorts have identified nonsynonymous SNPs in the extracellular fibronectin repeats associated with risk of developing thyroid, colorectal, head and neck squamous cell, and esophageal cancers [67, 70,[108][109][110]. More recently, a subclonal K1017N missense mutation in the non-catalytic cytoplasmic domain of PTPRJ was identified in a primary breast tumor with significant enrichment in a brain metastases and patientderived xenograft [111]. PTPRJ substrates that may mediate its tumor suppressive potential include ERK1/2, Akt, various receptor tyrosine kinases, and Src kinases [42, [112][113][114][115][116]. However, Ptprj knockout mice have normal development with no cancer predisposition and thus inactivation of this gene does not appear to be sufficient to induce tumorigenesis [65]. Across all TCGA studies published to date (10,951 cases from 33 tumor types in the TCGA PanCancer Atlas accessed via cBioPortal), the frequency of somatic PTPRJ point mutations and/or deep deletions is low-211/10,951 (1.9%, S12 Table) [117,118]. Only 21 somatic PTRPJ mutations are present in the TCGA human cutaneous melanoma data set consisting of 363 cases (a single homozygous deletion, five truncating mutations, and fifteen missense mutations). However, a related receptor-type protein tyrosine phosphatase, PTPRD, is thought play a role in regulation of STAT3 signaling and has been frequently implicated as a tumor suppressor in human cancers through inactivating somatic mutation, focal deletion or methylation in glioma, melanoma, neuroblastoma, colorectal, liver, head and neck, and lung cancers [119][120][121][122]. In human cutaneous melanoma, PTPRD is deleted or truncated in 9-12% of cutaneous cases, but has not been determined to occur at high frequency in rare histological subtypes [50,55,56,120,123].
Here, we present the first report of recurrent somatic truncating mutations in PTPRJ in a naturally occurring cancer. We have discovered seven cases (19%) of canine melanomas bearing somatic PTPRJ mutations. Canine and human PTPRJ orthologs share 70% sequence identity with a highly conserved C-terminus containing the protein tyrosine phosphatase catalytic domain (S6 Fig). Sequencing of PTPRJ across all 37 tumors revealed nine mutations in seven cases (seven mucosal and one uveal) comprising 19% of all tumors and 23% of mucosal cases. Six frameshifts or stop gains were discovered in addition to one splice site mutation, a C-terminal 10-amino acid deletion, and a single predicted damaging missense mutation. Two cases-ND10-190 and ND10-376 -contained two mutations each, consistent with bi-allelic inactivation of a tumor suppressor gene. Further, LOH was evident by analysis of adjacent SNPs in WGS data in case ND10-166 bearing the M110fs mutation (S9 Table). Although regional LOH on chromosome 18 was observed by SNP array in three of six cases bearing single mutations in PTPRJ, these regions were not observed to directly overlap the coding region of PTPRJ. Overall, the enrichment for PTPRJ truncating mutation in canine malignant melanoma bears intriguing implications both for a previously underappreciated role for this gene in human melanoma (e.g. through as-yet understudied roles for hemizygous deletion [124] and/or epigenetic modifications) and for the possibility of unique mechanisms of tumorigenesis across species.
Through deep integrated genomic analysis combining WGS, LI-WGS, RNA sequencing, aCGH, SNP arrays, and targeted Sanger sequencing we have determined that canine melanoma is driven by frequent dysregulation of MAPK and cell cycle/apoptosis pathways and, in some cases as is seen in our WGS cohort of predominantly Cocker Spaniels, extensive chromosomal instability. In keeping with prior comparative melanoma studies that have incorporated histology, targeted sequencing, and aCGH [32,36,38,41], this work highlights the striking resemblance of canine malignant melanoma to BRAFwt subtypes of human melanoma. Finally, we have additionally discovered a putative novel tumor suppressor that may reflect unique species-specific biology and/or may highlight a tumor suppressive axis more subtly altered and as-yet underappreciated in human melanoma. This work bears immediate relevance for development of improved diagnostic and treatment approaches in canine malignant melanoma and provides further evidence to credential the naturally occurring canine melanoma model for study of relevant genomic subsets of human melanoma.

Ethics statement
Samples were obtained under institutional review protocols at the Van Andel Research Institute in collaboration with local speciality veterinary clinics (protocol #08-06-14).

Immunohistochemistry
Two tissue microarrays (TMAs), designated Dog MEL A TMA and Dog MEL B TMA, consisted of 96 individual dogs and 131 tissue samples placed in duplicate and two tissue samples placed in quadruplicate (272 array spots). Multiple tumors from nine dogs were present on the array and multiple samples from varying sites within the same tumor were present for twelve dogs. Additionally, non-melanoma stromal or control normal tissues were included. TMAs were hematoxylin and eosin-stained and evaluated via routine immunohistochemical procedures for melanoma cocktail (anti-melan A, anti-melanosome, and anti-tyrosinase), and antibodies to vimentin, MDM2 and p53. Samples scoring positive for MDM2 staining were then confirmed for positive staining with melanoma cocktail and re-evaluated for p53 staining. Positive staining was counted if at least one of the two duplicate samples could be evaluated for both MDM2 and melanoma cocktail on the TMA. Antibodies were purchased from Santa Cruz Biotechnology or Cell Marque. A total of 98 dogs and 189 spots/samples (132 tumors) met these criteria for evaluation for MDM2 protein expression by IHC. Of these 98 dogs, 18 dogs (17%) had melanocytic tumors positive for MDM2 staining in 33 spots/samples (25%). MDM2 staining was predominantly cytoplasmic highest intensity at junction between epithelial and subepithelial (submucosa, dermis). Staining was observed in both malignant pigmented and amelanotic melanoma and benign melanocytomas. Most intense staining (4+ cytoplasmic and nuclear) was observed in a benign cutaneous melanocytoma from a boxer that had additionally a malignant melanoma (negative for MDM2 staining on the array) and multiple cutaneous mast cell tumors.

Nucleic acid extraction from tumor tissue and blood
Tissue was disrupted and homogenized in Buffer RLT plus (Qiagen AllPrep DNA/RNA Mini Kit), using the Bullet Blender, Next Advance, and transferred to a microcentrifuge tube containing Buffer RLT plus and 1.6 mm stainless steel beads or 0.9 mm-2.0 mm RNase free stainless steel beads. Blood leukocytes (buffy coat) were isolated from whole blood by centrifugation at room temperature and resuspended in Buffer RLT plus. All samples were homogenized, centrifuged at full speed, and lysates were transferred to Qiagen AllPrep spin columns. Genomic DNA and RNA were then purified following the manufacturer's protocol. DNA was quantified using the Nanodrop spectrophotometer and quality was accessed from 260/280 and 260/230 absorbance ratios. RNA was analyzed on the Agilent Bioanalyzer RNA 6000 Nano Chip to validate RNA integrity (RIN!7.0).

Library construction and next generation sequencing
Three μg of genomic DNA from each sample was fragmented to a target size of 300-350 base pairs (bp). Overhangs in the fragmented samples were repaired and adenine bases were ligated on. Diluted paired end Illumina adapters were then ligated onto the A-tailed products. Following ligation, samples were run on a 3% TAE gel to separate products. Ligation products at 300 bp and 350 bp were selected for each sample, isolated from gel punches, and purified. 2× Phusion High-Fidelity PCR Master Mix (Finnzymes; catalog#F-531L) was used to perform PCR to enrich for these products. Enriched PCR products were run on a 2% TAE gel and extracted. Products were quantified using Agilent's High Sensitivity DNA chip (catalog#5067-4626) on the Agilent 2100 Bioanalyzer (catalog#G2939AA).
Long insert whole genome libraries were constructed as previously described [126] with the following modifications: 1100ng inputs were used; following DNA fragmentation, a bead purification was performed at a 1:1.8 (sample volume: bead volume) ratio; a 1% size selection gel was used; and during library enrichment, 10 PCR cycles was used. Libraries were clustered onto Illumina V3 flowcells (San Diego, CA) using Illumina's TruSeq PE Cluster Kit V3 (cat#PE-401-3001) and sequenced for paired 100bp reads using Illumina's TruSeq SBS Kit V3 (cat#FC-401-3002, n = 3) on the Illumina HiSeq.
10 ng of total RNA was used to generate whole transcriptome libraries for RNA sequencing. Using the Nugen Ovation RNA-Seq System (cat#7100-08), total RNA was used to generate double stranded cDNA, which was amplified using Nugen's SPIA linear amplification process. Amplified cDNA was input into Illumina's TruSeq DNA Sample Preparation Kit-Set A (cat#FC-121-1001) for library preparation. In summary, 1 μg of amplified cDNA was fragmented to a target insert size of 300 bp and end repaired. Samples were then adenylated and indexed paired end adapters were ligated. Ligation products were run on a 2% TAE gel and size selected at 400 bp. Ligation products were isolated from gel punches and purified. Cleaned ligation products were input into PCR to enrich for libraries. PCR products were cleaned and quantified using the Agilent Bioanalyzer.
Tumor and normal libraries were prepared for paired end sequencing as described above. Clusters were generated using Illumina's cBot and HiSeq Paired End Cluster Generation Kits (catalog#PE-401-1001) and sequenced on Illumina's HiSeq 2000 using Illumina's HiSeq Sequencing Kit (catalog#FC-401-1001).

Next generation sequencing data analysis
All informatic tools, versions, and flags are shown in S13 Table. BCL to FASTQ file conversion was performed using Illumina's BCL converter tool. Read alignment was performed with BWA (Burrows-Wheeler Aligner) v.0.7.8 [127] using the canine reference genome CanFam 3.1. Aligned BAMs were indel (insertion/deletion) realigned and recalibrated using GATK v3.3.0 [128,129] and duplicate reads marked using Picard v1.128 (http://broadinstitute.github. io/picard/). Variants were called using Strelka v.1.0.13 [130], Seurat v2.6 [131] and MuTect v.1.1.4 [132] and calls were annotated according to dbSNP 151, SNPs on the Illumina Cani-neHD BeadChip, and SnpEff-3.5 [133]. Final somatic SNVs were called by at least 2/3 callers. LI-WGS data were utilized for CNV and SV detection. For CNV detection, read depths at every 100 bases across sequenced regions were determined. Next, normalized log 2 foldchanges between tumor and normal were calculated and a smoothing window applied. Tumor allele frequencies of known heterozygous germline SNPs were utilized to evaluate potential false positives and correct biases. Finally, the Circular Binary Segmentation algorithm [134] was used to correct log 2 fold-changes. For mutation burden metrics, a focal CNV is included if the log 2 change is > = |2|. SV detection was performed utilizing Delly v0. 76 [52]. A minimum tumor allele ratio of 0.10 and a minimum quality score of 20 is required for an SV to be called.
RNA sequencing data in FASTQ format was checked for quality using cycle-by-cycle quality plots and biases such as GC content. Reads were aligned to the canine reference genome CanFam 3.1 using STAR-2.4 to generate alignment files in BAM format [49]. Somatic SNVs were called with HaplotypeCaller (GATK v3.3.0) and verified in IGV. Transcript abundance in FPKMs (Fragments Per Kilobase of transcript per Million mapped reads) was obtained by CuffDiff v2.2.1 [134] and annotated using ENSEMBL (CanFam 3.1.68).

Pathway analysis
A list of 1,405 genes with single nucleotide variation or structural variation or copy number variation from the discovery cohort were analyzed using ClueGo4 [79], a Cytoscape plug-in, to create a functionally organized pathway network. Kappa scores were then used to measure association between the networks. Functional networks were created with a minimum Kappa score threshold of 0.5 and a minimum of 3 affected genes in every network forming at least 10% of the total associated genes in that particular network. The genes were assigned to the networks based on the predefined pathways from KEGG, REACTOME and Wiki Pathways. 97 pathways were obtained, all with Benjamini-Hochberg corrected p-value <0.05. These pathways were grouped together based on inter-term kappa score and named by the most significant pathway in the respective groups.  Table). All amplification reactions were performed using Platinum Taq DNA Polymerase #10966-034 (Life Technologies; Carlsbad, CA). Briefly, each primer pair was mixed with 10 ng of genomic DNA and subjected to the following cycling parameters: 94˚C for 2 min., 3

Array comparative genomic hybridization
Oligo array CGH (aCGH) was performed by co-hybridization of tumor (test) DNA and a common reference DNA sample, where the latter comprised an equimolar pool of genomic DNA samples from multiple healthy individuals of various breeds. DNA was labeled using an Agilent SureTag Labeling Kit (Agilent Technologies, Santa Clara, CA) with all test samples labeled with Cyanine-3-dCTP and the common reference sample labeled with Cyanine-5-dCTP. Fluorochrome incorporation and final probe concentrations were determined using routine spectrophotometric parameters with readings taken from a Nanodrop1000. Fluorescently labeled test and reference samples were co-hybridized to Canine G3 180,000 feature CGH arrays (Agilent, AMADID 025522) for 40 h at 65˚C and 20 rpm, as described previously [135,136]. Arrays were scanned at 3 μm using a high-resolution microarray scanner (Agilent,Model G2505C) and data extracted using Feature Extraction (v10.9) software. Scan data were assessed for quality by the 'Quality Metrics' report in Agilent's Feature extraction software (v10.5) (Agilent Technologies).

SNP array genotyping
SNP genotyping was performed using the Illumina CanineHD array (cat#WG-440-1001). Per manufacturer's protocol, 200ng of DNA was first denatured then neutralized with 0.1N NaOH before amplification at 37˚C for 24 hours. The amplified DNA was then enzymatically fragmented and precipitated using 100% 2-propanol before drying for one hour at room temperature. After resuspension the fragmented DNA was then denatured and loaded onto the CanineHD BeadChip and hybridized for 16 hours at 48˚C. BeadChips were washed, a single base extension of hybridized primers added followed by multi-layer staining of the primers. Arrays were then washed, coated with the XC4 reagent (Illumina) and dried under vacuum for one hour. Coated arrays were read on the HiScan system and data visualized using the Illumina GenomeStudio Genotyping 2.0 software with an average sample call rate of 97%.

aCGH and SNP array data analysis
For both aCGH and SNP arrays, copy number data were analyzed with NEXUS Copy Number v8.0 software (Biodiscovery Inc., CA, USA). For cross-platform comparisons, LI-WGS BAMs were also analyzed utilizing Nexus software. CNVs were identified using a FASST2 segmentation algorithm with a significance threshold of 5.5×10 −6. Aberrations were defined as a minimum of three consecutive probes with log2 tumor: reference value of >1.14 (high gain), 1.13 to 0.2 (gain), −0.23 to −1.1 (loss), <−1.1 (big loss). Recurrent CNVs within each subtype were determined within NEXUS using an involvement threshold of 50%. Significance of these regions was then determined in NEXUS using the GISTIC algorithm (to identify regions with a statistically high frequency of CNVs over background) with a G-score cut off of G>1.0 and a significance of Q<0.05. CNV frequency comparisons amongst sample groups were performed in NEXUS using Fisher's exact test with differential threshold of >50% and significance p<0.05. Significance of each probe between the two groups was calculated in NEXUS using a Mann-Whitney test for median comparison.  Table. Whole genome sequencing metrics. Ã The average number of times a base is read or spanned by mate paired reads. Calculated using the formula C = N(2L+I)/G where C is the physical coverage, N = number of aligned reads, L = read length (the 2 multiplier denotes paired end sequencing), G = size of canine genome and I = inter-read base pair distance for PE seqencing. ÃÃ The proportion of the genome that could be genotyped accurately at a minimum read depth of 20 at a single locus; PF = Passing Filter. (XLSX) S3 Table. Somatic coding mutations identified in canine melanoma discovery cohort Ã . Ã Sequencing approaches include next generation sequencing (short-insert whole-genome (SI-WGS), long-insert whole genome (LI-WGS), or mRNA-Seq) and Sanger sequencing. ÃÃ Sanger validation uninformative n/a = <4 reads detected at this locus; ND = no data.