Revisiting determinants of prognosis in cutanous melanoma

Introduction

Melanoma is increasing in incidence. In the United States in 2014, there were over 76,000 new melanoma cases and over 9,000 melanoma-related deaths [1]. Current American Joint Committee on Cancer’s (AJCC) staging for cutaneous melanoma is based on primary tumor thickness and presence of ulceration, mitoses, nodal spread, and distant metastases as determinants of prognosis. However, survival outcomes can vary widely even within the same stage. Given the clinical and biological heterogeneity of primary melanoma, new prognostic tools are needed to more precisely identify the patients most likely to develop advanced disease. Such tools would affect clinical surveillance strategies and aid in patient selection for adjuvant therapy. Here, we discuss the evolution of primary melanoma prognostic determinants, their limitations, and the current literature of melanoma prognostic markers from a molecular and immunologic perspective.

Evolution of Cutaneous Melanoma Staging Over Time

In the mid-twentieth century, melanoma staging, and thus prognosis, were driven by thickness and level of invasion. In 1969, Clark et al. [2] described five levels of primary melanoma tumor invasion that inversely correlated with overall survival (OS) in a study of 208 primary melanomas [2]. Breslow et al. [3] also observed that tumor thickness combined with Clark’s level produced the most accurate prognoses in a cohort of 98 patients with primary melanomas examined for Clark’s level, tumor thickness, and cross-sectional area (thickness times diameter). No melanoma recurrences or metastases from primary tumors < 0.76 mm thick in Clark’s levels II and III occurred within 5 years of follow-up, thus 0.76 mm was recommended as a cut-off point to stratify patients for lymph node dissection [3].

Since these seminal studies, melanoma staging has continued to evolve as patient cohorts have expanded and additional prognostic features have been described. Today, the AJCC oversees melanoma staging guidelines, which incorporate tumor thickness and other histopathological parameters into the “TNM” (tumor, node, metastasis) criteria used for most solid tumors. In 2002, the AJCC made significant changes to the melanoma staging system, including increasing thickness thresholds for T staging from 0.75, 1.5, and 4 mm to 1, 2, and 4 mm, and including ulceration status as a second determinant of T staging, based on a multivariate analysis of 13,581 primary melanoma patients [4–5]. Additionally, in a subset of 4,750 patients without clinically-detected nodal metastases, Clark’s level was found to be predictive of outcome in thin melanomas (<1 mm), whereas ulceration was the most prognostic factor in thicker melanomas (> 1 mm) [5]. Other AJCC staging updates at that time included shifting thick T4 melanomas from stage IIIA to stage IIC, inclusion of satellite metastases in N staging, and exclusion of the size of nodal metastases, which was no longer found to be prognostic [5, 6, 7].

In 2010, mitotic rate (≥ 1/mm²) replaced Clark’s level to characterize T1 melanomas when a multivariate analysis of 4,861 T1 primary melanomas demonstrated that thickness, mitotic rate, and ulceration significantly predicted survival outcomes, while Clark’s level did not [7]. Inclusion of mitoses in melanoma staging was considered practice-changing. Due to a 10% risk of sentinel lymph node (SLN) positivity for primary tumor thickness of 0.76–1 mm with mitoses, surgeons might now be prompted to offer SLN biopsy to patients with thinner melanomas if mitoses are present [8]. Additionally, a lymph node was now defined as positive if any isolated tumor cells were detected histologically or by immunohistochemistry (IHC), whereas at least 0.2 mm of tumor was previously required for positivity. This also changed prognostic discussions that clinicians have with these patients, who are now being upstaged to at least IIIA melanoma for even a single tumor cell noted in the lymph node, conferring a worse prognosis [7–8].

Limitations of AJCC Staging System

While the AJCC cutaneous melanoma staging system has evolved over time to more accurately reflect patient prognosis, improvements are still needed. TNM staging does not always reliably predict patient outcomes by stage. For example, stage I melanomas ≤ 1 mm comprise the majority (∼80%) of melanoma diagnoses reported to the SEER Cancer Registry and carry an excellent prognosis with 5 year OS rates of 92–97% [9–10]. However, 15% of these patients die from melanoma [10]. Furthermore, stage IIC melanomas have a worse 5-year OS (53%) compared to stage IIIA melanomas (78%) and stage III patients have highly variable 5-year OS rates, ranging from 78%, 59%, and 50% for stage IIIA, IIIB, and IIIC, respectively [9, 11].

It has been hypothesized that adding measurements of host anti-tumor immune response and/or molecular features to staging may improve primary melanoma prognostication. New tumor-, host-, and/or blood-based biomarkers to address the inconsistencies of an ever-evolving, imperfect staging system are critically needed. We review the literature on prognostic molecular and immunologic markers in primary cutaneous melanoma, their associations with clinicopathologic and survival outcomes, and their potential for incorporation into current staging models.

Tissue-based Molecular Genetic Markers

Identifying tissue-based prognostic molecular markers in melanoma has been challenging. One major limitation is that most primary melanomas are preserved as formalin-fixed paraffin-embedded (FFPE) rather than fresh frozen tissues due to the small size of the specimens. High quality genetic material is less commonly extractable from FFPE, precluding their use with early generation sequencing methods [12]. Relative stability of mRNA is also inconsistent in FFPE samples [13]. However, improved extraction methods, newer technologies, and novel, stable RNA molecules, such as microRNA, have allowed exome- or genome-wide mutational analysis or RNA expression profiling from small amounts of genetic material isolated from FFPE [12]. Such technical and biological advances have yielded new insights into melanoma biology and new molecules to assess prognostically.

Genes that were initially identified as genetically altered drivers of melanoma tumorigenesis (BRAF, NRAS, PTEN, etc.) were captured in candidate-based analyses focused on known proliferation and survival pathways. Next generation sequencing (NGS) has afforded an unprecedented depth of knowledge of the mutational landscape of entire melanoma genomes, which has confirmed and expanded on these findings [14, 15, 16, 17]. However, challenges identifying genetic drivers of melanomagenesis and progression remain. A comprehensive analysis of whole exome sequencing (WES) data from 21 tumor types (TCGA and Broad Institute) calculated that sequencing information for >5000 samples is required to identify all enriched mutated genes present in greater than 2% of melanomas at 90% power [18], highlighting the particular challenge of interpreting sequencing data for melanoma. Moreover, among the genes nominated from melanoma NGS studies, few reports have associated individual mutations with clinical outcomes or studied their use as prognostic biomarkers in cutaneous melanoma. None-the-less, there has been a movement to genetically define melanomas, and here we review associations with patient prognosis for commonly mutated genes in melanoma.

Expression of protein-coding genes

Expression level-based analyses of mRNAs and proteins provide immensely alluring molecular markers, as there are tens of thousands of potential markers. In theory, proteins have a distinct advantage of being the functional end products of protein-coding genes. Moreover, their stability, modifications, and localization multiply their potential as prognostic markers. However, tools to measure proteins genome-wide are still developing and protein-based (immunohistochemical) biomarker studies thus far reported in melanoma have largely taken candidate approaches. Several comprehensive reviews and/or meta-analyses have examined protein-based biomarker studies in melanoma [43–46], thus we refer readers to these excellent manuscripts instead of addressing them in depth here.

Briefly, we do note 3 important multi-marker prognostic signature studies that have been developed with the use of tissue microarray-based (TMA) immunohistochemistry [43, 47]. Gould Rothberg et al. built a 5-protein signature (high non-nuclear/nuclear ATF2, high nuclear p21, high total β-catenin, low non-nuclear/nuclear p16, and low total fibronectin; from 20 candidates evaluated) indicative of favorable prognosis (OS) using immunofluorescence-based automated quantitative analysis (AQUA) histochemistry in 192 AJCC stage II FFPE specimens [43]. They validated this scoring method in a set of 226 additional primary melanoma patients. In contrast, Meyer et al. developed a 7-marker signature (Bax, Bcl-X, PTEN, COX-2, β-catenin loss, MTAP loss, and presence of CD20+ B-lymphocytes, from 70 candidates evaluated) indicative of poor prognosis (RFS and OS) in training (n = 362) and validation (n = 225) patient cohort sets [46]. Both studies identified β-catenin as a protective prognostic indicator. Of the remaining 6 reported by Meyer et al., none were evaluated in Gould Rothberg et al. Of the remaining 4 reported by Gould Rothberg et al., p21 and p16 were evaluated by Meyer et al., though this study did not examine prognostic impact of nuclear staining only (p21) or non-nuclear/nuclear ratios (p16).

Additionally, in a cohort of 395 primary melanomas, Kashani-Sabet et al. confirmed a protein-based prognostic signature containing NCOA3, SPP1 (osteopontin), and RGS1, which they previously identified as prognostic in a cDNA microarray study [48]. A multimarker index score of expression was found to be an independent predictor of disease-specific survival (DSS) and sentinel lymph node status, and was validated in an independent patient cohort. Interestingly, osteopontin has been demonstrated as a prognostic factor across multiple studies, as we detail below.

In contrast to protein, mRNA-based biomarkers can be robustly and quantifiably measured genome-wide using several methodologies and should be easily translatable into a highly quantitative RT-qPCR-based clinical assay. Such studies, however, are limited by poor quality mRNA obtained in archival FFPE tissues [13, 49]. Moreover, contamination from infiltrating cells and surrounding microenvironment likely result in significant expression bias and complicate cross-study comparisons. There are numerous reports documenting mRNA expression profiling of melanoma cell lines, metastatic melanoma tissues or comparisons across melanoma progression (nevi, primary, metastasis) [50–55]. However, only a few studies have examined gene expression profiling or developed mRNA-based prognostic signatures from primary melanoma tissues of different outcomes (Table 2).

Winnepenninckx et al. reported mRNA expression profiles for 83 fresh-frozen primary cutaneous melanomas, in which they defined a set of 254 genes expressed differentially based on 4-year distant metastasis-free survival (DMFS) and identified high expression of minichromosome maintenance (MCM) genes as key predictors of reduced DMFS [55]. A follow-up study re-analyzed this data in the context of biological pathways and found that high expression of DNA repair genes was associated with poor DMFS [56]. Interestingly, another recent study, which defined a pro-invasive HOXA1 transcriptional signature, found that patients from the Winnepenninckx et al. study with a ‘high’ HOXA1 signature had poor DMFS, adding an intriguing underlying transcriptional mechanism to those original expression findings, though they did not note if DNA repair pathway genes were highly represented in their signature [57]. A second primary cutaneous melanoma expression profiling study identified 92 genes associated with OS [58], and a subsequent follow-up study, which further assessed these 92 genes, reported a 9-gene signature (6 protective and 3 unfavorable) associated with OS in a 91 patient cohort [59]. A recent report utilized serial analysis of gene expression (SAGE) to identify differentially expressed transcripts between 116 primary tumors that had not metastasized (stage I/II) and 72 primary tumors that had already spread at initial diagnosis (stage III/IV) [60]. Subsequent multivariate class prediction analyses of the top 50 differentially expressed genes identified CD24 and EVL as strong predictors of metastatic risk in an independent set of 28 primary melanoma samples. Conway et al. [61] performed DASL (cDNA-mediated annealing, selection, extension, and ligation) expression profiling of 502 cancer genes for >350 FFPE cutaneous primary melanomas from two patient cohorts, identifying high osteopontin (SPP1) expression associated with reduced RFS, which confirmed previous findings [50, 62–63]. Osteopontin has been implicated in cell motility, proliferation, and in activation of the NF-kappa B pathway, which may add mechanistic explanations of its prognostic capacity [62]. A subsequent re-analysis of these data by the same group identified increased expression of DNA repair genes as associated with poor recurrence-free survival [64], consistent with the described re-analysis of Winnepenninckx et al. data [56].

Collectively, these studies reported from tens to hundreds of genes associated with metastatic risk and/or survival outcomes in primary cutaneous melanoma. No reported genes overlap in more than two studies. Of the 50 genes reported in Gschaider et al. [60], 7 genes overlap with Winnepenninckx et al. [55], though, interestingly, the cellular source of these RNAs is likely to be keratinocytes, endothelial cells, or infiltrating immune cells, which would support that microenvironment and/or host response are key prognostic markers for patient outcomes in primary cutaneous melanoma (Supplementary Table 1). Five genes from a combined list from Conway et al. and Jewell et al. overlap with Winnepenninckx et al., and osteopontin overlapped between the Conway/Jewell and Gschaider et al. studies. None of the 92 genes originally reported to be associated with OS in Brunner et al. [58] and further assessed in a follow-up study [59] overlapped with genes reported in the other studies, though we note that none of the other studies defined genes directly associated with OS.

Two recent studies by Gerami and colleagues selected a set of 28 gene targets (melanoma metastasis-associated genes identified from data mining, previously defined prognostic uveal melanoma genes, and BAP1) for analysis as a prognostic molecular signature in melanoma [65–66] . In the first study, the gene set was measured in 268 primary melanoma patient samples, which were mostly randomly stratified to training (n=164) and validation (n=104) cohorts [66]. In the second, analysis of gene set measurements was restricted to patients who had undergone sentinel lymph node biopsy (SLNB) procedure. Prognostic modeling was performed using Radial Basis Machine (RBM) modeling to separate patients into two risk classes that were plotted in KM survival curves and compared with AJCC stage, Breslow depth, ulceration, mitotic rate, and age or SLN status in univariate and multivariate analyses. Each study found the defined signature to be an independent predictor of metastatic risk. These studies are the basis of a clinical prognostic test (DecisionDx) offered by Castle Biosciences, which we discuss below.

Expression of microRNA

A newer class of RNA, microRNA (miRNA), has been of particular interest to cancer researchers over the last decade. While fewer in number than mRNA and not translated to protein, miRNA are ideal potential biomarkers because they are quantifiable genome-wide and highly stable in FFPE tissues [67–68], although they are equally affected by sample purity as mRNA. MiRNA expression profiling has typically been performed using microarray (genome-wide, but inconsistently quantitative) or RT-qPCR (fewer miRNA, but highly quantitative), though newer technologies, including RNA-sequencing and Nanostring, allow highly quantitative genome-wide analysis. Early studies of global miRNA expression in primary melanoma did not focus on prognosis, but rather on identification of miRNA that play functional roles in melanoma tumorigenesis [69–73]. Our group reported 18 miRNA associated with post-recurrence survival [74], with a subset of 6 of these microRNA (mir-150, mir-342-3p, mir-455-3p, mir-145, mir-155, and mir-497) defined as a post-recurrence survival signature. Interestingly, miR-342-3p is located in an intron of its host gene EVL, which was recently reported to be associated with metastatic risk in primary melanoma [60]. A recent study replicated the prognostic importance of miR-150 and miR-142 (from the 18 associated with post-recurrence survival [74]) in stage III lymph node metastases and patient sera [75].

We recently reported identification of metastasis-suppressive miRNA that we originally identified as differentially expressed in array-based global miRNA profiling of primary cutaneous melanoma tissues (2 cohorts containing 92 and 119 patient samples) [76]. We found these miRNA to be associated with thickness, recurrence status, and/or RFS. From this data, we defined a 4-miRNA signature (miR-150, miR-15b, miR-16-5p, and miR-374b-3p) that, in combination with AJCC stage, was predictive of development of brain metastasis [77]. Our group also reported a series of 134 miRNAs differentially expressed in nodular (n = 56) vs. superficial spreading melanoma (n = 26) histological subtypes, controlling for stage [78]; however, the possibility that selected miRNA signatures have independent prognostic importance for these subtypes has not yet been explored.

Van Kempen et al. performed expression profiling of 159 miRNAs by RT-qPCR in 23 fresh-frozen melanoma samples [79]. 11 miRNAs associated with tumor thickness in this patient cohort. In a validation patient set, expression analysis of these 11 and further in situ hybridization revealed loss of miR-200 family (miR-200a and c) and miR-203 expression during melanoma progression and at the invasive front of primary tumors. Of the miRNA we found consistently correlated with tumor thickness in our patient cohorts (n = 27), 10 were measured in the van Kempen study. 4 of 10 (miR-200b and c, miR-203, and miR-205) correlated with thickness, and each maintained the same direction as in our data (Supplementary Table 2). Of the 11 miRNA the van Kempen study reported, there was evidence of a correlation with thickness for 9 in at least 1 of our data sets, and all correlation directions were maintained between studies.

Several other studies identified miRNA that play functional roles in the proliferative or invasive/metastatic capacity of melanoma cells, based on analysis of miRNA expression and associations with patient outcomes. Our group previously reported that high expression in primary melanoma tissues of miR-30d, a pro-metastatic miRNA that regulates glycosylation and immune suppression in melanoma, associates with advanced stage, increased thickness, nodular histological subtype, and poor RFS and OS [80]. Another recent study used in vivo selection in mice to generate highly metastatic melanoma cell line variants [81]. MiR-1908, miR-199a-3p, and miR-199a-5p, which were highly up-regulated in these metastatic variants and critical drivers of melanoma progression, were inversely associated with metastasis-free survival in a cohort of 71 primary melanoma samples, though multivariate analyses were not performed and patient characteristics were not reported.

Several reports have defined loss of miR-211 expression as a key feature of melanoma cells, and its over-expression is suppressive of melanoma cell proliferation and invasion in vitro [71, 73, 82]. Surprisingly, however, no studies have yet reported associations of miR-211 expression in primary tumors with patient outcomes. In this regard, one of our recently published datasets, which included a set of congenital nevi, is supportive of decreased abundance of miR-211 in primary melanoma compared to nevi [76]. However, miR-211 expression in these primary tumors did not clearly associate with prognostic parameters or patient outcome in these data (not shown). Finally, Satzger et al. reported that high expression of the pro-proliferative miRNA miR-15b in a set of 128 primary melanoma tissues significantly associated with poor outcomes (RFS and OS) in multivariate analyses [83].

Overall, these studies demonstrate potential utility of miRNA expression-based prognostic assays, however, there is limited overlap of the key miRNA reported as prognostic in these studies and sample sizes remain modest. miR-150 and miR-15b are the most consistently associated with prognosis in primary and metastatic melanoma, based on tissue and serum (described below) studies.

Immune Markers of the Host

Serum Markers

While the serum marker lactate dehydrogenase (LDH) is incorporated into the AJCC staging system for substaging of metastatic (stage IV) melanoma, no serum-based biomarkers have been identified that would meet criteria for incorporation into the staging system for earlier-stage melanoma (predictive of relapse) or to provide better prognostic or therapeutic guidelines in advanced melanoma.

Melanoma inhibitory activity protein (MIA) is produced and secreted by melanoma cells. In a cohort of 84 patients prospectively followed with stage I and II cutaneous melanoma, high vs. low serum MIA concentrations were associated with increased recurrence (66% vs. 5%) and reduced DFS at 4 years (30.3% vs. 94.5%) [100]. Serum levels of S100B, a protein involved in microtubule assembly and inhibition of the tumor suppressor p53, increase progressively in each stage of melanoma and associate negatively with survival [101–102].

C-reactive protein (CRP) is an acute phase reactant elevated in numerous clinical scenarios, including states of infection, inflammation, and malignancy. Measurement of elevated serum CRP level may be superior to serum LDH level in detecting patients who progress to metastatic melanoma [103]. Serum CRP levels measured at time of primary melanoma resection and then sequentially thereafter showed an association of elevated CRP with worse OS, MSS, and DFS in stage I and II patients in multivariate analyses,, suggesting serum CRP may be a valuable prognostic biomarker in melanoma [104].

Two studies by our group examined the utility of serum-based miRNA abundance for prediction and surveillance of recurrence in melanoma patients [105–106]. Friedman et al. defined a 5-miRNA signature from sera isolated from at-diagnosis blood of 80 primary melanoma patients that classified patients into high- and low-risk groups, which separated well in KM curves for RFS. In a follow-up study, we also developed a refined 4-miRNA sera-based signature (miR-150, miR-15b, miR-30d, and miR-425) adjusted for AJCC stage to define low- and high-risk patients, which stratified patients by RFS and OS in two cohorts (n = 201, n = 82) [105].

Finally, there is interest in measuring circulating tumor DNA (ctDNA) in advanced melanoma as an early indicator of response to treatment [107–108]. While ctDNA measurements are not currently used in clinical practice, prospective studies will address whether ctDNA can be used as a blood-based biomarker to detect melanoma recurrence.

Discussion

Our review has identified numerous reports of individual or sets of biomarkers that are prognostic or associate with prognostic factors in the primary melanoma cohorts in which they were studied. Only a few have been examined in multiple similar studies by independent groups and in independent patient cohorts. Collectively, assessed biomarkers encompass melanoma cell intrinsic and host-derived gene-, RNA-, and protein-based factors, as well as host-based cellular factors (e.g. TILs). Ideally, the biologic functions of these markers would mechanistically explain their prognostic impact, and as such, focus on markers with biological plausibility might be prudent; however, many putative biomarkers, which may be useful prognostically, have underlying biology that has not yet been clearly elucidated. In spite of the many studies reported, translation of genetic, molecular, or host-based factors into clinical usage in primary melanoma has not yet occurred.

Immunostaining-based molecular markers are routinely assessed for certain cancers, such as HER2 amplification in breast cancer [109]; however, none of the protein-based studies in primary melanoma have advanced into clinical practice. Cancer centers are rapidly incorporating genetic sequencing platforms into routine clinical practice to identify somatic variants in patient tumors, however, primary melanoma patients are unlikely to receive such testing until there is clarity of the mutations that clearly associate with patient outcome or until additional adjuvant therapies become available. RNA expression-based molecular markers will be more challenging to implement into routine practice and will likely require commercial development and validation of these assays as superior to the somatic mutational-based prognostic markers that have achieved broad uptake in the oncologic community. Examples of prognostic gene expression signatures already in practice include Mammaprint and OncotypeDX in early stage breast cancers [110–111] and OncotypeDX in stage II colon cancer [112]. The development of new biomarker-based systems for prognosticating clinical outcomes in early-stage melanoma can serve to select patients for clinical trials of adjuvant therapies.

Recently, Castle Biosciences released the DecisionDx Melanoma gene panel which predicts the risk (high (69%) or low (3%)) of distant metastasis at 5 years for stage I-III patients [66, 113]. This panel may be used in clinical practice, however it is not currently recommended by any standard treatment guidelines, including the National Comprehensive Cancer Network (NCCN). Definitive assessment of the sensitivity and specificity of this test should be conducted in prospective analyses of patients who are representative of the general primary melanoma patient population. Moreover, guidance on interpretation of DecisionDx results is of critical importance. Clinicians may suggest that patients predicted by this test to be at high risk for metastasis undergo more frequent clinical evaluations and surveillance imaging, particularly in patients who have negative sentinel lymph node biopsies. However, in settings in which adjuvant therapies are either contraindicated or not available, this categorization may increase patient anxiety without real benefit to the patient. Would increased surveillance of these patients result in earlier detection of recurrent disease? More importantly, would early detection of recurrence improve overall survival or help reduce melanoma mortality rates of these patients?

As previously noted [43], there are a number of challenges that have limited prognostic biomarker study interpretations and cross-study comparisons. Since 2005, publishing groups are recommended to follow the REMARK criteria as reporting standards [114], which aim to encourage transparency and improve study reporting. REMARK guidelines cover the reporting of study design, preplanned hypotheses, patient and specimen characteristics, assay methods, and statistical analysis methods. We advocate that groups performing future prognostic studies should consistently use REMARK criteria as a guide during study design, implementation, and analysis, as well as final reporting [115]. In addition to these guidelines, there are other technical considerations for molecular prognostic studies that are not frequently considered or discussed. For non-histological studies, sample isolation and purity are critical considerations. For tumor cell-intrinsic biomarkers, pure tumor sample is ideal for accurate measurements, cross-sample, and cross-study comparisons. In contrast, molecular-based measurements of immune cell infiltrates or micro-environmental factors from surrounding tissue may also represent good prognostic markers. As such, consideration of the cellular source(s) of the markers to be measured can greatly inform strategies for sample isolation and the level of tumor purity that is desirable. In contrast, the tumor- or host-derived nature of histological-based markers is clear, because cellular architecture is preserved. Moreover, for protein biomarkers cellular localization can be considered (e.g. nuclear staining, nuclear/cytoplasmic ratio, etc.), as well as post-translational modifications. However, the quantitative nature of the assessment being performed is a key concern for histological-based markers, with the most rigorous studies being based on automated quantitative analysis (AQUA). In addition, histological studies are typically limited to a single, 1-cell thick plane (or even small tumor areas within such planes, i.e. tissue microarrays [TMAs]) of a much larger 3-dimensional object. This limitation is also present in non-histological marker studies, but to a slightly lesser extent, since multiple sections (5–10) or whole TMA cores are typically required to extract sufficient material for analysis. In these studies, which typically rely on small portions of a much larger tumor, tumor-cell heterogeneity, if present, may result in sampling bias that has the potential to increase false negatives and false positives, and thereby confound study interpretations.

In addition to these considerations, markers are measured at single points in time and thus cannot adequately address any lead-time bias associated with the dimension of time. To the extent that AJCC stage represents some measure of time from tumor onset, it would be advisable to adjust for or stratify by tumor stage in multivariate analyses. Among other factors to control for, histological subtype (superficial spreading, nodular, acral), which vary in terms of pathologic features and may represent different biological entities, is an important consideration [116–117]. While the data in this review are focused on cutaneous melanoma, the rarer and more therapy-resistant melanoma types such as uveal and mucosal will also require further study to identify markers of both prognostic and predictive value.

Additionally, future prognostic studies must consider which primary outcomes to examine (PFS, DMFS, OS, etc.) and understand how these outcomes may be affected by frequency of clinical surveillance strategies, including imaging, and treatment strategies. Although DMFS may be associated with lead-time bias, it may provide more accurate measures of associations with putative prognostic markers than OS, which will increasingly be confounded by efficacious systemic therapies administered to advanced melanoma patients. The effect of a changing treatment landscape on defining prognostic molecular markers in future studies warrants serious consideration. As archival FFPE samples are exhausted from patient cohorts for whom clearly efficacious therapeutics were not available, future prognostic studies will depend on samples from patients treated with MAPK inhibitors, immunotherapeutic regimens, and other therapies or combinations, which will have varying responses and survival outcomes. However, prognostic factors that are independent of therapeutic response will be more difficult to define, due to the large number of factors that will need to be controlled for during study design or adjusted for in multivariate analyses. Collectively, these limitations suggest that collaborative, multi-institutional working groups that design and conduct well-reasoned, pre-defined studies evaluating multiple markers and/or marker types in large, well-defined patient cohorts would be the most efficient deployment of resources and the most likely to identify reproducible markers. Such studies could also be designed to allow multi-modal, multi-parameter risk signatures to be defined, which might dramatically enhance prognostication of primary melanoma patients. Any promising prognostic markers identified should ideally be validated in prospectively-collected tissue sets, preferably in the setting of a clinical trial.

Conclusion

We continue to believe that improved prognostic stratification has the potential to transform primary melanoma patient management, as clear differences in outcome exist for patients with histologically similar lesions. Our understanding of the particular factors (genetic mutation, expression alteration, host response, etc.) that are critical for predicting patient outcomes is incomplete. Overall, the studies we have considered in this review have not defined prognostic markers that could be readily incorporated into the current staging system. As such, efforts should be continued in these and other (e.g. proteomics) directions to maximize the likelihood of identifying clinically useful prognostic biomarkers. Each of these directions informs the others, and we may find that each is independently important to derive a more advanced prognostic model that stratifies patients accurately and robustly.

Supplementary Material