Despite recent advances in treatment of advanced stage melanoma, particularly with the expansion of the use of immune checkpoint inhibitors (6), melanoma continues to confer significant morbidity and mortality. In 2015, there were 59,782 deaths attributed to melanoma worldwide (7). Nearly 1.6% of all cancer diagnoses are melanoma, and the disease accounts for 0.64% of cancer deaths (8). In the United States specifically, it is expected that there will be over 100,000 new diagnoses of melanoma in 2020 with almost 7,000 deaths attributed to melanoma (9). In Europe, while cost of treatment varies widely based on country and stage, melanoma treatment can range from several thousand to tens of thousands of Euros per patient on average, and these costs are expected to increase with wider adoption of immune checkpoint inhibitors and targeted kinase inhibitors (10).

Varied and unique uncertainties surround clinical decision-making during the detection, diagnosis, and treatment of melanocytic tumors. We have reviewed and analyzed four critical decision points during the identification and treatment of melanoma for which biomarkers that reduce uncertainty have been investigated. These decision points are: (i) deciding whether to biopsy a melanocytic neoplasm, (ii) confirming histopathologic diagnosis, (iii) stratifying risk for lymphatic spread with consideration for SLNB and, (iv) selecting systemic therapy. While this review focuses on cutaneous melanoma, the considerations discussed could be adapted to potential biomarkers for mucosal and uveal melanoma as well.

One poignant example of where biomarkers could dramatically enhance melanoma care is with the histopathologic diagnosis of melanoma. Histopathologic diagnosis is the gold standard for melanoma diagnosis, but, despite the advent of immunohistochemistry (IHC) and standardization of diagnostic criteria (11–13), it remains a subjective medical art constrained by significant intra- and interobserver variability. While nuances exist, melanocytic proliferations largely exist on a spectrum from histologically benign to malignant, as described in the MPATH-Dx Classification scheme (Table 1) (11). This classification scheme breaks melanocytic lesions into benign (class I), moderately dysplastic nevus and Spitz nevus (class II), severely dysplastic nevus, atypical Spitz nevi (class III), AJCC T1a and T1b invasive melanomas (class IV), and AJCC T1b-T4a invasive melanomas (class V). In the absence of metastasis identified during sentinel lymph node biopsy (SLNB) or further staging, these primary melanomas are considered Stage I-II using the AJCC 8th Edition Pathological Staging Criteria (14).

Although lesions at the extremes of the spectrum (class I and class V lesions) are most likely to be reproducibly diagnosed as such, there remains significant diagnostic discordance among dermatopathologists for each class. In one study evaluating 187 pathologists, MPATH-Dx class I and V have intraobserver concordance rates of 77 and 67%, respectively, and interobserver concordance rates of 71 and 55%, respectively. When compared to consensus diagnosis, the accuracies of individual pathologists were 92% for class I and 72% for class V (15). Separating class II, class III, and class IV is a greater diagnostic challenge. Class II have the poorest concordance (35% intraobserver, 25% interobserver) and consensus accuracy (25%), followed closely by classes III (60% intra- and 45% interobserver consensus with 40% accuracy) and IV (63% intra- and 46% interobserver consensus with 43% accuracy) (15). Similar results have been seen in smaller studies (16, 17). To summarize, these studies concluded that pathologists concurred with their own previous evaluations in only two-thirds to three-quarters of cases, and, for most stages, reached the same diagnosis as their colleagues less than half the time.

Given the poor concordance in distinguishing class II through IV melanocytic lesions, several reports have sought to identify barriers to reproducibility. Unsurprisingly, those pathologists who are trained in dermatopathology, have a significant caseload of melanocytic lesions, and have more than 5 years of post-training experience have better reproducibility and accuracy (18). Dermatopathology training reduced misclassification rates of class II lesions by 10% and class III-IV lesions by 20%. While second and third opinions from trained dermatopathologists can reduce the misclassification rate of a general pathologist, additional peer review has a smaller, but still significant, benefit when a dermatopathologist performs the initial read (19). If these findings reflect the pathology community, then three highly trained dermatopathologists reviewing the same class II, III, or IV case, would be expected to have a consensus accuracy of 35, 52, and 62%, respectively (19). These studies indicate that training and experience are not sufficient to reach high levels of accuracy for intermediate lesions. The most beneficial, albeit modest, improvement in concordance was obtained by updating from the AJCC 7th edition diagnostic criteria to the 8th edition (12), suggesting that further refinement of diagnostic criteria could yet improve accuracy. However, diagnostic threshold variability between observers and a tendency to over call diagnoses due to medicolegal concerns (diagnostic drift) are expected to limit the histopathological diagnosis of melanocytic lesions regardless of diagnostic criteria improvements (20–22).

The above observations highlight the need for objective biomarkers to improve diagnostic accuracy. Biomarkers that bring clarity to the MPATH-Dx classifications would improve patient care by more accurately matching intervention risk to lesion prognosis and save patients and medical systems from expense and potential adverse outcomes related to unnecessary procedures. However, the poor concordance and accuracy of histologic melanoma diagnosis highlight the substantial challenge in developing such biomarkers. The development and validation of clinically useful biomarkers require a gold standard, or a true state, for comparison. If only unambiguous cases are selected for study, then the biomarker, by definition, provides no clinical value to augment current standard of care. Alternatively, if the gold standards are sufficiently ambiguous (for example, a consensus accuracy of 35%), then a theoretically perfect biomarker is expected to fail validation. Thus, for prospective biomarkers to yield useful test characteristics and address a clinically useful decision-making process, careful selection of the gold standard is required. In addition to validated clinical utility, practical considerations such as cost, invasiveness, tissue requirement, and ease of performing the test may influence biomarker use by clinicians.

The resources required to validate a candidate biomarker are substantial in terms of cost, time, and, most critically, patient volunteers. For example, an international coalition of melanoma experts reported that a randomized clinical trial evaluating the added value of prognostic gene-expression based biomarkers would require cohorts of 1,000 to 9,000 patients, depending on the trial design and hypothesis being tested (5). Therefore, it is critical to carefully consider all of the potential pitfalls of biomarker discovery, development, and implementation prior to committing limited resources to validation of any specific test. Considerations include not only the performance of a proposed biomarker, but also the size and diversity of the cohorts from which it was discovered and validated, and its potential to integrate with current practice and positively influence clinical decision-making. Prior to development, investigators should consider what test characteristics are required for a test to be useful, so that the field can determine how best to distribute limited resources for biomarker discovery and validation. These characteristics are described next.

The diversity of clinical circumstances surrounding individual melanomas ensures that there is no “one-size-fits-all” approach to the diagnosis and management of melanoma. Often, the first decision for a clinician in the evaluation of a pigmented skin lesion is to determine whether it is likely benign or has atypical clinical characteristics that warrant biopsy for further evaluation (Figure 1A). If a biopsy is performed, a pathologist must evaluate the likelihood of the lesion being malignant or benign (Figure 1B). For invasive malignant melanomas, a surgical oncologist must decide whether to offer SLNB to evaluate for early metastatic disease (Figure 1C), and selection of systemic therapy becomes necessary with local (e.g., a positive SLNB) or distant metastasis. Each of these decisions represent a junction where the clinician is faced with potential uncertainty and is forced to balance the odds of disease with a significant intervention. Clarification of uncertainty at these decision points might be achieved with objective biomarkers, as discussed in detail below. While these points do not comprise an exhaustive set of decision points in melanoma diagnosis and management that could benefit from biomarker development, tests that aid these decision points would lead to dramatic improvements in patient care. To build a foundation for this discussion, a summary of the fundamental theory of biomarker accuracy and utility is presented.

A clinically useful test must be capable of changing clinical decision making. In other words, usefulness is determined by the clinician ordering the test. Clinicians routinely encounter patients presenting with an ambiguous disease state. An illustrative example would be a sore throat that could be due to either Streptococcus pyogenes or a viral etiology. In this scenario, a rapid “strep” test or bacterial cultures may be ordered to resolve this ambiguity by increasing the clinician's certainty on whether the intervention of antibiotics would benefit the patient. A clinical test could resolve this uncertainty by decreasing the perceived likelihood of the disease state (exclusion), such as with a negative point-of-care rapid “strep” test. Alternatively, a clinical test could increase the likelihood of a disease state (confirmation), such as with a positive bacterial culture. If the test result shifts the clinician's certainty on the patient's diagnosis above or below their self-determined intervention threshold (e.g., a positive rapid “strep” test resulting in an antibiotic prescription), the test will have usefully influenced a clinical decision.

Similarly, predictive and prognostic tests may aim to determine disease trajectory and can assist in directing further intervention, such as follow up imaging or adjuvant therapy selection.

The potential utility of a biomarker can be estimated by first considering clinicians' certainty of a disease state using current standard of care (“prior odds”) and the certainty threshold that must be breached for intervention (“threshold”) (Figure 2A). It is important to recognize that since both values are derived from individual clinicians' perceptions, they are inherently imprecise and inconsistent. However, ranges for these values can be estimated through analysis of collective clinical decisions or surveying and these ranges are informative when considering the potential utility of a candidate test. When prior odds and intervention thresholds have been estimated, the slope of the line that connects them represents the change in the odds (also called the likelihood ratio or LR) required to change a decision. A clinical test that decreases the perceived likelihood of the disease state results in a negative likelihood ratio or LR- (Figure 2A, green line). A clinical test that increases the perceived likelihood of the disease state results in a positive likelihood ratio or LR+ (Figure 2A, blue line).

When determining whether a decision point might benefit from biomarker development, it is useful to estimate what values of LR would be required of a theoretical test for posterior odds to surpass the intervention threshold (Figure 2A). Likelihood ratios are calculated from the sensitivity and specificity of the test at one reference range or cut-off value [For information on how to calculate these and other established clinical accuracy metrics see (24) and (23)]. Sensitivity describes the ability of the test to discern the presence of a disease and specificity quantifies how often it correctly identifies disease absence. There is often a trade-off between sensitivity and specificity, particularly with test outcomes reported as a continuous variable. Since clinical accuracy is based on assignment of the test cut-off value (a value outside of the “normal” reference range that is flagged as an abnormal result), the sensitivity and specificity of a test can vary wildly based on where a laboratory assigns the cut-off values for normal. To address this issue, test characteristics can be plotted as a receiver operating characteristic (ROC) curve, which plots the sensitivity and specificity at different cut-off points. This also allows one to compare different tests/biomarkers by calculating the area under the curve (AUC). In general, the higher area under the curve reflects a better performing test, with an AUC of 1.0 having perfect sensitivity and specificity and an AUC of 0.5 being useless (Figure 2B). Since utility is determined by the clinician ordering and interpreting the test, there is no standardized AUC at which a test becomes clinically useful. Instead, positive and negative likelihood ratios can be derived from the slope between any ROC point and the (0,0) and (1,1) vertices, respectively (Figure 2B, point V). If a biomarker test application requires at least a LR+ of 10 and/or LR– of 0.1 to change clinical decision (as in the example presented in Figure 2A), then theoretically any point on the ROC curve that lies to the left (for LR+) or above (LR–) of the LR lines could be considered tests worth further development (Figures 2C,D). However, ROC curves with higher AUC will provide a larger range of potential test cut-offs that will meet these minimal LRs and will consequently have a higher sensitivity and/or specificity at those cut-offs (Figures 2C,D, ROC curves at point W and Y).

In the following sections we apply this theory to estimate the test characteristics required to change clinical decision making in the diagnosis and treatment of melanoma. Where available, we use a combination of published studies and surveys to estimate ranges for the prior odds and intervention thresholds. However, every providers' internal threshold of certainty for clinical decision making will vary in the context of patient-specific factors in addition to test characteristics. Thus, these ranges should not be considered exact quantifications but rather estimates to help determine which biomarkers may be worth investing community resources for further discovery and development.

Many factors are considered when assessing a pigmented lesion in clinic prior to recommending a biopsy. These include but are not limited to clinical impression of the lesion, history of the lesion (either as reported by the patient or through comparison to previous clinical images), the patient's history and risk for melanoma, total number of lesions, and dermatoscopic features. Since the decision to biopsy represents a gestalt of these different factors, it is not surprising that the level of specialty training and experience of the physician substantially alter the number of lesions biopsied for each melanoma identified. A recent meta-analysis of 351 studies determined the sensitivity and specificity of identifying melanomas during skin exams was 87.5 and 81.4%, respectively, for dermatologists, but only 79.9 and 70.9% for primary care physicians (25). Unfortunately, access to dermatologists for regular skin exams is neither uniform nor comprehensive, with over half of U.S. citizens having insufficient access to dermatologic care (26). Dermoscopy and reflectance confocal microscopy (RCM) are two non-invasive approaches that increase the number of melanomas biopsied, while reducing the number of benign lesions biopsied (27, 28). However, these techniques also require highly skilled interpretation, and use of RCM is largely limited to dermatologists who specialize in pigmented lesions. Dermoscopy is more broadly used in clinical practice, though it does not appear to increase diagnostic accuracy in the practice of non-experts (29) and has even been reported to reduce sensitivity of melanoma detection compared to clinical assessment alone (30). An objective biomarker could therefore improve the odds to biopsy a true melanoma, especially in situations where a specialized dermatologist and auxiliary equipment/techniques are not available. In contrast, any increases in certainty might be minimal when used by a highly specialized and experienced pigmented lesion dermatologist.

Ultimately, the cost of a skin biopsy to the individual patient, in terms of both morbidity and finance, is relatively low, but the cost of leaving an early-stage melanoma undiagnosed is exceptionally high. By applying the above theory of biomarker utility, the prior odds of malignancy can be estimated by considering the number of lesions biopsied to diagnose one melanoma. This ranges from ~2 for some dermatologists to ~30 for primary care physicians (31). Thus, according to current clinical practices, any lesions with prior odds lower than 1/2–1/30 of being malignant are less likely to be biopsied. A biomarker introduced at this stage could therefore provide utility if it conferred a significant degree of certainty that a lesion was not malignant and thus avoided biopsy. For certainty to be sufficiently increased such that the perceived odds of malignancy is reduced from 1/2 and 1/30 prior to the test to <1/30 after the test (posterior odds), an estimated likelihood ratio of 0.1 (LR-) is required (Figures 3A,B, green line). Conversely, for a melanocytic tumor eliciting a high suspicion of malignancy from conventional means (high prior odds), a biomarker is unlikely to change decision making, as the clinical threshold to perform the outcome (biopsy) has already been met (LR+ 1.0, Figures 3A,B, blue line). Although this analysis suggests that molecular biomarkers are unlikely to change clinical behavior when assessing solitary lesions, there are other clinical scenarios that challenge the assumptions and conclusions of this model. For example, when individual lesions present in sensitive areas, such as the face or genitals, the morbidity of biopsy and the clinical suspicion threshold required to biopsy are increased. In these cases, a modest LR– may suffice to shift the clinical decision from “biopsy” to “monitor.” Alternatively, due to increased morbidity and decreased patient tolerance of multiple skin biopsies, a patient with many suspicious lesions may benefit from a combination “rule in” and “rule out” molecular test (necessary characteristics estimated as LR+ = 5.0 and LR– 0.2, respectively, Figures 3C,D) to prioritize lesions to biopsy (32).

The pigmented lesion assay (PLA), developed by DermTech, is the only biomarker currently commercially available for clinical use addressing this decision (33, 34). This test is an adhesive patch-based gene expression test measuring expression of PRAME and LINC00518, with reported sensitivity of 91% and specificity of 69% (34). The non-invasive nature and high sensitivity of this test has allowed for clinical incorporation in certain specialized scenarios (32). However, a recent study concluded the test is neither widely nor routinely used by pigmented lesion experts, which is consistent with predictions of the above model (35). Beyond the test's performance characteristics, another limitation is the delay between sampling and result, which necessitates a subsequent biopsy visit if the test is positive. If a similarly sensitive point-of-care biomarker test could be developed as a rapid in-office test, it may be more widely adopted.

In summary, because the clinical default in a setting of uncertainty will usually be to biopsy an indeterminate lesion, biomarkers with a large LR– represent the opportunity to change clinical management by avoiding biopsy. For this purpose, a biomarker with maximum specificity, thus minimizing false negatives, is critical since the adverse consequences of missing the diagnosis is high and the morbidity and cost of a skin biopsy is relatively low. There are situations within a dermatology clinic where a biomarker with high sensitivity (LR+) may also prove clinically useful, but these are likely to be more specialized scenarios such as prioritizing lesions for biopsy in a patient with dysplastic nevus syndrome. Combined with practical limitations for such widespread use of non-invasive biomarkers used at this stage, including the ease of use, cost, and the timing of results, we speculate that current technologies for molecular profiling are unlikely to be influential in the routine evaluation of single lesions.

The established challenges of histopathologic analysis has motivated the search for biomarkers to assist in melanoma diagnosis. If prior odds based on original histology favor a more benign dysplasia, such as a dysplastic nevus, then it is possible that a moderate LR–, estimated at 0.2, may convince a clinician to monitor a biopsy site rather than pursuing an excision with margins (Figure 4). Conversely, if a biomarker applied to the same lesion demonstrated a significant LR+ of 10, then the result may convince a clinician to excise the area with larger margins (10–20 mm instead of 5 mm), despite relatively low prior odds. For a lesion with higher prior odds of possessing malignant potential but lacking bona fide features of melanoma, often termed an atypical melanocytic proliferation, a less compelling LR+ of 2 might be sufficient to upgrade a diagnosis to melanoma, whereas a significantly lower LR- of 0.1–0.01 might be required to decrease excision margins or avoid excision altogether. These analyses demonstrate that there could be great clinical utility of biomarkers examining the malignant potential of intermediate melanocytic lesions, particularly when the biomarker holds a significant likelihood ratio that modifies the prior odds by an order of magnitude or more.

Numerous biomarker tests have been developed to address this issue. Table 2 reports biomarker tests that are currently commercially available or used in selected referral centers and Table 3 (discussed below) reports additional candidate biomarker tests reported in the literature. Immunohistochemistry (IHC) biomarkers are usually the most easily adopted test with regards to tissue requirement, cost, and required laboratory equipment and expertise. One promising IHC marker is PRAME (Preferentially expressed Antigen in Melanoma), which has demonstrated an LR+ of 29–62.5 and an LR- of 0.13–0.253 when assessed with an independent validation cohort (39, 40). Considering that it shares 90% concordance with fluorescent in situ hybridization (FISH) and 92.7% agreement with the final diagnostic interpretation (40), adopting PRAME IHC could be an economical option for most laboratories to use for routine differentiation of nevi and melanoma. Comparatively, p16 IHC, which is used at some institutions to help distinguish between nevi and melanoma in challenging cases, has no published validation cohorts to identify the clinical characteristics of this test. As a result, the accuracy, reliability, and versatility of p16 is unclear. In addition, the utility of p16 expression depends on the type of lesion since it may help distinguish desmoplastic Spitz nevi from desmoplastic melanoma (37), but fails at distinguishing Spitz nevi from Spitzoid melanoma (38).

FISH is another tissue staining modality that has been used for over a decade to help distinguish nevi from melanoma. It has well-developed scoring criteria, several validation studies, has undergone quality improvement to reduce the number of erroneous Spitzoid classifications, and has been directly compared to both the myPath gene expression panel (GEP) and copy number variation (CNV) assay by comparative genome hybridization (CGH) (42, 43, 49–51). The clinical characteristics of FISH in melanoma diagnosis are highly competitive and cost-effective when compared to the other molecular techniques (Table 2) (43, 49–51). However, evaluating FISH staining requires expert interpretation and expensive microscopy equipment, which limits the adoption of the technique outside of reference laboratories.

The CNV by CGH assay has been used clinically for almost two decades and assesses large genomic mutations and deletions across the genome. Melanomas usually possess multiple chromosomal losses and gains (96.2% frequency), Spitzoid nevi frequently have a unique 11p amplification, and non-Spitzoid nevi usually have no detectable chromosomal changes (41). While a powerful technique to distinguish nevi from melanomas, there is no clear CNV-based definition of melanoma. As a result, the CNV cut-off that separates a benign vs. malignant lesion is often a gestalt decision from the dermatopathologists, which makes quantifying the clinical accuracy characteristics challenging. However, the price of CGH has fallen in recent years, the technology can be readily adopted by laboratories with experience in molecular diagnostics, and, it has 90% concordance with FISH in cases with unequivocal histopathology (49). In particularly challenging cases, CGH also outperformed both FISH and myPath in reaching concordance with a consensus histopathological diagnosis (51).

More recently, the Myriad myPath quantitative gene expression panel (GEP) and scoring algorithm was reported to have favorable test characteristics (AUC of 0.96, sensitivity of 90% and specificity of 91%) (44). This platform has subsequently been studied across prospectively submitted samples (45) in the evaluation of desmoplastic melanocytic proliferations (47) and in conjunction with long-term clinical outcomes (46). The most recent study is well-designed to assess function in histopathologically uncertain lesions and exemplifies how to use clinical outcome data as the gold standard instead of histopathology (48). Despite its impressive clinical characteristics, when assessed individually, the myPath methodology is not as competitive in comparison studies to FISH and CGH when histopathology is the gold standard (50, 51). Additional studies of the myPath test compared directly against FISH and CGH with long term follow up of clinical outcomes as the gold standard are warranted to determine the true relative utility of these tests.

In summary, the existing literature most highly supports the use of PRAME IHC, FISH, and CGH to aid dermatopathologists in routinely stratifying melanoma risk. While the clinical characteristics of CGH are uncertain, the presence of amplifications and deletions would likely prompt a conservative practitioner to upgrade the diagnosis to melanoma (presumed LR+ >2). FISH and PRAME IHC easily reach the modeled threshold of utility, even in challenging cases, but would not be sufficient to rule out melanoma (LR– is not <0.1–0.01) (50, 51, 63), and it is unclear if a CGH result revealing no amplifications or deletions would have a sufficient LR- to overturn a conservative melanoma diagnosis. The only available biomarker that could potentially have such a robust LR– is myPath, but such performance would need to be proven in a head-to-head comparison with the other testing modalities. As it stands, there is no currently used biomarker with a robust enough LR+ and LR– to be definitive for including or excluding a melanoma diagnosis with certainty, which leaves room for the further development of novel tests and more rigorous validation of current tests.

Since distinguishing early invasive melanoma from pre-malignant melanocytic lesions by routine histopathology remains a diagnostic challenge (15), numerous research groups have attempted to develop ancillary tests to augment melanoma diagnosis. Presented below and summarized in Table 3 are a set of studies that examined promising biomarkers. All of these studies utilize methodologies obtainable by academic reference laboratories, all have at least one test characteristic defined, and most have been assessed in at least one independent test cohort.

One study used IHC to develop a semi-quantitative scoring system based on the percent expression of Ki-67 (MIB-1), p16, and HMB45 expression in melanocytes. Of these commonly available stains, Ki-67 has the best individual sensitivity and specificity, but the addition of p16 and HMB45 increased this to 97.4% sensitivity and 97.3% specificity with an AUC of 0.987. Notably, the IHC panel accurately predicted the metastatic evolution of tumors that were initially classified as melanocytic or Spitzoid nevi, and it appropriately downgraded one melanoma to a benign nevus (62). While these results are impressive, this test requires stain scoring which is notoriously subjective. Therefore, maintaining a high interobserver concordance with three IHC stains could be challenging, especially with scoring borderline p16 and HMB45 cases. Another proposed diagnostic marker is 5-hydroxymethylcytosine (5-hmC), which is decreased in melanoma compared to nevi, keratinocytes, and lymphocytes. Staining variation limits the sensitivity and specificity of 5-hmC scoring system to 92.7 and 97.7%, respectively, though there is a characteristic biphasic staining pattern between the melanoma and nevi components of a transformed nevus that could be clinically useful according to a pilot study (61).

Melanomas lose primary cilia after transformation whereas they are retained in benign nevi (64–66). The presence of primary cilia have been shown to differentiate typical and atypical Spitzoid nevi from malignant tumors with a sensitivity of 81% and specificity of 65% (AUC 0.84), but, with the incorporation of cytologic atypia, hyperchromatism, and asymmetry in making a diagnosis, the combined specificity and AUC were increased (sensitivity 72%, specificity 89%, AUC 0.92) (60). This is an improvement over standard diagnosis of Spitzoid tumors, which are classically difficult to distinguish between benign and malignant subtypes. Recent pilot studies have also suggested visualization of primary cilia using IHC may also provide similar utility in other melanoma types (67–69).

Discrimination between nevus and melanoma may also be accomplished using differences in gene expression. The most accessible method of analysis is to perform quantitative real time polymerase chain reaction (qRT-PCR) on a single transcript target from tissue or serum. For instance, silver homolog (also called SILV, PMEL, gp100, or ME20M) expression from FFPE tissue is significantly lower in melanomas compared to nevi, and it has favorable characteristics (AUC 0.94) as a single biomarker (57). Similarly, a qRT-PCR panel for 6 microRNAs (miRNA) in serum demonstrated positive and negative likelihood ratios of 20 and 0.07, respectively, in distinguishing uveal melanoma and nevi (59). Additional work remains to be completed to confirm clinical utility of these studies as neither contained validation cohorts and both had small sample sizes. The importance of a validation cohort is highlighted by our 2019 study where we developed a miRNA ratio trained model from miRNAs enriched in microdissected melanoma cells compared to nevus cells. While the initial training model had an AUC of 0.98, the clinical accuracy of the qRT-PCR validation cohort containing whole FFPE sections dropped to a sensitivity of 81%, specificity of 88%, and AUC of 0.90 (56). Likewise, a microarray study on microdissected FFPE tissue found that the expression profiles of 36 genes could be used to identify melanoma with sensitivity (90%) and specificity (88%) in the validation cohort (58). While the gene expression panels described above have test characteristics that are theoretically compatible with a useful clinical test, their reliability is uncertain until larger studies confirm clinical utility.

Of the different classes of proposed biomarkers, DNA methylation testing is not widely available outside of reference laboratory testing for X chromosome disorders and specific genes, such as MLH1 and MGMT. In melanoma, promoter methylation status of two genes have been examined as potential biomarkers. The Ras associated domain family 1 isoform A (RASSF1A) gene is a tumor suppressor found to be methylated in melanomas compared to healthy subjects (AUC of 0.905) by cell-free DNA (cfDNA) analysis (55). While cfDNA allows for a non-invasive blood draw and detects in situ melanomas with an AUC of 0.945, no studies have been published validating the RASSF1A marker. In comparison, the promoter for CLDN11, a component of the tight junction, undergoes CpG island methylation in approximately 50% of primary melanomas compared to 3–6% of dysplastic nevi. CLDN11 promoter methylation status was determined by extraction from FFPE tissue and methylation-specific PCR in a large cohort of melanomas and nevi. This yielded a 52% sensitivity and 94% specificity (AUC of 0.806), suggesting utility for ruling in melanoma in the context of a suspicious dysplastic nevus. Another group developed a CpG site microarray panel and, ultimately, a 40-CpG next generation sequencing (NGS) panel to assess the methylation of status of multiple genes from microdissected FFPE slides. While both of these studies had characteristics compatible with a useful clinical test (sensitivity up to 96.6%, specificity up to 100%, AUC up to 0.996), they are limited by small cohort sizes and lack of long-term patient outcome data (52, 53).

Overall, there are several potential biomarkers that could be used to distinguish early melanoma from dysplastic nevi. All of the above tests have LRs (Table 3) that could be beneficial in distinguishing atypical melanocytic proliferations from melanoma based on our model (Figure 4). Of these biomarkers, we speculate that the most accessible, least complex, minimally invasive, and relatively affordable test that could be readily adopted into the laboratory is 5-hmC IHC (61). Prior to its adoption, however, large cohort studies are required to confirm the promising initial results. Of all of the tests included in Tables 2, 3, the 40-CpG classifier next generation sequencing test for DNA methylation has the best overall combined characteristics (52). If future multi-institutional studies show good concordance and validation of the test characteristics, the 40-CpG classifier test could be useful for assessing the most challenging cases of atypical melanocytic proliferations. However, the major limitation of this test is the NGS approach: ample tumor tissue will be required, the method is expensive, and turn-around-time will likely be at least 8–14 days. Thus, other tests that require less tissue, cost, time, and expert interpretation, such as the multi-IHC panel or ciliation index, also warrant further consideration (62, 68).

In addition to indications for biopsy and diagnostic accuracy, potential biomarkers were assessed for other important clinical decision points related to the staging and treatment of invasive melanomas (Figure 1C, Table 4).

For Stage III melanomas that have metastasized to sentinel lymph nodes and for those in which nodal or in transit metastatic disease has been surgically resected, adjuvant therapy using anti PD-1 immunotherapy or BRAF/MEK inhibition is currently recommended for consideration based on recent clinical trial data showing improved relapse-free survival (RFS) (91, 108). There is significant interest in using adjuvant therapy in Stage II high-risk melanomas, as evidenced by a large, double-blind, prospective, randomized clinical trial currently evaluating the safety and efficacy of immune checkpoint blockade in resected Stage II melanoma (KEYNOTE-716: https://clinicaltrials.gov/ct2/show/NCT03553836). Recently an 11-gene GEP was published that demonstrated the ability to stratify patients with stage II melanoma into high GEP score groups and low GEP score groups, with the latter having significantly higher melanoma specific survival and relapse free survival rates (109). The authors of this study speculate that patients with higher GEP scores may be better candidates for adjuvant therapy due to the increased risk for relapse and death, potentially reflecting ambitions to test this hypothesis in a future study (109).

If an actionable BRAF V600 mutation is identified by sensitive and specific molecular testing (110), then BRAF plus MEK inhibitor therapy is an option for first-line therapy in resected Stage III and in Stage IV melanoma patients (90). Regardless of BRAF mutation status, immune checkpoint inhibitors (ICI) against PD-1 and CTLA-4 are frequently used due to improved progression-free and overall survival compared to both chemotherapy and combined BRAF/MEK inhibition (111–113). However, ~65% of Stage IV metastatic melanomas do not significantly respond to ICI therapy (114–116), indicating an unfilled niche for future biomarkers to discriminate responders from non-responders prior to therapy initiation. Furthermore, accurately predicting which Stage III, Stage II, and even Stage I patients, who comprise the majority of cases of melanoma specific mortality (117), have the potential to benefit from adjuvant therapy is a major goal of predictive biomarkers in melanoma.

With the development of new ICIs, kinase inhibitors, combination therapy regimens, and other novel treatment modalities, there will be an increasing need for biomarkers to match patients to optimal adjuvant therapies. Successful predictive biomarkers will have the potential to decrease patient morbidity by avoiding side-effects of less helpful treatments, boost response rates by identifying the most appropriate therapy upfront, and increase the cost-effectiveness of treatment by avoiding expensive ICIs when not indicated. When there are multiple treatment options considered first- or second-line treatments, there is even more value in the role of biomarkers guiding therapy selection, however the dichotomous framework that we have presented for other key decision-making points in the treatment of melanoma is less applicable. The complex decisions surrounding adjuvant therapy selection would necessitate multi-dimensional LR+/LR- analyses not attempted in this review due to exponentially increasing complexity. We await the results of ongoing studies and clinical trials in this field and refer the reader to several recent reviews that cover the impressive breadth of proposed biomarkers for use with ICI (118–121).