Melanoma is cancer of the melanocytes — the cells that make pigment. Usually it arises in skin, and it is the dangerous one among the common skin cancers, for a single reason: it can spread while it is still small.
That single asymmetry organises the whole disease. A basal-cell carcinoma can sit on a face for years and be cut out. A melanoma a few millimetres across can already have seeded the lymph nodes. Which is why the most important number in melanoma pathology is not how wide the lesion is but how deep it goes.
The second thing to know is that melanoma is where modern cancer immunotherapy was proven. Metastatic melanoma was, until about 2010, close to untreatable — a diagnosis measured in months. Within a decade two entirely different approaches, targeted inhibition of a mutated kinase and blockade of the brakes on T cells, had both extended survival, and the second produced something oncology had rarely seen: a substantial group of patients with advanced disease still alive years later, off treatment. Melanoma led because it is unusually visible to the immune system, and what worked here was then carried into lung, kidney, bladder and head-and-neck cancer.
Start here: ultraviolet light, and what it does to DNA
Ultraviolet radiation — chiefly UV-B — damages DNA directly, producing cyclobutane pyrimidine dimers and 6-4 photoproducts, along with strand breaks; cells counter these with photoreactivation, base- and nucleotide-excision repair and mismatch repair [1]. Melanoma is what happens when that damage lands in a melanocyte, escapes repair, and accumulates.
But the relationship between sun and melanoma is not simple, and a 2005 study showed why in an unusually clean way. Comparing 126 melanomas from four groups differing in UV exposure — skin with chronic sun-induced damage, skin without it, acral sites (palms, soles, under nails) and mucosal sites — the frequencies of BRAF mutation and of genomic copy-number change differed significantly between groups, and samples could be classified into the four groups with 70 percent accuracy from copy-number changes alone [2]. These are genetically distinct diseases that share a cell of origin. Melanomas on chronically sun-damaged skin are not the same molecular entity as melanomas on intermittently exposed skin, and acral and mucosal melanomas — which occur on sites that get almost no sun — are different again.
The order in which the mutations arrive has also been read directly. Sequencing 293 genes across 150 areas of 37 primary melanomas and their adjacent precursor lesions found that unequivocally benign lesions carried BRAF V600E exclusively; intermediate lesions were enriched for NRAS and additional drivers; TERT promoter mutations were present in 77 percent of intermediate lesions and melanomas in situ, appearing much earlier than expected; biallelic CDKN2A loss emerged only in invasive melanoma; and PTEN and TP53 mutations only in advanced primaries [3].
That is a staged progression written into the genome — and it explains something clinically important. A common mole carries the same BRAF V600E mutation as a melanoma. The mutation is not the disease. What separates them is everything that comes after.
Pillar 1: measurement and diagnosis
Recognition, and where it is going
Melanoma is one of the few cancers a person can see. Recognition rests on noticing a pigmented lesion that is changing, or that looks unlike the others on the same skin, and the practical route to diagnosis is excisional biopsy of the suspicious lesion so the pathologist can measure the whole thing.
The most striking recent development is that this visual task turns out to be learnable by machines. A convolutional neural network trained on clinical images achieved dermatologist-level classification of skin cancer [4]. A follow-up asked a better question — not machine versus clinician but machine plus clinician. Using 11,444 dermoscopic images to train a CNN, then having 112 dermatologists from 13 German university hospitals and the network independently classify 300 biopsy-verified lesions into five categories, the combined classifier outperformed either alone [5]. That framing is the right one, and it is the one most likely to reach patients.
An honest gap: this review's substrate was assembled by citation crawl during an OpenAlex outage, seeded from the immuno-oncology literature. It covers the molecular and therapeutic literature densely and the clinical dermatology literature — the ABCDE rule, dermoscopy training, screening programmes, excision margins — thinly. Where this review is quiet on those, it is because the sources were not recoverable, not because they do not exist.
The pathology that decides everything
Once the lesion is out, the pathology report drives the rest of care. Two revisions of the AJCC staging system, built on very large datasets, established which features matter.
The 2001 analysis pooled 17,600 patients with complete data from thirteen centres and cooperative groups and found that in the T category tumour thickness and ulceration were the most powerful predictors of survival, with level of invasion mattering only within thin (≤1 mm) melanomas; in the N category, the number of metastatic nodes, whether nodal disease was clinically occult or apparent, and ulceration of the primary; and in the M category, that non-visceral metastases carried better survival than visceral [6] [7]. The 2009 revision, from a multivariate analysis of 30,946 patients with stage I–III disease and 7,972 with stage IV, reaffirmed thickness, mitotic rate and ulceration as dominant in localised disease and made mitotic rate the criterion defining T1b [8]. An earlier critical reanalysis had already found that the best statistical fit for thickness cut-offs was at 1 versus 2 versus 4 mm [9].
The eighth edition refined this further: thickness recorded to the nearest 0.1 mm rather than 0.01; T1a redefined as under 0.8 mm without ulceration and T1b as 0.8–1.0 mm or under 0.8 mm with ulceration, with mitotic rate dropped as a T criterion; "microscopic" and "macroscopic" nodal disease renamed clinically occult and clinically apparent; stage III split into four subgroups; and a new M1d category for central nervous system metastasis [10]. It drew on more than 46,000 patients from ten centres.
Breslow thickness is the depth of invasion in millimetres. It is the single most important prognostic measurement in localised melanoma, and the figure below is about why.
Sentinel lymph node biopsy
The technique was developed in 1992: inject vital dye at the primary site, identify the first node on the direct drainage pathway, remove it, and examine it for occult tumour. In the original series the sentinel node was identified in 194 of 237 basins, and metastases were found in 21 percent — visible on routine staining in 12 percent and only on immunohistochemistry in a further 9 percent [11]. The concept is that lymphatic spread is orderly, so the first node tells you about the rest.
It became the standard nodal staging procedure, and its prognostic power is very large: in a multi-institutional series of 580 successful mappings, sentinel node status was the most significant prognostic factor for both disease-free and disease-specific survival, and thickness and ulceration added no further prognostic information once the node was known to be positive [12]. Two further points have since been settled: what predicts a positive node — thickness, ulceration, male sex and absent tumour-infiltrating lymphocytes, the last being strong enough that a brisk lymphocytic infiltrate dropped the probability of a positive node to 3.9 percent against 26.2 percent when lymphocytes were absent [13] [14] [15] — and what to do about it, since completion lymph-node dissection after a positive sentinel node was not associated with increased melanoma-specific survival in 1,934 randomised patients [16]. The node is measured, not necessarily cleared.
Molecular testing
For anyone with advanced disease, BRAF mutation status is now a treatment-selecting test. BRAF mutations were catalogued across human cancers in 2002 [17], with the same group reporting activating mutations in 66 percent of melanomas [18] — a higher figure than the advanced-disease series above, and the range across studies is real. The reason to test is in Pillar 2.
Centerpiece: a simple simulatable model of why depth decides
The AJCC survival-by-thickness tables are not carried in any abstract in this substrate, so this figure does not fit a thickness curve — that would mean inventing numbers. Instead it takes the mechanism by which depth kills and decomposes survival through it.
Depth matters because depth buys access to the lymphatics. A melanoma confined to the epidermis has no route out; one that has invaded far enough to reach dermal lymphatic vessels does. So the prognostic weight of Breslow thickness should run almost entirely through one quantity: the probability that occult tumour has already reached the regional nodes.
MSLT-I, the randomised trial of sentinel-node biopsy versus nodal observation in 1,269 patients with intermediate-thickness (1.2–3.5 mm) melanoma, reports every term needed to test that [19]. Five-year melanoma-specific survival was 90.2 ± 1.3 percent with a tumour-negative sentinel node and 72.3 ± 4.6 percent with a positive one. Sentinel micrometastases were present in 16.0 percent (122 of 764).
The model is one line with no free parameters:
S(p) = p × 72.3 + (1 − p) × 90.2 = 90.2 − 17.9p
Survival falls linearly in the probability of nodal spread, at 17.9 percentage points per unit — the whole prognostic weight compressed into one number.
The check the arithmetic was never given. MSLT-I reports its two stratum survivals and its positivity rate in one place and its population five-year melanoma-specific survival somewhere else, as 87.1 ± 1.3 percent. Those three numbers were never combined in the paper. Doing so gives 0.160 × 72.3 + 0.840 × 90.2 = 87.34 percent, against the reported 87.1 — a quarter of a percentage point, with nothing fitted.
A second check, from the trial's own design. The two randomised arms detect occult nodal disease by completely different routes: immediate pathological examination of a sentinel node, or waiting for the nodes to declare themselves clinically. Sentinel positivity was 16.0 percent (122/764); nodal relapse in the observation arm was 15.6 percent (78/500) [19]. Half a percentage point apart. That agreement is what makes the sentinel node credible as a measurement of biology rather than an artefact of looking very hard with immunohistochemistry — a real concern given that the original series found 9 percent of positives only on immunostaining [11].
And an external one. Four independent series report sentinel positivity of 14.7 percent (85/580) [12], 16.0 percent [19], 17.6 percent (156 of 887 mapped) [13] and 22.1 percent (252/1,138, melanomas ≥ 0.75 mm) [14] — a narrow band across two decades and different institutions, bracketing the trial figure without ever having been fitted to it.
The teaching point. Because survival is linear in nodal-spread probability, and nodal-spread probability rises with depth, catching a melanoma thin is not a marginal gain — it moves the patient along a straight line toward the 90.2 percent end of it. Excision of a thin melanoma is genuinely curative. Every millimetre of delay is depth, and depth is probability of spread.
Three honest limits. This is a decomposition, not a thickness curve — the model shows how depth acts, not the quantitative relationship between millimetres and survival, because the source for that was not recoverable here. MSLT-I enrolled only intermediate-thickness melanomas, so the stratum survivals are conditioned on that band and would not be identical in a thin or thick population. And the linearity is a property of the law of total probability given two fixed strata; real nodal disease is graded rather than binary, and sentinel-node tumour burden itself carries prognostic information [20].
One further finding deserves stating because it constrains what the figure means. MSLT-I improved five-year disease-free survival (78.3 versus 73.1 percent) but melanoma-specific survival was similar between the arms (87.1 versus 86.6 percent) [19]. Knowing about the nodal disease earlier is powerful prognostic information; it is not, by itself, a cure. The figure is a model of prognosis, not of benefit from the procedure that measures it.
Pillar 2: treatment
Early disease
For localised melanoma the treatment is wide local excision of the primary [21], and the prognosis is the reason the rest of this section exists: five-year melanoma-specific survival with a tumour-negative sentinel node was 90.2 percent [19]. Everything above is about identifying whose disease is not localised.
The first revolution: targeting the mutated kinase
Roughly half of advanced melanomas carry an activating BRAF mutation — the figure quoted in the MEK-inhibitor trials is 50 percent [22] [23], and a clinicopathologic series of melanoma patients found 48 percent mutated, of whom 74 percent carried V600E, 20 percent V600K and 6 percent other genotypes [24]. The mutation locks the MAP kinase pathway on. Inhibiting the mutant kinase produced responses in a proof-of-concept study [25], and then a phase 3 trial of vemurafenib against dacarbazine in 675 patients with previously untreated BRAF V600E-mutant metastatic melanoma found six-month overall survival of 84 percent versus 64 percent, with a 63 percent relative reduction in the risk of death and 74 percent in the risk of death or progression [26]. For a disease with no effective systemic therapy, that was extraordinary.
The problem was durability: responses were dramatic and then resistance emerged, generally within months. The answer was to block the pathway at two points. Combining a BRAF inhibitor with a MEK inhibitor downstream improved on BRAF inhibition alone [27] [28] [29] [22], and in a phase 3 trial of 704 patients, dabrafenib plus trametinib against vemurafenib gave 12-month overall survival of 72 versus 65 percent (HR 0.69) with median progression-free survival of 11.4 months [30] [31]. Pooled five-year data from 563 patients across COMBI-d and COMBI-v showed progression-free survival of 21 percent at four years and 19 percent at five [32]. Encorafenib plus binimetinib is a further combination [33].
So targeted therapy works fast, works only in BRAF-mutant disease, and holds long-term in a minority. That last figure — roughly one patient in five progression-free at five years — is a real plateau, and it is why the sequencing question against immunotherapy matters.
The second revolution: releasing the brakes on T cells
That melanoma is visible to the immune system was established long before it could be exploited: a gene encoding an antigen recognised by cytolytic T lymphocytes on a human melanoma was identified in 1991 [34], and further melanoma antigens recognised by tumour-infiltrating lymphocytes followed [35]. The therapeutic idea came from mouse work in 1996 showing that blocking CTLA-4 enhanced antitumour immunity [36] — CTLA-4 being a brake on T-cell activation rather than anything the tumour makes.
Ipilimumab, an anti-CTLA-4 antibody, was the first therapy ever to improve overall survival in metastatic melanoma. In 676 previously treated patients, median overall survival was 10.0 months with ipilimumab plus a gp100 peptide vaccine and 10.1 months with ipilimumab alone, against 6.4 months with the vaccine alone (HR 0.68) [37]; it was subsequently combined with dacarbazine in untreated disease [38]. The headline numbers look modest. What mattered was the shape of the curve rather than its median — a tail of patients who did not relapse.
Anti-PD-1 antibodies followed, and were better. PD-1 is a second brake, engaged in the tumour microenvironment; the first-in-human study of anti-PD-1 across cancers showed activity and immune correlates [39], as did anti-PD-L1 [40], with melanoma-specific results shortly after [41] [42]. Nivolumab beat chemotherapy in untreated BRAF wild-type melanoma [43], and pembrolizumab beat ipilimumab head-to-head [44] [45] [46].
Combining both brakes produced the largest effect and the clearest evidence of durability. In CheckMate 067, at a minimum follow-up of 60 months, median overall survival was not reached (more than 60 months) with nivolumab plus ipilimumab, 36.9 months with nivolumab alone, and 19.9 months with ipilimumab alone (HR 0.52 for the combination against ipilimumab) [47] [48] [49] [50]. A median overall survival that has not been reached at five years, in a disease whose median was measured in months a decade earlier, is the single most consequential number in this review. The cost is toxicity. Releasing a brake on T cells releases it everywhere, and immune-related adverse events follow: endocrine events after ipilimumab have been characterised specifically [51], and treating ipilimumab-induced hypophysitis with high-dose glucocorticoids has itself been associated with worse outcomes [52].
An oncolytic viral therapy, talimogene laherparepvec, is also approved and has been combined with checkpoint blockade [53] [54].
Moving treatment earlier
If checkpoint blockade works in metastatic disease, it should work better with less disease to clear. Adjuvant therapy after complete resection has been tested extensively. Interferon alfa-2b was the old standard [55]. Adjuvant ipilimumab improved recurrence-free and then overall survival against placebo in resected stage III [56] [57]. Adjuvant nivolumab beat adjuvant ipilimumab in 906 patients with resected stage IIIB, IIIC or IV disease [58]. Adjuvant pembrolizumab against placebo in 1,019 patients with resected stage III gave one-year recurrence-free survival of 75.4 versus 61.0 percent [59]. And for BRAF-mutant stage III, adjuvant dabrafenib plus trametinib is an alternative [60].
Neoadjuvant therapy — giving the drug before surgery rather than after — is the more interesting development, and melanoma is where it was tested most rigorously. Neoadjuvant checkpoint blockade in high-risk resectable disease [61] and a direct comparison of neoadjuvant against adjuvant ipilimumab plus nivolumab in macroscopic stage III [62] were reported together, with a subsequent trial identifying an optimal combination dosing schedule [63]. The rationale is immunological rather than surgical: with the tumour still present, there is antigen for the primed T cells to recognise, so the immune response generated is broader.
Cancer vaccines
Personalised neoantigen vaccines target the mutations unique to one patient's tumour. An immunogenic personal neoantigen vaccine was reported in melanoma patients in 2017 [64], shown to induce persistent memory T-cell responses and epitope spreading [65], and combined with anti-PD-1 in a phase Ib trial [66]. A dendritic-cell vaccine was shown to broaden the diversity of neoantigen-specific T cells [67], and neoantigen-specific lymphocytes can be identified prospectively in peripheral blood [68].
The mRNA route has a longer history in melanoma than the recent coverage suggests: direct injection of protamine-protected mRNA was trialled in melanoma patients in 2009 [69], and mRNA-electroporated dendritic cells in 2013 and 2016 [70] [71]. The current generation combines an individualised mRNA vaccine with a checkpoint inhibitor, on the logic that the vaccine supplies the target and the checkpoint inhibitor removes the brake. This is the most-watched area in the field, and the honest position is that it is promising and not yet established.
Pillar 3: what is unresolved
Who responds, and why. Checkpoint blockade produces durable benefit in a substantial minority and nothing in the rest, and predicting which is unsolved. Mutational load is associated with benefit from CTLA-4 blockade but is not sufficient alone [72]; genomic and transcriptomic features of response to anti-PD-1 have been characterised [73]; tumour mutational burden associates with outcome across solid tumours [74] [75]; PD-L1 expression on tumour cells is prognostic [76] and colocalises with inflammatory infiltrate [77]; myeloid-derived suppressor cells [78] and neutrophil-to-lymphocyte ratio [79] have been proposed. Three papers published together in 2017 reported that the gut microbiome modulates response to PD-1 blockade in melanoma patients [80] [81], with earlier work linking the intestinal microbiome to checkpoint-blockade colitis risk [82] and to CTLA-4 blockade efficacy in mice [83]. None of these is yet a test that changes practice.
Sequencing. For a patient with BRAF-mutant metastatic melanoma, both targeted therapy and immunotherapy work, by unrelated mechanisms. Which to give first — fast, reliable, temporary responses against slower, less reliable, potentially permanent ones — is a genuine clinical dilemma the trials above do not settle.
The immunologically cold subtypes. Acral and mucosal melanomas are molecularly distinct [2], carry BRAF mutations far less often, and respond poorly to the therapies that transformed cutaneous disease. The revolution described here has largely bypassed them.
Prevention and early detection. Since survival is linear in nodal spread and nodal spread rises with depth, the largest available gain is detecting melanoma thinner. Whether population screening achieves that without large-scale overdiagnosis is not settled here, and the substrate does not cover it well enough for this review to claim otherwise.
Dig deeper in lmmol
Checkpoint blockade was proven in melanoma and then exported, which is why the oncology siblings are worth reading together. Lung cancer is where it moved next and now shapes first-line treatment; both diseases also illustrate the same targeted-therapy logic of testing for a driver mutation before choosing a drug. Breast cancer and prostate cancer offer the sharpest contrast on early detection — both have organised screening programmes and an overdiagnosis debate, while melanoma has neither in most countries despite depth being decisive; the sentinel-node concept itself was developed in parallel in melanoma and breast cancer [11]. Colorectal cancer shares the mutational-burden story, since it is the mismatch-repair-deficient tumours with high burden that respond to checkpoint blockade [74]. And psoriasis is the instructive inverse: the same skin, and a disease treated by suppressing immune pathways indefinitely, against a cancer treated by releasing them — which is also why checkpoint inhibitors so often cause autoimmune toxicity [51]. The full collection is at health.