Knowledge hub

AI and Melanoma: From Earlier Detection to Personalised Treatment

Artificial intelligence is beginning to support different stages of the melanoma journey—from identifying potentially concerning skin changes to selecting patient-specific targets for experimental cancer vaccines. This guide explains what AI can currently do, what remains under investigation and where human clinical oversight remains essential.

Written by ScanSkinAI Editorial TeamReviewed by the ScanSkinAI Clinical Review TeamPublished Last updated

Executive summary

  • AI can support skin-risk screening and lesion monitoring, largely by making it easier to look at the same spot consistently over time.
  • AI screening does not diagnose melanoma. It produces risk awareness and a suggested next step.
  • Dermatologists and laboratory testing remain responsible for diagnosis, normally through dermoscopic examination and, where indicated, biopsy and histopathology.
  • Separately, AI is being investigated for personalised melanoma treatment design.
  • Some experimental systems analyse tumour mutations and rank neoantigens for inclusion in an individual cancer vaccine.
  • These vaccines remain investigational unless and until approved by the appropriate regulator.

The melanoma care journey

Awareness → AI-supported screening → Clinical assessment → Diagnosis → Personalised treatment and monitoring. AI can contribute at several of these stages; a qualified healthcare professional is required at the ones that decide what happens to you.

  1. 1. Awareness

    Where AI may contribute: AI-assisted education and risk questionnaires can help someone understand personal risk factors and what a concerning change looks like.

    Human oversight: No clinician needed to learn — but risk information is general, not personal medical advice.

  2. 2. AI-supported screening

    Where AI may contribute: Computer-vision models analyse a photograph of a lesion and return a risk-awareness output plus a suggested next step.

    Human oversight: A screening output is not a diagnosis. Any flagged or persistent change should be shown to a clinician.

  3. 3. Clinical assessment

    Where AI may contribute: Triage tooling and stored photo history can help a clinician see how a lesion has behaved between appointments.

    Human oversight: A qualified clinician takes the history, examines the skin and performs dermoscopy.

  4. 4. Diagnosis

    Where AI may contribute: Research systems assist with dermatopathology workflows, but they do not replace laboratory reporting.

    Human oversight: Diagnosis is made by clinicians, normally with biopsy and histopathology.

  5. 5. Personalised treatment and monitoring

    Where AI may contribute: Investigational systems analyse tumour sequencing to rank candidate neoantigens for individualised vaccines; photo monitoring supports follow-up.

    Human oversight: Treatment selection, trial eligibility and recurrence surveillance are decided by an oncology team.

AI-supported melanoma screening

Computer-vision systems analyse a photograph of a lesion and score visual characteristics that clinicians also consider. The signals below are the ones most image-based systems work with.

Asymmetry

Whether one half of the lesion mirrors the other.

Border irregularity

Notched, scalloped or poorly defined edges.

Colour variation

Multiple shades of brown, black, red, white or blue-grey in one lesion.

Visible structural patterns

Network, streak, dot and blotch patterns visible in the image.

Change recorded over time

Differences between the same lesion photographed weeks or months apart.

Difference from your other lesions

The "ugly duckling" idea — a spot that does not resemble your usual pattern.

Image-based AI cannot confirm or exclude melanoma. A low-risk result does not mean a lesion is harmless, and a flagged result does not mean it is cancer. Both are prompts to act, not answers.

AI screening vs dermatologist diagnosis

These are complementary, not competing. One lowers the cost of looking; the other decides what a finding means.

  • What it produces

    AI screening
    Risk awareness and a suggested next step
    Clinical diagnosis
    A clinical diagnosis and a treatment pathway
  • Inputs used

    AI screening
    One or more photographs, plus any answers you provide
    Clinical diagnosis
    Medical history, symptoms, full-skin examination and dermoscopy
  • Confirmation

    AI screening
    Cannot confirm or exclude melanoma
    Clinical diagnosis
    May require biopsy and histopathology to confirm
  • Access

    AI screening
    Remote, on demand, from a phone browser
    Clinical diagnosis
    Appointment-based, in person or via teledermatology
  • Tracking over time

    AI screening
    Supports longitudinal photo comparison between visits
    Clinical diagnosis
    Documents change at each clinical review
  • Responsibility

    AI screening
    Supports prioritisation only
    Clinical diagnosis
    Holds clinical responsibility for the diagnosis

If you want a human opinion on a photograph without waiting for an in-person appointment, a photo-based dermatologist review sits between the two.

How personalised melanoma vaccines work

This is an area of active research, not a treatment you can request today. In outline, the process runs as follows.

  1. 1

    A sample of the patient's tumour and their healthy cells are sequenced.

  2. 2

    Mutations that are specific to the tumour, and expressed by it, are identified.

  3. 3

    AI or predictive algorithms rank the mutation-derived protein fragments — called neoantigens — by how likely they are to be presented by that person's immune system.

  4. 4

    A limited set of selected neoantigens is encoded into a vaccine manufactured for that individual.

  5. 5

    The vaccine aims to train the immune system to recognise cancer cells carrying those targets, usually alongside other therapy such as a checkpoint inhibitor.

The AI does not manufacture the immune response or cure the cancer. It helps researchers select potential targets for inclusion in the vaccine. Predicting which targets will work is genuinely hard: strong predicted binding does not guarantee that a tumour cell presents that peptide, and tumour heterogeneity and immune escape can defeat apparently good targets.

Current leading programmes

The table below summarises publicly documented programmes as of 22 August 2026. Where a figure comes from a company announcement rather than a peer-reviewed publication, that is stated. None of these therapies is commercially approved.

  • Intismeran autogene (previously V940 / mRNA-4157)

    Organisation
    Moderna and Merck
    Format
    Personalised mRNA
    Cancer studied
    Melanoma (adjuvant, after resection)
    Computational selection
    Developer describes integrated AI algorithms that process tumour and blood sequencing and predict up to 34 patient-specific neoantigens.
    Clinical phase
    Phase 3 (INTerpath-001); randomised Phase 2b KEYNOTE-942 completed
    Evidence strength
    Strongest in the class. Randomised Phase 2b published, plus a company announcement on 19 August 2026 that Phase 3 recurrence-free and distant-metastasis-free survival endpoints were met.
    Limitations
    Phase 3 data are topline company announcements only — detailed effect sizes, subgroups and full safety data were not public as of 22 August 2026. Investigational; not approved.
  • EVX-01

    Organisation
    Evaxion
    Format
    Personalised peptide vaccine
    Cancer studied
    Advanced melanoma
    Computational selection
    PIONEER / AI-Immunology platform selects patient-specific neoantigens; a peer-reviewed Phase 1 documents the AI-based selection.
    Clinical phase
    Phase 2 (NCT05309421), single-arm
    Evidence strength
    The clearest published link between AI-predicted targets and measured T-cell responses. Phase 2 response and biomarker figures are largely company disclosures and conference material.
    Limitations
    Small, single-arm study; historical comparisons cannot remove selection bias. Investigational.
  • Autogene cevumeran (BNT122 / RO7198457)

    Organisation
    BioNTech and Genentech
    Format
    Personalised mRNA-lipoplex
    Cancer studied
    Melanoma and other solid tumours
    Computational selection
    Developer describes predictive algorithms that identify multiple patient-specific neoantigens for an individualised mRNA product.
    Clinical phase
    Randomised development across several indications
    Evidence strength
    Credible, mature platform. Public melanoma outcome evidence and algorithm-specific validation currently trail intismeran autogene.
    Limitations
    Indication-specific efficacy still developing; the AI contribution is not isolated from the rest of the system.
  • GNOS-PV02

    Organisation
    Geneos
    Format
    Personalised DNA vaccine (up to 40 neoantigens)
    Cancer studied
    Advanced hepatocellular carcinoma (liver), not melanoma
    Computational selection
    Neoantigens selected from tumour DNA/RNA and matched normal sequencing; less clearly positioned publicly as an AI-first platform.
    Clinical phase
    Phase 1/2, 36 patients, combined with pembrolizumab and IL-12
    Evidence strength
    Peer-reviewed clinical signal reported in Nature Medicine, including an objective response rate near 30% with some complete responses.
    Limitations
    Early, small and non-randomised; outside melanoma, so not directly transferable.
  • NeoVax / NeoVaxMI

    Organisation
    Dana-Farber and the Broad Institute
    Format
    Personalised long-peptide vaccine
    Cancer studied
    Melanoma
    Computational selection
    Next-generation sequencing with machine-learning algorithms for neoantigen identification.
    Clinical phase
    Phase 1; NeoVaxMI (2025) used a multi-adjuvant formulation
    Evidence strength
    Scientifically influential immune proof-of-concept, including evidence of durable immune memory.
    Limitations
    Designed mainly to establish safety, feasibility and immunogenicity rather than comparative efficacy. Limited comparative outcome evidence.

Evidence descriptions reflect public information available on 22 August 2026 and are not a head-to-head comparison. Proprietary algorithms are not disclosed in enough detail for independent technical comparison. Any figures should be re-verified against the primary sources before republication.

mRNA-4157 (V940) melanoma vaccine

Intismeran autogene — previously known as V940 and, before that, as mRNA-4157 — is the most widely discussed personalised melanoma vaccine candidate. It is an individualised neoantigen therapy: tumour and healthy tissue are sequenced, computational models rank the mutations most likely to be visible to the immune system, and a bespoke mRNA product encoding those targets is manufactured for that one patient. It is given alongside a checkpoint inhibitor rather than instead of one.

  • It is investigational. As of 22 August 2026 it is not approved for sale in any market and is only available through clinical trials.
  • Its developers reported a randomised Phase 2b signal in resected high-risk melanoma when combined with pembrolizumab, versus pembrolizumab alone.
  • A Phase 3 melanoma study was announced on 19 August 2026 as having met its recurrence-free and distant-metastasis-free survival endpoints. That announcement was topline only, with full data not yet published or peer reviewed.
  • The AI component selects candidate neoantigens. It does not generate the immune response, and predicted targets do not always translate into presented peptides on tumour cells.

If you are asking about mRNA-4157 because of a personal melanoma diagnosis, the route in is a trial referral through your treating oncologist — not a prescription. Nothing on this page is treatment advice, and ScanSkinAI has no involvement in these trials.

Personalised cancer vaccines for melanoma

A personalised cancer vaccine is not a preventive vaccine. It is a therapy built after diagnosis, from the genetic profile of one person's tumour, designed to train that person's immune system to recognise cancer cells it has so far tolerated. Melanoma is the lead indication for this approach because melanoma tumours typically carry a high mutation burden, which gives the target-selection algorithms more material to work with.

Two design families dominate current melanoma work: individualised mRNA products encoding a panel of patient-specific neoantigens, and peptide-based products delivering the selected targets directly. Both rely on computational prediction to rank which mutations are worth including, and both are being tested in combination with checkpoint inhibitors rather than as standalone treatments.

Where this sits today: promising randomised signals in resected high-risk melanoma, one topline Phase 3 readout, and no commercially approved personalised melanoma vaccine anywhere. Manufacturing time, cost and access remain unresolved practical barriers even if approvals follow.

What the evidence currently shows

What we can reasonably conclude

  • AI-assisted selection of personalised vaccine targets is already operating inside human clinical trials.
  • Computationally selected targets can repeatedly generate patient-specific T-cell responses.
  • Some combinations — notably a personalised mRNA vaccine plus a checkpoint inhibitor — have produced encouraging melanoma trial results, including a randomised Phase 2b signal.
  • Independently of treatment, image-based AI can support risk awareness and structured lesion monitoring.

What has not yet been proven

  • That the AI component alone caused the clinical benefit — a successful trial validates the whole chain: sequencing, algorithm, construct, formulation, manufacturing and companion therapy.
  • That these vaccines work for every patient.
  • That they cure melanoma.
  • That response rates from small, single-arm trials can be compared directly between programmes.
  • That any programme discussed here is commercially approved; each remains investigational at the date of writing.

The future of AI-enabled melanoma care

The interesting question is not whether any single model is clever, but whether the stages of care can be joined up so that information travels with the patient.

Public awareness

Plain-language risk education that reaches people before a lesion changes.

Earlier screening

Low-friction photo screening that prompts action sooner rather than replacing a clinic visit.

Clinical triage

Helping services prioritise the people most likely to need urgent assessment.

Dermatologist assessment

Structured photo history handed to the clinician at the point of examination.

Lesion monitoring

Consistent, comparable images over months instead of memory and guesswork.

Genomic analysis

Tumour and germline sequencing interpreted at a scale manual review cannot match.

Treatment selection

Ranking candidate targets for investigational, individualised therapies.

Recurrence monitoring

Follow-up imaging and tracking after treatment, under clinical supervision.

The defensible asset in any of these stages is not the phrase "AI-powered". It is a validated chain running from analytical validation, through prospective testing, to controlled and reproducible outcomes.

ScanSkinAI's role

ScanSkinAI supports preventive skin-health awareness through smartphone-based screening, lesion analysis and monitoring. It is not a personalised cancer-vaccine platform and does not replace clinical diagnosis or treatment.

Continue exploring

Frequently asked questions

No. Image-based AI produces a risk-awareness output, not a diagnosis. It cannot confirm or exclude melanoma. A diagnosis is made by a clinician, normally after dermoscopic examination and, where indicated, a biopsy with laboratory histopathology.

Accuracy varies by system, image quality, lesion type and skin tone, and figures from one study do not transfer to another. Any single number should be read alongside the population it was measured in. Our own published performance figures and their limitations are set out in the ScanSkinAI accuracy report and methodology paper linked in the references.

It is an investigational therapy designed for one individual. Their tumour and healthy cells are sequenced, tumour-specific mutations are identified, and a limited set of the resulting protein fragments — neoantigens — is encoded into a vaccine intended to help their immune system recognise cancer cells carrying those targets.

AI is used to rank candidates. Algorithms predict which mutation-derived peptides are likely to be processed and presented by that patient's HLA molecules, scoring them on predicted binding, expression, clonality, immunogenicity and manufacturability. The AI does not create the immune response or treat the cancer; it helps researchers choose which targets to include.

No. Intismeran autogene, previously called V940 or mRNA-4157, remains investigational. Its developers announced on 19 August 2026 that a Phase 3 melanoma study met its recurrence-free and distant-metastasis-free survival endpoints, but that announcement was topline only and does not constitute regulatory approval.

That has not been shown. Trials to date report immune responses and, in some randomised settings, improved recurrence-free survival when combined with other therapy. Improved outcomes in a trial are not the same as a cure, and results vary between patients.

No. ScanSkinAI supports preventive skin-health awareness through smartphone-based screening, lesion analysis and monitoring. We do not diagnose, we do not treat, and we are not a cancer-vaccine platform. The vaccine research described on this page is educational and is not a service we offer.

See a clinician promptly if a mole is changing in size, shape or colour, is asymmetric or has an irregular border, itches, bleeds, crusts or fails to heal, or simply looks different from your other spots. If you have a personal or family history of skin cancer, many moles or heavily sun-exposed skin, ask about a regular clinical schedule rather than waiting for a change.

References

  1. Moderna — AI and the V940 / intismeran autogene design workflow
  2. Merck and Moderna — Phase 3 INTerpath-001 topline announcement, 19 August 2026
  3. The Lancet — KEYNOTE-942 randomised Phase 2b study
  4. BioNTech — predictive algorithms and the autogene cevumeran programme
  5. Nature Medicine — GNOS-PV02 Phase 1/2 study
  6. Dana-Farber Cancer Institute — NeoVax machine-learning neoantigen programme
  7. Evaxion — EVX-01 / PIONEER Phase 1 peer-reviewed study and Phase 2 immune-biomarker updates
  8. ClinicalTrials.gov — EVX-01 Phase 2, NCT05309421
  9. MHRA — impact of AI on the regulation of medical products

Links point to the publishing organisation. Specific study pages, effect sizes and approval status should be verified against the primary source at the time of reading; topline company announcements are not equivalent to peer-reviewed publication or regulatory approval. Our own performance figures are documented in the accuracy report and screening methodology paper.

Medical disclaimer

This page is educational and does not provide medical advice, diagnosis or treatment. ScanSkinAI is a screening and monitoring tool, not a diagnostic device, and it does not offer cancer treatment. All therapies described here are investigational. Always consult a qualified healthcare professional about any concerning or persistent skin change, and discuss clinical trial options with your oncology team.