Personalized Melanoma Vaccines Using Tumor-Specific Targets: How They Work and Clinical Results

The generation of a personalised neoantigen vaccine is a complex, multi-step bioinformatic and manufacturing process (Figure 1). The workflow begins with the collection of tumour tissue and a matched normal sample (e.g., peripheral blood). Tumour and normal DNA are subjected to whole-exome sequencing (WES) or whole-genome sequencing (WGS) to identify somatic (tumour-specific) mutations. Simultaneously, tumour RNA sequencing (RNA-seq) is performed to confirm that these mutations are expressed and to quantify their transcript abundance (8, 13).

Following variant calling, the patient’s Human Leukocyte Antigen (HLA) haplotype is determined. Bioinformatic algorithms (e.g., NetMHCpan) are then used to predict which of the thousands of mutated peptides (neoepitopes) will bind with high affinity to the patient’s specific MHC Class I or Class II molecules. Rule-based pipelines that prioritize predicted HLA binding often exhibit low positive predictive value for true T-cell immunogenic neoantigens; benchmarking studies show that many top-ranked candidates fail downstream validation. They frequently fail to account for subsequent critical steps in antigen processing and presentation, such as proteasomal cleavage, peptide transport via TAP (Transporter associated with Antigen Processing), and the stability of the final peptide-MHC complex, leading to a high false-positive rate (12, 14).

To overcome the limitations of simple binding-affinity algorithms, a new generation of Artificial Intelligence (AI) and machine learning models has been developed. These tools, such as imNEO, DeepNeoAG, and ImmuneMirror (Table 1), integrate multi-omic data, including mass spectrometry (MS)-verified immunopeptidomes, gene expression levels, and features of T-cell recognition. These models aim to provide a more composite prioritization score rather than a simple binding score (12, 15).

While these models report superior performance, with Areas Under the Curve (AUCs) often exceeding 0.85 in in vitro validation datasets (Table 1), a critical “validation gap” persists. The high AUCs reported by these models primarily reflect their accuracy at predicting in vitro peptide-MHC binding, an essential but insufficient proxy for in vivo T-cell activation and, more importantly, clinical efficacy. The translational link between a high prioritization score and a patient’s recurrence-free survival (RFS) has not yet been prospectively established (12).

Terminology used in this review: Binding refers to predicting peptide–MHC affinity/stability in vitro. Presentation refers to predicting whether a peptide is generated and displayed on the cell surface (processing, transport, and ligand elution), which is closer to in vivo biology but still not an immune response. Immunogenicity refers to demonstrable T-cell recognition/activation (functional assays and/or clinical immunomonitoring), and therefore should not be inferred from binding AUC alone.

An additional application of AI-guided prioritization that warrants explicit consideration is the selection of short peptides, or minimal epitopes, designed primarily for HLA class I presentation. In contrast to synthetic long peptides, which often require endosomal uptake and may preferentially expand CD4⁺ T-helper responses, short-peptide strategies can be used to enrich for candidate neoepitopes with high predicted HLA-I binding, favourable processing features, and greater likelihood of eliciting cytotoxic CD8⁺ T-cell responses. In this context, AI models are not merely ranking peptide–MHC affinity, but are increasingly being used to refine epitope length, anchor-residue suitability, presentation probability, and, in some cases, T-cell recognition features. These short-peptide approaches therefore represent a mechanistically distinct design strategy within the broader peptide-vaccine landscape, and their inclusion helps explain why peptide platforms should not be treated as uniformly CD4⁺-dominant (12, 14, 15). Nevertheless, short-peptide approaches are not universally superior, because their performance remains constrained by HLA restriction, peptide stability, and the risk of incomplete helper T-cell support.

This validation gap is exacerbated by what has been termed the “neoantigen algorithm reproducibility crisis”. Many AI models exhibit performance bias due to their training data. For example, models trained on datasets enriched for specific HLA alleles or on melanoma-specific sequences (which are abundant) may not generalize well to other cancer types or patients with different HLA haplotypes. This lack of standardization and poor transferability remains a major scientific and regulatory hurdle (12).

Although Table 1 summarizes predictive performance, underlying architectural diversity contributes significantly to the observed spread in AUCs. DeepNeoAG utilizes a recurrent convolutional neural network trained on CEDAR peptide datasets, focusing on motif recognition independent of HLA allele context. ImmuneMirror instead applies ensemble gradient-boosting with integrated binding-stability and ligand-elution data, achieving higher biological interpretability. imNEO extends this framework by introducing TCR–epitope co-features, improving recall but at risk of melanoma-biased overfitting (12).

While the field has made significant progress, two structural weaknesses underlie the reproducibility crisis in neoantigen prediction.

First, most machine learning pipelines are trained on heavily biased datasets—particularly overrepresented HLA-A02:01* alleles and melanoma-derived immunopeptidomes. As a result, model accuracy often collapses when applied to rarer HLA haplotypes or non-melanoma tumours. Second, benchmark datasets such as IEDB or CEDAR lack unified standards for peptide length, affinity thresholds, or negative-sample definition, inflating in vitro AUCs without reflecting true immunogenicity (12).

Emerging solutions emphasize cross-cohort benchmarking and synthetic augmentation of rare HLA alleles to enhance generalizability. In parallel, regulators have begun framing AI-based pipelines as “software as a medical device (SaMD)” components under Chemistry, Manufacturing and Controls (CMC) standards. This shift allows algorithmic validation to become a formal part of regulatory review, potentially transforming the current ad hoc research tools into auditable, clinical-grade systems (16–18).

Messenger RNA (mRNA) vaccines, typically encapsulated in lipid nanoparticles (LNPs), are delivered directly into the cytoplasm of cells, primarily antigen-presenting cells (APCs) such as dendritic cells. Once inside, the mRNA is translated by the host cell’s own ribosomes, producing the neoantigen protein endogenously (i.e., inside the cell) (8, 13).

This intracellular origin is immunologically critical. Endogenously synthesized proteins are processed by the proteasome into short peptides. These peptides are then transported by TAP into the endoplasmic reticulum, where they are loaded onto MHC Class I molecules. The peptide-MHC-I complex is then trafficked to the cell surface.

Presentation on MHC class I provides a major route for CD8⁺ cytotoxic T-lymphocyte (CTL) priming. This helps explain why mRNA vaccine trials often show strong CD8⁺-skewed responses, although accompanying CD4⁺ responses can also contribute meaningfully to anti-tumour immunity. This direct priming of CTLs—the immune system’s primary tumour-killing cells—provides a strong mechanistic rationale for the synergy observed between mRNA vaccines and anti-PD-1 ICIs. The vaccine primes an army of new tumour-specific killers, and the ICI releases the PD-1 “brake,” allowing them to execute their function (8, 9, 19).

In contrast, synthetic long peptide (SLP) vaccines, which are co-administered with an adjuvant (e.g., Poly-ICLC, Montanide) to stimulate APCs, are exogenous antigens. They are taken up from the extracellular space by professional APCs via endocytosis or phagocytosis (7, 13). It is also important to distinguish SLP vaccines from short-peptide or minimal-epitope formulations, which are often intentionally designed for MHC class I loading and CD8⁺ T-cell activation and therefore should not be mechanistically collapsed into the same category.

These exogenous peptides traffic through the endolysosomal pathway, where they are processed and loaded onto MHC Class II molecules. The peptide-MHC-II complex is then presented on the APC surface.

Presentation on MHC class II is a major mechanism through which SLP vaccines activate CD4⁺ T-helper cells, and this likely contributes to the high-frequency CD4⁺ responses reported in several peptide-vaccine trials. However, this pathway should not be interpreted as exclusive. After uptake by professional APCs, exogenous peptide antigens can also enter the MHC class I pathway through cross-presentation, thereby generating CD8⁺ T-cell responses under favourable biological and adjuvant conditions. Accordingly, peptide-based vaccines should be viewed as platforms that often favour CD4⁺ helper immunity but remain capable of inducing mixed CD4⁺/CD8⁺ responses, with the balance depending on peptide design, APC subset engagement, adjuvant choice, and antigen-processing efficiency. This mechanistic difference may underlie the more variable efficacy signals observed with peptide-based platforms (7, 13, 20).

Although the majority of current pipelines emphasize MHC-I–restricted CD8⁺ T-cell epitopes, the contribution of CD4⁺ T-cell responses via MHC-II presentation remains underexplored.

Prediction algorithms such as NetMHCIIpan 4.1 now allow high-throughput identification of HLA class II–restricted epitopes, though their accuracy remains lower than class I counterparts (14).

Incorporating MHC-II predictions may help explain why several peptide vaccine trials demonstrated robust CD4⁺ responses without corresponding clinical benefit, suggesting a need for balanced epitope selection (21, 22).

The choice of platform involves a trade-off between the desired immune response, manufacturing speed, cost, and logistical stability. mRNA vaccines, for example, offer rapid manufacturing but require a stringent cold chain, whereas peptides are more stable but have a longer synthesis time. These differences are summarized in Table 2 (13, 16).

A quantitative meta-analysis of neoantigen vaccine trials is precluded by the profound heterogeneity across studies. Trials differ in vaccine platform (mRNA vs. peptide), adjuvant used (e.g., Poly-ICLC, Montanide, CDX-1140), patient population (adjuvant stage III or IV vs. metastatic), comparator (ICI monotherapy vs. single-arm), and primary endpoints (RFS vs. ORR) (9, 11, 21–23).

Presenting these data in a single “forest plot” or calculating a “weighted average” effect size, as has been attempted, is statistically invalid and highly misleading. Combining heterogeneous endpoints like Hazard Ratios (HRs), which are time-to-event measures, and Objective Response Rates (ORRs), which are dichotomous proportions, on a single visual axis is methodologically unsound. Furthermore, attempting to “approximate” an HR from a Kaplan-Meier proportion ignores censored data and violates the proportional hazards assumption, rendering the estimate uninterpretable. Any weighting by simple sample size (), rather than by the inverse variance of the effect estimate, is not a valid meta-analytic technique. n

Therefore, this review will not conduct a meta-analysis. Instead, it adheres to the Synthesis Without Meta-analysis (SWiM) guidelines by presenting a critical narrative synthesis, with results from the six key trials summarized descriptively in Table 3. A risk-of-bias assessment (Table 4) highlights that all single-arm trials included suffer from a high risk of selection and reporting bias (24).

The clinical evidence for neoantigen vaccines in melanoma is derived from a small number of key Phase I/II trials (Table 3). A critical re-examination of the data from these trials, including recent 2024–2025 updates, reveals a more complex and nuanced picture than previously reported (9, 11, 21–23, 25, 28).

To clarify the relationship between vaccine type, prediction pipeline, and outcome, Figure 2B stratifies the key melanoma trials by platform and AI model sophistication.

Although the clonal architecture of tumours has been recognized as a major determinant of vaccine efficacy, only a minority of ongoing trials explicitly incorporate clonal analysis in antigen selection. Although tumour clonality is increasingly recognized as important, explicit clonality-informed antigen selection is not consistently reported across trials, and thresholds/pipelines remain non-standardized.

In practice, clonality-informed antigen selection depends not only on mutation detection by NGS, but also on how variant allele fraction is interpreted after accounting for tumour purity, local copy-number status, sequencing depth, and sampling bias. As a result, there is no universally accepted NGS-based cutoff that cleanly separates clonal from subclonal neoantigens across studies. For this reason, clonality should be treated as a probabilistic prioritization feature rather than a binary rule and ideally interpreted together with RNA expression and antigen-presentation likelihood.

Incorporating clonal frequency estimates into antigen prioritization pipelines may improve the likelihood of targeting truncal neoantigens that persist across metastases. However, real-time implementation remains constrained by the computational cost and lack of standardized thresholds for clonality calls (26, 27).

Trials employing deep-learning–guided neoantigen selection (e.g., EVX-01, imNEO) cluster in the upper-right quadrant, reflecting higher ORR and durability, even with comparable baseline TMB and disease stage (Supplementary Table 1) (28).

The corrected trial data (Table 3) reveals distinct patterns.

For mRNA Platform (KEYNOTE-942), the randomized Phase 2b KEYNOTE-942 trial provides the highest-quality evidence to date. The combination of the mRNA vaccine mRNA-4157 with pembrolizumab resulted in a clinically and statistically significant improvement in RFS. The 3-year update confirmed a durable benefit, with an RFS HR of 0.510 (49% risk reduction) and a distant metastasis-free survival (DMFS) HR of 0.384. In the randomized phase 2b KEYNOTE-942 trial, the observed 49% relative reduction in recurrence or death was demonstrated within a cohort of completely resected high-risk stage IIIB-IV melanoma. Because recurrence risk differs across TNM categories, direct comparison with external populations should be interpreted cautiously unless stage-matched subgroup data are available. This result strongly supports the hypothesis that vaccine-induced cellular immunity, including a substantial cytotoxic T-cell component, can add tangible benefit to checkpoint blockade (9, 19, 25).

For Peptide Platforms (A Heterogeneous Picture), they show significant variability.

While neoantigen-based vaccines have demonstrated promising efficacy in select melanoma cohorts, predictive biomarkers that delineate responders from non-responders remain underdefined. A multidimensional biomarker strategy is therefore essential to guide patient selection (Table 5), monitor vaccine-induced immunity, and optimize clinical outcomes.

a. Predictive biomarkers

Tumour mutational burden (TMB) has been consistently correlated with the abundance of neoantigens; however, its predictive precision is limited by intertumoural heterogeneity and non-immunogenic passenger mutations. Clonality of neoantigens—particularly those derived from truncal mutations—has shown stronger association with durable responses compared to subclonal targets. In parallel, baseline interferon-γ (IFN-γ) signalling and MHC-I expression levels serve as key immunocompetence markers, influencing both antigen presentation and cytotoxic T-cell priming efficiency (2, 26, 29–31).

b. Dynamic monitoring biomarkers

Dynamic immune readouts have become critical in capturing real-time therapeutic efficacy. Circulating tumour DNA (ctDNA) dynamics and longitudinal T-cell receptor (TCR) repertoire tracking are being evaluated as potential response markers in neoantigen-vaccine studies. These approaches enable quantification of clonal expansion and persistence of vaccine-primed T cells, offering a more granular readout of immune kinetics than static assays (12, 13).

c. Integrative modelling and composite scores

Recent efforts have focused on integrating genomic, transcriptomic, and immunophenotypic metrics into composite predictive models. Frameworks based on AI-driven ensemble predictors combine TMB, neoantigen clonality, and immune signatures to stratify patients into high- and low-likelihood responders. Such composite indices could evolve into standardized metrics for clinical trial stratification and regulatory validation (12).

In summary, predictive biomarkers must evolve beyond TMB-centric paradigms toward dynamic, integrative models capturing both tumour-intrinsic and immune-contextual variables.

A primary mechanism of vaccine failure and tumour immune escape is intra-tumour heterogeneity (ITH). A tumour is not a monolith but an evolving ecosystem of distinct subclones. Mutations can be “clonal” (truncal), meaning they arose early and are present in all tumour cells, or “subclonal” (branched), existing in only a subset of cells (27).

Although clonality is already recognized as important, its impact cannot be overstated. A landmark meta-analysis of 12 checkpoint inhibition trials provided a definitive insight: high subclonal TMB confers no clinical benefit to CPI therapy, whereas high clonal TMB is significantly correlated with better overall survival (p = 2.9×10⁻⁷) (26).

The implication for vaccine design is profound. A vaccine targeting a subclonal neoantigen may successfully eliminate the subclones that express it, but this merely “prunes the tree” and selects for the growth of other subclones that lack the target antigen, leading directly to relapse. Therefore, truncal or highly clonal neoantigens are generally attractive vaccine targets, although effective vaccine design must also account for subclonal architecture, immunoediting, and dynamic antigen loss. The failure to prioritize clonal neoantigens, a bioinformatic challenge, is a likely explanation for the modest efficacy seen in trials like NCT04072900↗ (21, 26, 27).

The highly immunogenic, “hot” TME of melanoma is permissive for immunotherapy. Trials in “colder,” more immunosuppressive tumours, such as glioblastoma (GBM) and ovarian cancer (OvCa), provide crucial insights into failure mechanisms. However, the analysis in the original draft contained critical misinterpretations of key studies (32–34).

The success of neoantigen vaccines in melanoma, and their relative struggles elsewhere, is not arbitrary. It is a direct function of fundamental differences in tumour immunobiology. TMB provides the targets, but the TME dictates the strategy. A comparison of these three distinct cancers (Table 6) illuminates the landscape of challenges (26, 34).

Beyond subclonal diversity, multiple molecular mechanisms can undermine vaccine-induced immunity. Loss of antigen presentation—via β2-microglobulin (β2M) mutations, MHC-I heavy chain loss, or TAP1/2 transport defects—renders tumour cells invisible to cytotoxic T lymphocytes despite robust priming (6, 26).

Similarly, chronic antigen exposure may drive T-cell exhaustion characterized by PD-1^highTIM-3⁺LAG-3⁺phenotypes, reducing effector function (6).

Under selective pressure, tumours may undergo immunoediting, progressively deleting target neoantigens or downregulating interferon-γ signalling pathways (6, 26).

Addressing these escape routes will require rational combination strategies—such as pairing vaccines with oncolytic viruses, STING agonists, or epigenetic modulators—to restore antigen presentation and reinvigorate exhausted T cells (13, 39).