• Introduction
• Why the First Generation Failed
• The Three Mechanistic Walls
• The Neoantigen Quality Problem
• The Cross-Presentation Bottleneck
• Current Clinical Data
• The Setback Column
• The Commercial Problem
• The Honest Forecast
• References

Cancer vaccines: thirty years of humbling biology, and finally a credible signal

The idea was always intellectually irresistible. Cancer cells carry somatic mutations that generate peptide fragments, presented on MHC class I molecules, that the immune system has never encountered and should recognize as foreign. The challenge is that recognizing and eliminating are very different things, and tumors have had evolutionary time to blur the line between them.

Why the first generation failed

Cancer Vaccine Development: 1990s to Projected First Approvals

From the early nineties through to roughly 2015, the field cycled through wave after wave of active specific immunotherapy candidates: peptides, dendritic cells, viral vectors, whole-cell preparations. They generated impressive T cell responses in preclinical models and mostly nothing useful in the clinic. The primary target class, tumor-associated antigens (TAAs) such as CEA and HER2, was also the primary structural flaw: because these proteins are expressed, at lower levels, on normal tissue, the immune system had already been conditioned to tolerate them through central and peripheral deletion of high-affinity clones. You were asking the adaptive immune response to unmute itself against a signal it had spent years suppressing.¹ Identifying tumor-specific antigens (TSA) was possible if technically challenging, but even then not all cells in a cancer growth expressed that target, leading to a reliance on epitope spreading and other mechanisms to (hopefully) mop up the rest.

Sipuleucel-T's 2010 FDA approval was simultaneously the field's greatest vindication and most cautionary data point. In the pivotal IMPACT trial, the dendritic cell-based autologous immunotherapy demonstrated a 4.1-month improvement in median overall survival versus placebo in castrate-resistant prostate cancer (N Engl J Med 2010;363:411-422), but failed to reduce PSA or time to objective progression: a mechanistically incoherent dissociation that the field still debates. By 2014, Dendreon filed for bankruptcy. The European Commission withdrew marketing authorization in 2015. A product that proved immune-mediated antitumor activity was possible couldn't survive the economic and operational reality of three leukapheresis cycles at $93,000 per course with a mechanism of action the oncology community couldn't confidently articulate.^2,3,4,5,6

The three mechanistic walls

The neoantigen quality problem

Tumor mutational burden is an imperfect proxy for immunogenic neoantigen load. TMB measures mutation count; what determines clinical relevance is whether a given peptide is processed by the proteasome, transported by TAP, loaded onto the patient's specific HLA allotype, presented at threshold density, and then recognized at sufficient TCR affinity to drive a functional immune response. Computational pipelines anchored in HLA-binding affinity prediction return substantial false-positive rates because they do not capture the full chain of events between a mutation and an immunogenic pMHC complex (peptide-MHC complex presented on the cell surface). Features including gene expression level, variant allele frequency, peptide stability in the HLA groove, and TCR-pMHC binding energetics are all partially predictive but none individually sufficient. In practical terms, a vaccine encoding 20-34 neoantigens is likely to contain a meaningful fraction of peptides that never generate a T cell response, wasting limited vaccine real estate.^13,14,15

Intratumoral heterogeneity compounds this. McGranahan and colleagues demonstrated in Science (2016;351:1463-1469) that NSCLC patients with high clonal neoantigen burden and low neoantigen heterogeneity had significantly better survival and deeper responses to anti-PD-1 therapy than those with predominantly subclonal neoantigens, establishing clonality as a mechanistically grounded biomarker. Targeting subclonal neoantigens leaves clonal architecture intact and exerts selective pressure that can eliminate vaccine-targeted subclones while allowing antigen-negative variants to expand. Targeting clonal neoantigens arising from early truncal mutations attacks every tumor cell by design, but identifying them requires deep multi-region sequencing and the bioinformatics infrastructure to reliably distinguish clonal from subclonal events. Spatial heterogeneity invisible to the biopsy used to design the vaccine remains an irreducible problem.^16,17,18,19

A parallel approach avoids the individualized manufacturing challenge entirely: shared neoantigen vaccines targeting recurrent driver mutations (KRAS G12D/G12V/G12C, TP53 hotspots) that are common across patients and tumor types. These "off-the-shelf" designs are commercially more tractable and compatible with conventional batch manufacturing. However, the clinical data for shared-antigen cancer vaccines are less mature than for the personalized programs, and the immunogenicity of common driver mutations is constrained by the same tolerance and presentation barriers outlined above. This analysis focuses on the personalized model because it has generated the most advanced clinical evidence to date, including the two registration-relevant datasets discussed below.

The cross-presentation bottleneck

Upstream of all the above, mRNA-based vaccines require the antigen they encode to be cross-presented on MHC class I by dendritic cells, specifically the cDC1 subset, which expresses XCR1 and CLEC9A and is specialized for cross-priming CD8+ T cells. Cross-presentation is dependent on CD4+ T cell help via CD40L-CD40 signaling, sensitive to the local inflammatory milieu, and inherently less efficient than direct endogenous antigen presentation. Recent work on adjuvants that expand and activate cDC1s, including Flt3L-FlaB hybrid constructs, has shown that augmenting this subset can substantially improve vaccine-induced generation of stem-like memory (Tscm) and precursor-exhausted (Tpex) CD8+ T cell phenotypes, which are the populations most capable of sustained long-term antitumor surveillance. The goal is not simply to prime a T cell response. It is to prime T cells with the phenotypic durability to persist in an immunosuppressive environment over months to years: a different and harder specification.^20,21,22,23

What the current clinical data actually show

The parallel renal cell carcinoma neoantigen vaccine trial from Dana-Farber (also published in Nature, February 2025) in 9 patients showed zero recurrences at up to 3+ years in early follow-up: a small but mechanistically coherent result in a tumor type where adjuvant checkpoint inhibition already has an established role.²⁶

The open question across these datasets is what predicts response. In the PDAC cohort, responders were distinguished by the magnitude of neoantigen-specific CD8+ T cell clonotype expansion, successful transition to a memory phenotype, and persistence of effector function over years of follow-up. Circulating tumor DNA (ctDNA) clearance has emerged as a potential early biomarker for vaccine efficacy in the adjuvant setting, with several groups investigating whether post-vaccination ctDNA dynamics can identify responders before clinical recurrence events occur. No validated companion diagnostic exists for personalized cancer vaccines, and the development of one is complicated by the individualized nature of the product: unlike a targeted therapy where a single biomarker predicts response to a fixed drug, vaccine response depends on the interaction between each patient's unique neoantigen repertoire, HLA allotype, immune repertoire, and tumor biology. Without a response-prediction biomarker, clinical trials must enroll broadly and commercial models must price for heterogeneous efficacy across the treated population.²⁷

The setback column

Against these signals, the failures warrant honest accounting. The accelerated approval pathway for intismeran autogene was not pursued in September 2024, with the FDA indicating that Phase 2b data were insufficient and raising questions about the AI-driven neoantigen selection component of the manufacturing process: a regulatory signal that the agency is applying heightened scrutiny to AI-enabled clinical decision-making in biologics manufacturing. Autogene cevumeran failed to demonstrate benefit in first-line metastatic melanoma. INTerpath-007 in cutaneous squamous cell carcinoma was discontinued after enrolling 46 patients. The Lancet in June 2025 published a review concluding that therapeutic cancer vaccine trials remain heterogeneous in design, target populations, and endpoints, making cross-trial inference difficult: a structural problem for evidence synthesis. Phase 3 INTerpath-001 in melanoma is the pivotal readout, with a decision window expected in late 2026.^28,29,30

The regulatory path forward is a traditional BLA with recurrence-free survival as the primary endpoint, not an accelerated approval based on a surrogate. Given the responder/non-responder split observed across trials, regulators may eventually require a companion diagnostic or biomarker-based patient selection strategy, though no validated predictive biomarker currently exists. The CMC regulatory challenge is itself novel: each dose is a chemically unique product manufactured for one patient, raising questions about batch definition, release testing standards, and manufacturing consistency that existing biologics frameworks were not designed to address. The FDA's September 2024 signal about AI-driven neoantigen selection suggests that the computational pipeline used to design each individualized vaccine will face regulatory scrutiny as a controlled manufacturing step, not merely a research tool.

The strategic lesson emerging from the pattern of successes and failures: tumor burden is the governing variable. Vaccines appear to be effective in the adjuvant micrometastatic setting, where a small residual cancer cell population meets a primed immune response with a functioning effector arm. In established metastatic disease, where the immunosuppressive TME is fully elaborated, the tumor mass is high, and immune exhaustion is advanced, vaccine-primed T cells face a suppressive environment they cannot overcome with available combination partners. This has profound implications for addressable patient populations and commercial scope.^16,28

The commercial problem

Personalized mRNA Vaccine: Per-Patient Manufacturing Workflow

Per-patient manufacturing costs for personalized mRNA cancer vaccines currently exceed $100,000, driven by the individualized workflow: multi-region tumor biopsy and NGS, bioinformatics neoantigen-HLA prediction, GMP mRNA synthesis, formulation, QC release testing, and cold-chain logistics to the treatment center, all within a manufacturing window of approximately six to nine weeks that is constrained by the clinical urgency of the adjuvant setting. Modular automation and AI-driven pipeline compression may reduce timelines and costs at scale, but the fundamental architecture of individualized manufacturing prevents the unit economics of standard biopharma batch production from ever fully applying.^30,31

Existing HTA frameworks were designed for products with fixed drug compositions, defined comparative populations, and stable clinical outcome distributions. A product that is chemically unique per patient, whose clinical benefit partially depends on whether that specific patient's immune system mounts a T cell response, and whose manufacturing begins at surgery, does not fit the conventional cost-effectiveness model cleanly. The ICER arithmetic at current price assumptions, for an adjuvant indication in patients who may survive for many years on other regimens, is structurally challenging. NICE currently operates at a £25,000-35,000/QALY threshold, rising to £50,000 for end-of-life indications; personalized cancer vaccines at launch pricing are almost certainly going to land well above both thresholds in many HTA jurisdictions.^32,33

The Provenge story is the historical template here, and the field would do well to treat it as such rather than as ancient history. A first-in-class immunotherapy, approved on a survival endpoint that payers could not translate into daily clinical practice, at a price point that broke the payer-provider relationship, in a disease with competitive alternatives, resulted in market withdrawal in Europe and corporate bankruptcy in the US. The next generation of personalized cancer vaccines is arriving at a price point that, based on per-patient manufacturing costs alone, will substantially exceed Provenge's, in a more complex manufacturing model, to payers who are currently under political pressure to control specialty drug costs. Outcomes-based contracting and risk-sharing schemes will likely emerge as the access mechanism of choice, though the industry's track record with outcomes-based agreements to date has been mixed, with few examples of durable implementation at scale for complex biologics. Negotiating these for a product with individualized manufacturing, heterogeneous patient-level response rates, and no established biomarker for response prediction is operationally unprecedented.^3,4,2

Equity deserves explicit statement. A therapy requiring expert surgical center access, multi-region tumor sampling, NGS-based neoantigen calling, decentralized GMP manufacturing, and cold-chain logistics to specialized infusion centers is structurally inaccessible to patients outside tertiary academic centers in high-income settings. The BMJ has flagged pooled procurement and equity-based tiered pricing as the policy mechanisms most likely to widen access in lower-income healthcare systems, but neither is currently operational for this product class.³⁴

The honest forecast

First approvals in adjuvant melanoma are plausible in the 2027-2029 window, contingent on INTerpath-001. The indication will be narrow (roughly 10,000-15,000 patients per year are diagnosed with resectable stage III/IV melanoma in the US; initial uptake will be limited to NCI-designated cancer centers with the manufacturing logistics infrastructure, further constraining the addressable population), the pricing will be the most contentious oncology launch since CAR-T (Kymriah launched at $475,000 in 2017; Yescarta at $373,000; both remain commercially viable seven years later but face persistent reimbursement friction outside the US), and uptake will be concentrated in academic centers with the infrastructure to manage the manufacturing logistics. Payers will push back hard, outcomes-based contracts will emerge, and the equity access gap will be real and significant.

The competitive field extends well beyond the Moderna/Merck and BioNTech programs discussed above. Gritstone Bio's GRANITE and SLATE programs use a heterologous prime-boost approach combining self-amplifying RNA with adenoviral vectors. Nykode Therapeutics is developing VB10.NEO, a DNA-based platform targeting both shared and individualized neoantigens. CureVac and Transgene (TG4050, viral vector) are pursuing individualized approaches on different platform architectures. BioNTech's FixVac platform targets shared tumor-associated antigens with an off-the-shelf design, representing a parallel strategy to their individualized autogene cevumeran work. This platform diversity is both a scientific strength, because different delivery systems may prove optimal for different tumor types and clinical settings, and a source of interpretive difficulty, because cross-trial comparison is complicated by differences in antigen selection, delivery platform, formulation, dosing schedule, combination partner, and patient population.

The longer-term trajectory depends on three convergences: manufacturing costs falling to commercially viable levels through automation and process standardization; better predictive biomarkers to identify which patients will mount clinically meaningful immune responses (the PDAC responder/non-responder split, where 8 of 16 patients generated vaccine-induced T cells, is both the most encouraging and most sobering signal in the field, though the response rate will vary substantially by indication and tumor biology); and validated combinations that address the TME suppressive network beyond the PD-1 axis alone.^30,1

What to watch

The field has persisted through thirty years of failure by building better science. The mechanistic case is now the most compelling it has ever been. Whether that science can survive commercial reality, and reach the patients who need it, is the question that no Phase 3 trial answers.

References