Rethinking Cancer Vaccine Carriers: A Bottleneck-First Test of Antigen Delivery Platforms

Therapeutic cancer vaccines have moved from a distant aspiration to a clinical reality, with personalized neoantigen vaccines now demonstrating that carefully selected tumor antigens can prime tumor-reactive T cells in patients. Yet a persistent and underappreciated problem shadows every success: the carrier that delivers the antigen. A new review published in Materials Today Bio argues that the field has been evaluating delivery platforms the wrong way, celebrating increased cellular uptake or tumor shrinkage without asking whether the carrier variable under study actually caused a defined delivery improvement that translated into antigen-specific immunity. The authors, led by Shuai Wang and Dingkang Liu, propose a bottleneck-first framework that begins with a specific biological limitation—protecting the antigen, reaching the right antigen-presenting cells, entering productive processing pathways, or maintaining function across repeat doses—and then asks whether a controllable carrier feature changes a matched delivery endpoint and a functional outcome.

The framework classifies candidate carriers across seven material matrices plus a carrier-free assembly mode: lipid, polymer, coordination network, other inorganic or mineral matrix, biological membrane or vesicle, protein scaffold or capsid, and nucleic acid nanostructure. Each matrix carries its own engineerable variables and interpretive constraints. Lipid nanoparticles and liposomes offer control over lipid identity, component ratios, and the apparent pKa of the formulated particle, supporting nucleic acid protection and transfection, but they bring formulation-dependent innate immune activation and the risk of anti-PEG antibody responses. Polymer systems enable sustained release and intracellular transfer, while metal-organic frameworks provide stimulus-coupled release at the cost of metal-dependent toxicity and product heterogeneity. Biological membranes and extracellular vesicles inherit source-cell trafficking behaviors but suffer from batch heterogeneity that makes mechanistic attribution notoriously difficult.

The evidence hierarchy across these platforms is strikingly uneven. Lipid nanoparticles stand alone in possessing randomized human evidence through V940 (mRNA-4157), the individualized neoantigen therapy tested in the KEYNOTE-942 phase IIb trial in resected melanoma. In that study of 157 patients randomized to V940 plus pembrolizumab or pembrolizumab alone, the primary recurrence-free survival analysis met its prespecified one-sided criterion with a hazard ratio of 0.561, though the two-sided P value of 0.053 fell just short of conventional significance. Sponsors later announced that the phase III INTerpath-001 trial met its recurrence-free and distant-metastasis-free survival endpoints, without disclosing effect estimates. Crucially, the review notes, this regimen tests the addition of the whole vaccine product, not an isolated lipid variable, so it cannot establish which carrier feature contributed to the benefit.

By contrast, most other platforms remain preclinical for cancer vaccination. Polymer, coordination network, mineral, and carrier-free systems have generated encouraging mouse data but no human therapeutic trials. Protein cages and virus-like particles have documented scaffold-specific CD4 T-cell responses in prophylactic HIV vaccination, and DNA origami searches found no human vaccine trial as of August 2026. A 22-patient single-arm study of dendritic cell-derived exosomes in non-small-cell lung cancer missed its progression-free survival target, with four-month progression-free survival of just 32 percent and no objective responses, though the design prevents attribution to the exosome carrier itself. The review classifies selected comparisons as supported, partial, untested, or refuted, and remarkably few achieve full support because the causal chain from carrier variable to delivery endpoint to antigen-specific function is rarely complete.

One of the review’s sharpest analytical tools is its dissection of surface engineering claims. The authors separate the target cell-receptor axis, which determines which cell subset a carrier engages, from the anatomical interface axis, which determines whether the carrier reaches the required tissue at all. A single modification can affect both, so a targeting claim must identify which axis is being manipulated and treat the other as a potential confounder. The heavy-chain ferritin cage, for example, binds transferrin receptor 1 and undergoes receptor-mediated uptake, but transferrin receptor 1 is also expressed by normal proliferating cells and erythroid-lineage cells, so receptor expression confers no tumor specificity. Increased association with cells is insufficient evidence of improved vaccination when receptor dependence is not demonstrated or when uptake terminates in degradation rather than presentation.

The review also introduces a critical distinction among combination configurations that is frequently blurred in the literature. Physical co-delivery means antigen and immune modulator share one particle; same-tissue modulation keeps them physically separate but locally co-localized; systemic co-administration combines separate treatments at the regimen level. Each requires claim-matched evidence and should not be relabeled as another category. Endosomolytic polymersomes that increased cyclic dinucleotide uptake and STING activity, for instance, do not provide physical co-delivery evidence because no selected antigen was co-loaded. Similarly, systemic checkpoint blockade paired with an mRNA vaccine is regimen-level co-administration, not carrier-level coupling, and the review warns that mechanistic compatibility on paper does not predict clinical efficacy, as illustrated by the failed phase III trial of talimogene laherparepvec plus pembrolizumab in melanoma.

Beyond priming, the review identifies the tumor microenvironment as a second, separable delivery-to-immunity bottleneck. A carrier may successfully deliver antigen to lymph-node antigen-presenting cells and expand antigen-specific T cells while leaving endothelial entry, movement through collagen-rich stroma, and function within suppressive myeloid niches entirely unchanged. Human urothelial tumor data link transforming growth factor-beta signaling in fibroblasts to exclusion of CD8 T cells from tumor parenchyma, and dynamic imaging of viable human lung tumor slices shows macrophage-mediated trapping reducing T-cell motility and access to tumor islets. If priming persists without intratumoral access or activity, the unresolved bottleneck is effector-cell access, not antigen delivery, and tumor control alone cannot identify which bottleneck was actually resolved.

Repeat dosing and manufacturing constraints receive equally rigorous treatment. Anti-PEG antibodies boosted by SARS-CoV-2 mRNA vaccines did not impair neutralizing responses in one study, and a small pharmacokinetic study of a single booster found baseline anti-PEG antibodies did not track mRNA exposure, but the authors caution that seropositivity alone is neither failure nor safety. Carrier-directed helper responses can be substantial: in prophylactic HIV vaccination, scaffold-specific CD4 responses were detected in the vast majority of recipients, and in mouse models ferritin peptides dominated the responding T-cell population. Whether such responses help, coexist harmlessly, or interfere through epitopic suppression depends on measured effects on later-dose delivery and target-antigen function. Personalized manufacture compounds these questions, since each patient’s vaccine is a new sequence-specific product requiring lot release, consistency, and post-change comparability testing that must fit within clinical treatment windows.

The review closes with a bounded role for artificial intelligence. Machine learning can prioritize carrier compositions, fusion sites, linker lengths, and antigen candidates, but the authors document sobering failure modes: delivery rankings for 281 lipid nanoparticles in cultured cells did not predict their rankings across tissues in vivo, and species-dependent functional mRNA delivery undermines cross-species extrapolation. They propose four experimental gates—domain, product, delivery, and function/toxicity—that any AI-prioritized design must pass before advancement. The overarching conclusion is demanding but clarifying: no carrier is optimal across all bottlenecks, and a delivery advantage is only supported when a defined variable changes a prespecified transport, release, or processing step linked to an antigen-specific outcome that persists across repeat doses and independently manufactured lots. Increased uptake and tumor regression, the authors insist, are simply not enough.

Subject of Research: Evaluation of antigen delivery platforms for therapeutic cancer vaccines using a bottleneck-first evidence framework

Article Title: A bottleneck-first evaluation of antigen delivery platforms for therapeutic cancer vaccines

Article References: Wang, S., Zhang, M., Tian, J., Wu, J., Lu, S., Liu, L., & Liu, D. (2026). A bottleneck-first evaluation of antigen delivery platforms for therapeutic cancer vaccines. Materials Today Bio, 41, Article 103715. https://doi.org/10.1016/j.mtbio.2026.103715

Image Credits: AI Generated

DOI: 10.1016/j.mtbio.2026.103715

Keywords: cancer vaccines, antigen delivery, lipid nanoparticles, neoantigens, nanomedicine, mRNA vaccines, protein cages, virus-like particles, extracellular vesicles, tumor microenvironment, immunotherapy, drug delivery