Case Study 26.2 — Thirty Years of Failure
Why therapeutic cancer vaccines did not work, and what neoantigen approaches do differently
There is a version of the history of therapeutic cancer vaccines that treats it as a cautionary tale about scientific overconfidence, and there is a version that treats it as a heroic prelude to a breakthrough. Both are too tidy.
The accurate version is more interesting and more useful: a field made a specific, identifiable scientific mistake, repeated it for roughly three decades across many programs, diagnosed it, and is now testing whether fixing it is sufficient. That is not a story about carelessness. It is a story about how difficult it is to notice that your central assumption is wrong when every intermediate result says it is working.
This case study is the counterweight to Case Study 26.1. Read them together or neither.
Part 1 — The shape of the record
Across the 1990s, 2000s, and 2010s, therapeutic cancer vaccines reached large randomized trials in multiple tumor types, using multiple antigen classes and multiple delivery platforms. Among them:
- A cancer-testis antigen — a protein normally expressed only in germ cells and re-expressed by some tumors — tested in large randomized trials in lung cancer and melanoma.
- A mucin-derived antigen delivered in a liposomal peptide formulation, tested in a large randomized trial in non-small-cell lung cancer.
- A telomerase-derived peptide, tested in pancreatic cancer.
- A mutant growth factor receptor peptide — notably, a genuine tumor-specific target — tested in glioblastoma.
- Numerous dendritic cell, whole-tumor-lysate, and shared-epitope peptide products across melanoma, renal, prostate, and colorectal cancers.
All of the large randomized programs listed above failed to meet their primary endpoints. Some missed narrowly, some decisively. Subgroup signals appeared and, where pursued, generally did not replicate.
And here is the part that should unsettle you, because it is the part that made the failures so hard to anticipate: in many of these programs the vaccines demonstrably did what they were designed to do. Patients were vaccinated; antigen-specific T cells appeared; immune monitoring assays turned positive. The immunology worked. The medicine did not.
That dissociation — immunological success with clinical failure, reproduced across decades and platforms — is the single most important fact in this case study, and it is the reason the chapter insists so hard on keeping the immune readout and the clinical endpoint on separate lines.
The exceptions, and what they teach
Three things sit outside the failure pattern and are worth knowing precisely.
An autologous cellular immunotherapy for metastatic prostate cancer received approval on the strength of an overall survival benefit — without a corresponding improvement in progression measures. That discordance was and remains debated. It demonstrated two things at once: that a therapeutic cancer immunotherapy could extend survival, and that immunotherapies can behave strangely against endpoints designed for cytotoxic drugs. Its individualized manufacturing was also logistically heavy and expensive, which is a directly relevant precedent for anything individualized that follows.
An engineered oncolytic virus injected directly into melanoma lesions holds approval and sits adjacent to the vaccine category — it provokes local immunity as part of its mechanism.
A bacterial preparation instilled into the bladder has been standard therapy for early bladder cancer for decades. It is an immunostimulant rather than a vaccine, and it long predates the molecular era. It is the oldest durable evidence that provoking immunity locally can control a tumor, and it was discovered without any of the mechanistic apparatus in this chapter.
The largest success is not in this category at all
The most effective cancer vaccines in existence are prophylactic antiviral vaccines: hepatitis B vaccination prevents a large fraction of liver cancer, and human papillomavirus vaccination prevents cervical and other cancers. Neither is a peptide vaccine — one is a recombinant surface protein, the other a virus-like particle — and neither treats an existing tumor. They prevent the chronic infection that causes the cancer, in healthy people, years in advance.
That is not a footnote. It is the field's clearest demonstration that the prophylactic problem and the therapeutic problem are different problems, and that we have solved one of them far better than the other.
Part 2 — The diagnosis: four reasons it failed
Reason 1 — Tolerance (the deep one)
Almost every antigen used in those programs was a self protein. Overexpressed, aberrantly expressed, re-expressed, but self.
Your immune system spent your development systematically removing T cells that recognize self peptides with high affinity, and restraining those that slipped through. That is central and peripheral tolerance, and it is not a bug to be engineered around; it is the reason you are not currently attacking your own tissues.
So these vaccines were recruiting from a repertoire that had been purged of exactly the clones they needed. The T cells that could be raised were the leftovers — low-affinity, easily exhausted, easily suppressed. The field was not failing to raise responses. It was raising the wrong quality of response, and the assays used to measure success could not tell the difference.
This is the mistake, and it took a long time to see because every intermediate readout said things were going well.
Reason 2 — Adjuvants and formats
Early peptide vaccines frequently used minimal epitopes — 8- to 10-residue peptides — which load directly onto class I molecules on any cell they encounter, including cells with no costimulatory capacity. That is the recipe for tolerance rather than immunity (§26.4).
Adjuvant choices compounded this. Long-lived water-in-oil depots were common; preclinical work later showed that such depots can retain induced T cells at the injection site, where they are deleted rather than dispatched. Meanwhile the aluminum-salt adjuvants that work beautifully for antibody-directed prophylactic vaccines are poor at generating the cytotoxic responses a cancer vaccine needs.
Reason 3 — No way to keep the T cells working
Suppose a vaccine did generate good T cells. They then had to enter a tumor microenvironment built to shut them down: inhibitory ligands, suppressive cells, metabolic hostility. Checkpoint inhibitors did not exist for most of this history. A vaccine of that era was generating soldiers and sending them into a building filled with sedative gas.
Reason 4 — HLA restriction narrowed everything
Trials commonly enrolled only patients carrying one particular common class I allele. That limited eligible populations, skewed the studied population toward particular ancestries, and meant that even a successful product would have addressed a fraction of patients.
Part 3 — What neoantigen approaches do differently
Three of the four diagnosed causes have specific, mechanistically targeted answers. Be precise about which:
| Cause of failure | What individualized neoantigen approaches do |
|---|---|
| Tolerance | Targets are genuinely foreign — sequences absent from the normal proteome, so no central tolerance was ever imposed and the high-affinity repertoire is intact. This is the decisive change. |
| Adjuvants and formats | Synthetic long peptides instead of minimal epitopes; defined TLR-targeting adjuvants instead of depot emulsions; nucleic acid platforms with intrinsic innate stimulation. |
| No way to sustain the response | Checkpoint blockade now exists and is given alongside. A vaccine that generates T cells plus a drug that keeps them functional address two halves of the same failure. |
| HLA restriction | Improved but not solved. Epitopes are selected against the individual's own HLA rather than a designated common allele — but prediction quality across all HLA backgrounds is not established, and the constraint has been redistributed rather than removed. |
That table is an argument, not a result. Every row describes a change that plausibly addresses a diagnosed cause. None of them demonstrates that the change is sufficient. The history establishes a low prior for this class of intervention; the mechanistic changes are a reason to think this attempt differs in ways that matter; and randomized confirmatory trials are how we learn which consideration dominates.
📊 Evidence Rating
Claim: Therapeutic cancer vaccines, as a general historical class, improve survival in patients with established solid tumors. Rating: ❌ to ⚠️ (as of 2026) Reason: A long record of large randomized trials failing to meet primary endpoints across multiple tumor types, antigens, and platforms — often while demonstrably inducing the intended immune response — against a small number of genuine successes, including one approved autologous cellular immunotherapy with a survival benefit in metastatic prostate cancer. What would change it: each confirmatory randomized success in a defined population moves the corresponding claim upward for that population.
Scope warning. This rates the historical class. It is not a rating of neoantigen approaches, which differ in the specific ways tabulated above. Using this ❌ to dismiss neoantigen vaccines is downgrading by association — the error rule 4 forbids. Using neoantigen optimism to retroactively rescue the historical class is the same error pointed the other way.
Part 4 — The transferable lesson
The most useful thing in this history is not "cancer vaccines failed." It is the anatomy of how a careful field can be wrong for thirty years:
A wrong central assumption can be invisible when every intermediate measurement is favorable. The assumption was that raising a measurable T-cell response against a tumor antigen would translate into tumor control. Immune monitoring assays confirmed the first half repeatedly, which made the assumption feel validated rather than tested.
The surrogate was measuring the thing the investigators controlled, not the thing patients needed. This is a general failure mode. You will meet it again in Chapter 35 and everywhere in this book that a biomarker stands in for an outcome.
The correction came from immunology, not from trial design. No amount of better statistics would have rescued a self-antigen vaccine. What changed the field's prospects was understanding why the responses being generated were the wrong ones — and that understanding produced neoantigens.
And the correction has not yet been validated. Which is exactly what 🔬 means.
Discussion Questions
1. The case study argues that the field's central mistake was invisible because intermediate readouts were favorable. Design an immune monitoring readout that would have exposed the tolerance problem earlier. What would it measure, and why would it have been hard to interpret at the time?
2. Of the four diagnosed causes of failure, the chapter treats tolerance as decisive and HLA restriction as improved-but-unsolved. Argue the case that one of the other three — adjuvants or the absence of checkpoint blockade — was actually the binding constraint. What evidence would settle it?
3. An approved autologous cellular immunotherapy showed an overall survival benefit without a progression benefit. Take that discordance seriously: what mechanisms could produce it, and what does it imply about using progression-based endpoints to evaluate immunotherapies?
4. The most successful cancer vaccines are prophylactic antiviral vaccines that are not peptide vaccines and do not treat existing disease. Does including them under the heading "cancer vaccine" clarify or obscure? Argue both sides, then commit.
5. Someone proposes that the thirty-year failure record should lower our expectations for individualized neoantigen vaccines. Someone else replies that the record is irrelevant because the targets are categorically different. Evaluate both positions. Is there a formulation that takes the history seriously without treating it as decisive?
6. Suppose the confirmatory neoantigen trials report and the results are negative. Write the two paragraphs that would need to be added to this case study — and then write the two that would be needed if the results are strongly positive. Notice which version was easier to write, and consider what that tells you about your own priors.