Case Study 26.1 — From Biopsy to Injection

Reading the individualized neoantigen vaccine approach as a study design

This case study is not about a molecule. It is about a procedure — the sequence of operations that turns a patient's tumor into a product made for that patient and nobody else — and about how to read the evidence such a procedure generates.

That distinction is the whole reason this case study exists. Everywhere else in this book, when you evaluate a claim, there is a defined chemical entity on the other end of it. Semaglutide is semaglutide. A trial of semaglutide in one population tells you something about semaglutide in another, because the molecule does not change. Here, the molecule changes every time. Two patients in the same trial arm receive products with no sequence in common. What the trial tests is the pipeline that made both.

Reading that correctly requires a different kind of attention, and this is where we practice it.

Part 1 — The procedure, step by step

Step 1: Tissue

Two samples are required: tumor and matched normal, the latter usually blood.

The tumor sample must be adequate in quantity and quality for exome sequencing and, ideally, RNA sequencing. This sounds like a formality and is not. Archival tissue from a diagnostic biopsy taken months earlier may be too small, too degraded, or too depleted by other tests. Ineligibility at this step is common, and it is not random — it correlates with how the patient's disease was worked up, which correlates with where they were treated, which correlates with a great deal else. Any trial population assembled this way has been filtered before randomization ever happens.

The matched normal sample is what makes the whole thing possible. Without it you cannot distinguish a somatic mutation — acquired by the tumor — from an inherited variant present in every cell the patient owns. Targeting an inherited variant would mean vaccinating a person against themselves.

Step 2: Sequencing and variant calling

Whole exome sequencing of both samples; RNA sequencing of the tumor where possible. Compare. What is present in tumor and absent in normal is somatic.

The RNA matters more than it sounds. A mutation in a gene the tumor does not express produces no protein, and therefore no peptide, and therefore no target. Expression filtering removes a large fraction of nominal candidates.

Step 3: HLA typing

The patient's own class I and class II alleles are determined, usually from the same sequencing data. Everything downstream is computed against these alleles. This is the step that makes the product personal in a second, independent sense: not only are the mutations the patient's own, but the selection of which mutations to use is made against the patient's own presentation machinery.

Step 4: Prediction and ranking

For each somatic mutation, candidate peptides spanning the altered residue are generated and scored on several axes at once:

• Will the peptide be produced by cellular processing?
• Will it bind this patient's HLA?
• Is the gene expressed, and at what level?
• Is the mutation clonal — in every tumor cell — or subclonal?
• Does the peptide look immunogenic, by whatever model the group uses?

This is the step where the approach is most vulnerable, and it is worth being blunt about why. Binding prediction has improved enormously and is genuinely good. Immunogenicity prediction — will this peptide actually raise a response in this person — is a much harder problem and remains substantially less accurate. In practice, only a minority of predicted epitopes elicit detectable T-cell responses when tested. A pipeline that selects twenty targets should be expected to generate responses against fewer, sometimes considerably fewer.

Step 5: Selection and manufacture

A set is chosen — programs have used from roughly ten to a few dozen targets per patient — and the product is made: either the individual peptides synthesized to pharmaceutical standard, or a single nucleic acid construct encoding all selected epitopes end-to-end.

Then release testing: identity, purity, sterility, potency, on a batch whose entire market is one person.

Elapsed time from tissue to product is measured in weeks. That timeline is a clinical variable, not a logistical footnote. It is why the adjuvant setting — after surgery, low tumor burden, no evident disease — is where this approach is most plausible: it is the setting in which a wait costs least.

Step 6: Administration and monitoring

Given with an adjuvant, and in current trials alongside a checkpoint inhibitor. Monitoring runs on two tracks that must not be conflated: immune readouts from blood, and clinical outcomes from scans and follow-up.

WHERE THE PIPELINE CAN FAIL

  step                        failure mode                       detectable when?
  ─────────────────────────── ────────────────────────────────── ────────────────
  tissue                      inadequate or unrepresentative     immediately
  variant calling             subclonal targets chosen           at recurrence
  HLA typing                  rare allele, poor reference data   at immune readout
  prediction                  epitopes that bind but do not      at immune readout
                              raise a response
  manufacture                 too slow for the disease           at progression
  administration              wrong adjuvant, wrong site         at immune readout
  in the patient              suppressive microenvironment       at progression
  in the tumor                antigen loss, HLA loss             at recurrence

  Note the pattern: the failures that matter most are detected LAST.

Part 2 — Reading the evidence this procedure has generated

🔬 Read the Study — the two generations, and what each design can support

Generation one: single-arm, first-in-human (published from 2017).

Design. Small studies in melanoma — each with fewer than twenty patients — administering either synthetic long peptides with a TLR-agonist adjuvant or an RNA-based construct, with immune monitoring as the principal readout.

What they legitimately established. That the pipeline can be executed end to end in a clinically usable timeframe, and that vaccinated patients develop T-cell responses against neoepitopes that were undetectable beforehand. Both are real findings and neither is trivial. Feasibility studies exist to answer feasibility questions, and these answered them.

What they cannot establish. Anything about efficacy. There is no comparator. Patients in these studies had heterogeneous disease, received other therapy, and were selected for enrollment by criteria correlated with prognosis. A favorable outcome in such a study is compatible with a large treatment effect, a small one, and none.

The trap. Reports of these studies frequently pair an immune result with a clinical observation in the same sentence — "patients developed neoantigen-specific T cells, and most remained recurrence-free at follow-up." Both halves may be true. The sentence structure implies a relationship the design cannot support.

Generation two: randomized, vaccine added to checkpoint blockade.

Design. All patients receive a checkpoint inhibitor; patients are randomized to also receive the individualized vaccine. Clinical endpoints, typically recurrence-based in the adjuvant setting.

Why this design is right. It is the only way to isolate the vaccine's contribution from the checkpoint inhibitor's. The comparator is active therapy, which is both the ethical requirement and the analytical difficulty.

Why it is hard. Detecting an increment on top of an effective therapy requires more patients and longer follow-up than detecting an effect against nothing. Recurrence endpoints take years to accumulate. And if benefit concentrates in a subgroup, an overall analysis may dilute it while a subgroup analysis that finds it will be difficult to trust.

Status as of 2026. The melanoma work in this generation has been encouraging enough to support larger confirmatory trials, and those trials are running. Programs are also underway in other tumor types. No individualized neoantigen cancer vaccine has general approval anywhere. I am not quoting effect sizes here; if you need them, take them from the primary publication with the confidence interval attached.

Honest limitations, stated plainly

1. The product is a process. External validity works differently here. A positive trial in melanoma establishes something about the pipeline as applied to melanoma, in that disease setting, with that manufacturing operation. Whether it transfers to a tumor type with a different mutational landscape, or to a manufacturing site with different practices, is a genuinely open empirical question.

2. Prediction is the weak joint. The step that makes individualization possible is also the least accurate step, and its accuracy is not evenly distributed — reference data and algorithms have been trained on unevenly sampled populations, and performance across all HLA backgrounds is not established.

3. Combination confounds interpretation. Randomization solves attribution in principle. It does not make the resulting estimate precise, and it does not tell you whether the vaccine would do anything without the checkpoint inhibitor.

4. Selection precedes randomization. Patients must have adequate tissue, survive the manufacturing interval in a condition to be treated, and meet trial criteria. That is a healthier, better-resourced population than the one a therapy would eventually serve.

5. The endpoint is usually not survival. Recurrence-free survival is a reasonable and often necessary endpoint. Translating it into a survival benefit is an inference.

6. Tumors with few mutations offer little raw material. This is structural, not temporary.

7. Cost and access are not solved. A therapy whose manufacturing is individualized has an economics problem that does not shrink with scale in the ordinary way.

📊 Evidence Rating

Claim: Individualized neoantigen vaccines improve clinical outcomes (recurrence-free or overall survival) in patients with solid tumors. Rating: 🔬 (frontier, as of 2026) Reason: The mechanism is established and the approach is proceeding properly through randomized trials with encouraging early signals, but no confirmatory randomized trial has established durable clinical benefit, nothing is approved, and combination with checkpoint inhibitors makes attributing benefit to the vaccine component genuinely difficult. What would change it: completed, adequately powered randomized trials reporting survival or durable recurrence-free survival benefit in defined tumor types and settings, with the vaccine arm separable from the checkpoint inhibitor's contribution by design.

Part 3 — What this case study is really teaching

Three transferable habits.

Read the design before the result. For every claim in this space, the design determines the ceiling on what can be concluded. A single-arm study cannot yield an efficacy conclusion no matter how good the numbers look, and no amount of mechanistic elegance repairs that.

Keep the immune readout and the clinical endpoint on separate lines. The history of this field is full of interventions that produced exactly the immune response they were designed to produce and changed nothing that mattered to patients. When those two readouts appear in the same sentence, split them.

Distinguish "the idea has not been shown to work" from "the idea does not work." Those are different statements, and 🔬 is the symbol for the first. This approach may turn out to be one of the significant advances in oncology of this era. It may also join the long list in Case Study 26.2. As of 2026, the evidence does not settle it — and the people running the confirmatory trials are the ones doing something about that.

Discussion Questions

1. The case study argues that "the product is a process" changes how external validity works. Work through a concrete example: a pipeline succeeds in melanoma and is then applied to a tumor type with a much lower mutation count. Which specific steps would you expect to behave differently, and what would you want to see before treating the melanoma result as supportive?

2. Ineligibility at the tissue step is described as "not random." Trace the consequences: how could tissue-based eligibility criteria bias a trial population, and would that bias tend to make the vaccine look better or worse? Does randomization fix it?

3. Immunogenicity prediction is identified as the weakest joint in the pipeline. Suppose it improved dramatically tomorrow. Which of the seven listed limitations would that resolve, which would it leave untouched, and would it change the chapter's rating?

4. Consider the sentence: "Patients developed neoantigen-specific T cells, and most remained recurrence-free at follow-up." Rewrite it so that both facts are preserved and no relationship is implied. Then explain why the original version is so persistent in scientific writing, not only in journalism.

5. The adjuvant setting — after surgery, minimal residual disease — is argued to be the most favorable for this approach. Construct the counterargument: what is harder about proving benefit in that setting than in advanced disease? Consider event rates, follow-up duration, and what happens to a trial when most patients in both arms do well.

6. A patient's family member asks you why, if the vaccine is made from their own tumor's mutations, it needs to be tested in a randomized trial at all — "it is personalized, so of course it fits." Answer them. Your answer should be honest about what personalization does and does not guarantee, and should not require them to already understand tolerance, HLA, or checkpoint blockade.