mRNA Cancer Vaccines: Where the Trials Stand

A few years ago I sent a draft review to a co-author, an immunologist who has spent most of her career on T cell priming. I'd written a paragraph about individualized mRNA vaccines that was, in retrospect, fairly breathless. She sent it back with one comment in the margin: "You've described the manufacturing. Where's the immunology?"

She was right, and it stuck with me. Most of what gets written about mRNA cancer vaccines is really about logistics — sequencing turnaround, cold chain, the engineering achievement of making a bespoke drug for one patient in six weeks. That's genuinely impressive. But it isn't the part that will determine whether these vaccines work, and it isn't the part that phase 3 trials are testing.

So this is my attempt to write the piece she was asking for. What the mechanism actually requires, what the trial data support at each level of evidence, and where I think the field is being sloppy.

The question, stated precisely

The claim under test is narrower than "vaccines can treat cancer." It's something like: in a patient whose bulk disease has been resected, can you expand a polyclonal T cell response against tumor-specific mutations fast enough and broadly enough to eliminate micrometastatic disease that checkpoint blockade alone would miss?

Note what that framing does. It puts these agents in the adjuvant setting, against minimal residual disease, in combination with a PD-1 inhibitor — which is exactly where the serious trials sit. It also means the primary question is one of relapse-free survival in a population where most patients are already cured by surgery. That's a hard place to show an effect, and an even harder place to show it quickly.

The mechanism, and the three places it can fail

The pipeline is familiar by now. Whole-exome and RNA sequencing of tumor and matched normal tissue identifies somatic mutations. An algorithm ranks candidate neoepitopes by predicted HLA binding, expression, and clonality. A subset — typically on the order of twenty to thirty — is encoded into a single RNA construct, formulated in a lipid nanoparticle or lipoplex, and given repeatedly alongside anti-PD-1.

There are three failure points, and I'd argue the field has solved roughly one of them.

1. Picking the right epitopes

This is the weak link. Class I binding prediction has improved substantially with mass-spectrometry-trained models, but the fraction of predicted neoepitopes that elicit a detectable T cell response in a patient remains modest. In the published vaccine series, a meaningful minority of encoded epitopes generate measurable responses, and the responses skew toward CD4 rather than CD8 more than most people expected when these programs started.

Whether that's a bug is genuinely unsettled. CD4 help may be doing real work through licensing of dendritic cells and epitope spreading rather than direct killing. But if your mechanistic model is "vaccine-induced CD8 clones find and kill micrometastases," the immunophenotyping data don't straightforwardly support it. Worth reading the neoantigen prediction and validation literature directly rather than taking summary claims on trust.

2. Delivering RNA to the right antigen-presenting cells

Intramuscular LNP delivery and intravenous lipoplex delivery to splenic APCs are different strategies with different innate signatures. Unmodified uridine-containing RNA engages endosomal TLR7/8 and provides built-in adjuvanticity; nucleoside-modified RNA damps that down and relies more on the carrier. Both approaches are in the clinic. We don't have a head-to-head trial, and I'd be surprised if we get one.

This matters because the innate sensing of the formulation may contribute as much to outcomes as the encoded antigens. There's preclinical work showing that RNA-lipoplex platforms can sensitize tumors to checkpoint blockade in ways not fully attributable to the specific sequences carried. That's an uncomfortable confound for a personalization narrative.

3. Working in a suppressive tumor microenvironment

Even a well-primed clone has to traffic, infiltrate, and function. This is why combination with PD-1 blockade is not an add-on but a structural assumption, and why these vaccines are being tested after resection rather than against bulky disease.

What the evidence actually supports

Randomized phase 2: the melanoma signal

The strongest randomized evidence comes from the adjuvant melanoma setting, where an individualized neoantigen mRNA vaccine plus pembrolizumab was compared with pembrolizumab alone in resected high-risk disease. The trial randomized roughly 150 patients, 2:1, and reported a relapse-free survival advantage for the combination, with a hazard ratio in the region of 0.55 to 0.6 and distant metastasis-free survival moving in the same direction. Benefit appeared across tumor mutational burden and PD-L1 subgroups, which is reassuring if it holds.

I want to be careful about how much weight that bears. A randomized phase 2b of this size gives you a wide confidence interval, an unblinded relapse endpoint assessed by investigators, and no overall survival data worth interpreting. It's a signal that justified phase 3. It is not an effect size you should quote as if it were established. The published melanoma vaccine literature is still small enough to read in an afternoon.

A second-year PhD student emailed me last spring asking whether she could cite that hazard ratio as "a roughly 44% reduction in recurrence risk." Technically yes. But I told her to write the sentence with the sample size and the phase in it, because the moment you strip those out, the number starts behaving like a fact. She rewrote it. The paragraph got better.

Early-phase: pancreatic cancer, and the most interesting biology

The pancreatic adenocarcinoma work is, to my mind, scientifically the most informative even though it is the weakest evidence. A small phase 1 study of an individualized mRNA vaccine given after surgery with chemotherapy and atezolizumab found that about half of patients mounted high-magnitude vaccine-induced T cell responses, and that those responders had markedly longer recurrence-free survival than non-responders. Follow-up reporting suggests some of those clones persist for years.

That is a correlation within a single arm, in a tiny cohort, with obvious confounding — patients who mount strong immune responses may differ in ways that independently predict outcome, including residual disease burden and splenectomy status. It is not evidence of efficacy. It is evidence that the mechanism can engage in a tumor type with low mutational burden and a famously cold microenvironment, which few people predicted. The randomized follow-on is the trial that matters; track the pancreatic neoantigen vaccine reports and the corresponding registrations on ClinicalTrials.gov rather than conference coverage.

Shared-antigen and off-the-shelf constructs

Parallel programs encode fixed tumor-associated antigens — melanoma differentiation antigens, HPV16 E6/E7 for head and neck cancer, KRAS driver mutations. These are cheaper, faster, and testable in larger populations. Evidence so far is early-phase: immunogenicity, tolerability, scattered responses. Nobody should be making efficacy claims here yet, but if a fixed construct ever matches a personalized one, the economics of the field change overnight.

What we do not have

No meta-analysis worth the name. No completed phase 3. No overall survival data. No validated surrogate linking vaccine-induced T cell magnitude to clinical benefit at the individual patient level. Anyone pooling across these heterogeneous platforms right now is generating a number, not evidence.

Where I think the field is being careless

Cross-trial comparison. I reviewed a manuscript last year that tabulated response rates across four vaccine platforms in four tumor types with four different backbones and drew conclusions about which construct was better. I recommended rejection, and I'd do it again. The control arms aren't comparable, the immune assays aren't comparable, and the patient populations differ in resection status and prior therapy.

Treating RFS as settled currency. In adjuvant melanoma, relapse-free survival benefit has historically tracked reasonably well with longer-term outcomes, but "reasonably well" is doing a lot of work. In other tumor types that relationship is less established. A phase 3 that hits RFS and misses OS will be argued about for a decade.

Ignoring the control-arm problem. These trials compare vaccine plus PD-1 against PD-1 alone, usually without a sham injection schedule. Investigator-assessed recurrence in an unblinded trial is not immune to expectation effects, particularly when imaging intervals are protocol-driven but clinical suspicion isn't.

Underweighting attrition. Manufacturing takes weeks. Patients relapse during that window, or deteriorate, or decline. How those patients are handled in the analysis population materially affects the estimate, and it's often buried in supplementary material. Read the CONSORT diagram before you read the Kaplan-Meier curve.

What the phase 3 trials need to show

Here's what I'll be looking for when the large adjuvant melanoma and non-small cell lung studies report, and what I'd push a student to look for too:

• A hazard ratio that survives blinded central review, not just investigator assessment, with a confidence interval that excludes trivial benefit rather than just excluding one.

• Consistency across the prespecified subgroups that matter mechanistically — tumor mutational burden, PD-L1, HLA diversity. If benefit concentrates only in high-TMB patients, personalization is doing something different from what's claimed.

• Immune correlates reported as prespecified analyses, not post hoc. Magnitude of vaccine-induced response versus outcome, with the obvious confounders modeled.

• ctDNA dynamics. Minimal residual disease detection is becoming the natural stratifier for adjuvant trials, and the ctDNA-guided adjuvant literature is maturing quickly. A vaccine trial that shows clearance of molecular residual disease has a far more interpretable mechanism story than one that only reports time-to-event curves.

• Honest reporting of manufacturing failures and delays, including patients randomized but never dosed.

• Immune-related toxicity in a curative-intent population. Adding an immunostimulant to anti-PD-1 in patients who may already be cured raises a different risk-benefit calculus than in metastatic disease.

How I'd follow this if it isn't your main field

Most people reading this aren't running vaccine trials. You're adjacent — you work on antigen processing, or sequencing methods, or you're a clinician who'll eventually be asked about this in clinic. The practical problem is that the literature moves across immunology, oncology, bioinformatics, and regulatory science simultaneously, and reading it piecemeal means you lose the thread between a preclinical mechanism paper and the trial that depends on it.

What works for me is keeping a single running document per question — one for epitope prediction performance, one for adjuvant trial design — where the search, the notes, and the draft argument all live in the same place. That's roughly the workflow PACR's research workspace is built around: you search, read, summarize, and write without exporting context between tools, so the reasoning you did in week one is still attached to the paper in week six. For a fast-moving area where your interpretation keeps shifting, that continuity matters more than any individual feature. If you want to try that pattern on your own reading list, you can set up a workspace and point it at the searches above.

What I actually think

I think the melanoma phase 3 reads out positive, modestly, and that the field over-reads it. I think the more consequential result will come from a low-mutational-burden tumor type, because that's where the mechanism is doing something we can't explain by "more mutations, more targets." And I think within five years the interesting debate won't be whether personalized vaccines work, but whether personalization earns its cost relative to a well-chosen fixed construct plus a better adjuvant.

My co-author's question still applies. When you read the next trial report, skip the turnaround times. Go to the immunogenicity supplement and ask what the T cells actually did.