AI Neoantigen Cancer Vaccines Are Heading to Phase 3

For decades, the dream of a personalized cancer vaccine sat just out of reach. Tumors mutate uniquely in every patient, and identifying which of those mutations the immune system can actually attack required computation that did not exist outside research labs. In 2026, that bottleneck is gone. AI-driven neoantigen prediction now powers a wave of mRNA cancer vaccines moving toward FDA approval, with five-year follow-up data showing durable benefit in melanoma, pancreatic, and colorectal cancers. The first regulatory submissions land this year.

This is not the recurring “cancer cure breakthrough” headline. The scientific machinery — tumor sequencing, neoantigen prediction, mRNA manufacturing, immune checkpoint pairing — is real, validated in randomized trials, and being scaled by drug companies that already shipped billions of mRNA doses during COVID. What changed is the AI layer that picks which mutations to target. That choice used to take weeks of expert curation. Today it takes hours, and the picks are better.

Why neoantigens, and why this finally works

A neoantigen is a fragment of a mutated protein that exists only inside cancer cells. Because the mutation does not appear in healthy tissue, the immune system can in principle recognize it as foreign and destroy any cell carrying it — without harming normal cells. This is the holy grail of oncology: a treatment specific to the tumor, with the precision of a targeted drug and the durability of an immune response.

The technical problem is staggering. A single solid tumor can carry hundreds to thousands of mutations. Only a small minority produce peptides that bind tightly to the patient’s specific HLA molecules, get presented on the cell surface, and actually trigger T-cell recognition. Picking the wrong neoantigens wastes the vaccine; picking the right twenty produces durable tumor-specific immunity.

Older neoantigen pipelines used hand-curated rules, simple binding affinity calculators, and expert immunologist review. They worked, but slowly and inconsistently. Modern pipelines use deep learning models trained on mass spectrometry datasets containing millions of confirmed peptide-HLA binding events, immunopeptidomics from real tumors, and T-cell receptor repertoire data from clinical responders. The models predict not just binding affinity but presentation likelihood, immunogenicity, and clonality — the same multi-step cascade an actual antigen-presenting cell performs.

The result: vaccine candidates today contain fewer dud peptides and more functional ones. That difference shows up directly in trial outcomes.

The melanoma data that flipped the field

The data point that flipped the field was the KEYNOTE-942 trial of mRNA-4157, Moderna and Merck’s individualized cancer vaccine, in high-risk resectable melanoma. The Phase 2b data, now extended to five-year follow-up, shows the combination of mRNA-4157 plus pembrolizumab cut the risk of recurrence or death by 49% versus pembrolizumab alone.

Eighteen-month recurrence-free survival hit 78.6% in the combination arm versus 62.2% on monotherapy. At five years, the curves remain separated, which is what oncologists actually care about — late relapses are the killers in melanoma, and a vaccine that prevents them changes the disease’s natural history. The Phase 3 trial, INTerpath-001, is fully enrolled. Moderna has signaled regulatory submission in 2026.

Each vaccine is bespoke. The patient’s tumor and matched normal tissue are sequenced, mutations are called, and AI ranks the top candidates for inclusion. Up to 34 neoantigen-encoding mRNA sequences are then synthesized into a single multi-epitope construct, lipid-formulated, and shipped back to the clinic. Total turnaround used to be nine weeks. Automated GMP platforms have compressed it to under four weeks for several developers.

Pancreatic cancer: the harder proof

Melanoma was always the easier target. It has high mutational burden, responds well to checkpoint inhibitors, and serves as the historical proving ground for immuno-oncology. Pancreatic ductal adenocarcinoma is the opposite — low mutation count, brutal stromal barrier, dismal five-year survival. If neoantigen vaccines work in pancreatic cancer, they likely work everywhere.

BioNTech’s BNT122 (autogene cevumeran), partnered with Genentech, ran a Phase 1 study in post-resection pancreatic cancer patients. Sixteen patients received personalized vaccines encoding 20 neoantigens each, paired with atezolizumab and standard chemotherapy. Eight developed strong vaccine-induced T-cell responses. In that responder subset, recurrence-free survival hit 100% at 18 months. Long-term follow-up presented at recent oncology meetings showed six of those eight responders still alive at six years, with several remaining recurrence-free — an unprecedented result for a cancer where median survival is typically measured in months.

The Phase 2 randomized trial, IMCODE 003, is now enrolling roughly 260 patients. If it confirms the Phase 1 signal, pancreatic cancer treatment changes structurally for the first time in a generation.

The AI stack underneath

The pipeline that produces a personalized vaccine looks roughly like this:

1. Sample processing — Paired tumor-normal whole-exome sequencing at over 100x coverage, plus tumor RNA-seq to confirm the mutated genes are actually expressed. HLA typing of the patient is done in parallel.
2. Variant calling — Standard bioinformatics identifies somatic single-nucleotide variants, indels, and frameshift mutations unique to the tumor.
3. Peptide generation — In silico translation produces all candidate 8-11mer peptides spanning each mutation site for MHC class I and longer peptides for class II.
4. Neoantigen prediction — This is where AI does the work. Deep learning models like NetMHCpan, MHCflurry, and proprietary BigMHC variants score each peptide for HLA binding affinity, presentation probability, and predicted immunogenicity. Newer transformer-based models trained on real immunopeptidomics datasets outperform earlier rule-based scorers by significant margins.
5. Ranking and selection — Candidates are filtered for clonality (preferring trunk mutations present in most tumor cells), expression level, and predicted T-cell response. Ten to thirty-four candidates make the cut.
6. mRNA design and manufacturing — Selected neoantigens are concatenated into a single mRNA construct with optimized linkers, capped, purified, and encapsulated in lipid nanoparticles using automated GMP equipment.

The whole loop is converging on the kind of fast, automated workflow we’ve covered in adjacent diagnostics — see how AI lab-on-a-chip is now picking cancer drugs in hours and how AI companion diagnostics match patients to existing therapies. Personalized vaccines are the natural extension of that same trend toward per-patient oncology decisions.

What’s still hard

The science is converging faster than the infrastructure can keep up. Three problems define the next phase.

Cost. A personalized vaccine currently costs north of $100,000 to manufacture per patient. The economics work for late-stage melanoma, where alternatives are expensive and outcomes are poor. They get harder for adjuvant settings in earlier-stage disease, and harder still for global access. Manufacturing automation, shared sequencing infrastructure, and improved mRNA stability are pushing costs down, but the curve has to bend further.

Manufacturing capacity. Producing thousands of bespoke mRNA constructs per month requires GMP capacity that did not exist before COVID-era investment. It exists now, but scaling to tens of thousands of patients per year — never mind hundreds of thousands — will require a build-out comparable to the early COVID vaccine rollout, this time in personalized batches of one.

Regulatory frameworks. The FDA and EMA are working out how to approve a product where every dose is technically different. The vaccine “platform” gets approved; individual patient doses are produced under that umbrella. This works for cell therapies like CAR-T, where similar precedent exists — and where AI is now cracking the $400,000 CAR-T manufacturing problem. The vaccine playbook borrows heavily from that template.

Who’s building what

Beyond Moderna/Merck and BioNTech/Genentech, a deep bench of programs is in motion. Gritstone bio’s individualized neoantigen vaccines, integrated with checkpoint inhibitors, target colorectal and lung cancers. Geneos Therapeutics is running Phase 2 trials of its DNA-based personalized vaccines in hepatocellular carcinoma. Achilles Therapeutics is taking a different route, manufacturing tumor-infiltrating lymphocytes targeted to clonal neoantigens.

On the platform side, sequencing companies and bioinformatics vendors are positioning themselves as the picks-and-shovels layer. Bigger pharma — Roche, Novartis, AstraZeneca — are watching from the sidelines or quietly licensing AI prediction tech. Expect M&A activity once the first approval lands.

What it means for patients

If the Phase 3 readouts hold and FDA submissions succeed, the first AI-designed personalized cancer vaccines will be commercially available in late 2026 or 2027 — initially in adjuvant melanoma, with pancreatic, colorectal, and renal indications likely to follow within two to three years. Coverage will start in academic centers and large oncology networks with the sequencing infrastructure to support per-patient manufacturing. Insurance reimbursement will be the next battleground.

For patients with high-risk early-stage cancers, the question to ask oncologists in 2026 is simple: am I a candidate for a clinical trial, or for one of these vaccines once approved? For patients in metastatic disease, combination strategies — vaccine plus checkpoint inhibitor plus targeted therapy — are where the most active research sits. We’ve seen this kind of multi-modal AI-guided treatment selection already in adjacent areas, including AI-driven drug resistance prediction and AI digital twins guiding multiple myeloma treatment. Personalized vaccines slot into that same precision-oncology stack.

The bottom line

Personalized cancer vaccines are not speculative anymore. The AI prediction layer works, the mRNA manufacturing works, and the clinical data is mature enough to justify FDA submissions. The next 18 months will determine whether this becomes the next major modality in oncology — alongside checkpoint inhibitors, CAR-T, and ADCs — or whether scaling and reimbursement bottlenecks slow the rollout to a crawl. Both outcomes remain possible. The science, for the first time in this story, is not the limiting factor.

That alone is news.