Personalized Cancer Vaccine Clears its First Phase 3 Trial

Ronni Holmvig Strøm · 2026-08-23

Moderna's stock closed up about 177% on Wednesday, on a press release that carried no efficacy numbers. The trigger was an interim readout of a Phase 3 melanoma trial, INTerpath-001, which Moderna and Merck say met its endpoints, with the actual data held for an upcoming medical meeting. The market was pricing the first randomized late-stage win for a therapy whose individual doses are designed, one patient at a time, by an algorithm.

The regimen pairs Merck's Keytruda with intismeran, the personalized mRNA drug formerly known as mRNA-4157. The trial enrolled 1,137 patients with surgically removed stage IIB to IV melanoma, and met its primary endpoint, recurrence-free survival, and a key secondary, distant metastasis-free survival. Built on the same mRNA platform as the Covid vaccine, intismeran does something the earlier product never had to. Every dose is a different molecule.

The Algorithm Inside the Production Loop

For each patient, Moderna sequences the tumor's DNA and RNA, compares it against healthy blood, and reads the patient's HLA type, the molecular machinery that decides which protein fragments the immune system can even see. That produces a search problem. A melanoma tumor carries hundreds of mutations, and Moderna's own patent puts the candidate mutanome at roughly 100 to 10,000. Almost all of them are useless as targets. The job is to find the small subset that genuinely belongs to the tumor, is expressed by it, can be displayed by that patient's particular HLA molecules, and will provoke a strong enough immune response to matter.

Moderna's system reviews the full mutation list and predicts up to 34 neoantigens most likely to trigger a T-cell attack, and those become the design spec for a single synthetic mRNA. Merck describes the contents of intismeran plainly, as "algorithmically derived" neoantigen sequences.

A caution the moment deserves: this is not "a neural network designed the drug." Moderna has not disclosed its production model, which it calls a proprietary algorithm developed with Merck. What can be said without stretching is that Moderna itself calls the system AI, its patent explicitly encompasses machine-learning methods from random forests to neural networks, and algorithmic target selection is intrinsic to the product. There is also older, blinded evidence the selection works. In a 2020 study across colorectal tumors where the reactive targets were already known, Moderna ran the mRNA-4157 algorithm across the sequencing data and captured 26 of 64 experimentally confirmed neoantigens, hitting at least one demonstrated target in 18 of 29 patients.

The Product Is the Pipeline, Not the Vial

Most "AI-designed drug" stories share one shape. A model helps find molecule X, molecule X gets manufactured, everyone receives molecule X. Intismeran breaks that shape. The algorithm looks at your tumor and designs your medicine, then looks at the next patient and designs a chemically different one. The product is the whole loop: sequencing, the predictive model, an mRNA compiler, and a manufacturing line that turns the output into a physical dose. The tumor supplies the input, the algorithm writes the program, and the factory compiles it.

Patient 1,137 does not receive batch 1,137 of one medicine. They receive drug 1,137, with its own sequence, its own chain of identity, and its own clinical deadline. Moderna built a system called Maestro to orchestrate exactly that, coordinating manufacturing capacity, quality control, scheduling and shipping across thousands of individualized batches. The turnaround runs a few weeks from biopsy to patient. Personalized cancer vaccines have been scientifically plausible for years. The wall was always making a unique pharmaceutical product for thousands of people without the logistics collapsing, and a second layer of AI is what holds that wall up.

What the Trial Actually Validated

The hard efficacy numbers are still the ones from June: a five-year Phase 2b readout showing a 49% reduction in the risk of recurrence or death and a 59% reduction in distant metastasis or death, both against Keytruda alone. The Phase 3 effect size, the subgroups, and overall survival are not public. Keytruda is doing real work in the regimen, surgery had already removed the visible tumor, and none of this is "AI cured cancer."

What the randomized trial did do is subtler and larger. Each patient in the treatment arm received a different, computationally designed therapy keyed to their own tumor, and the trial still beat its control. In melanoma, that validates an architecture in which algorithmic prediction sits between diagnosis and drug manufacture. A cancer therapy whose individual design depends on AI has now succeeded in Phase 3.