The Algorithm Inside Moderna's Cancer Vaccine, and the Pharma AI Race It Reframes
- David Borish
- 1 day ago
- 6 min read
On August 19, 2026, Moderna and Merck announced that their personalized mRNA cancer vaccine met the main goal of a large late-stage trial in melanoma. The INTerpath-001 study enrolled 1,137 patients whose cutaneous melanoma had been surgically removed but who faced a high risk of recurrence. Patients received either intismeran autogene combined with Merck's Keytruda or Keytruda alone. The combination produced statistically significant improvements in recurrence-free survival and in distant metastasis-free survival, a key secondary endpoint. The companies reported no new safety signals. It was the first time an individualized mRNA neoantigen therapy cleared a Phase 3 trial.
Moderna shares rose more than 120% on the day, the largest single-session gain in the company's history, lifting the stock to a three-and-a-half-year high. Merck gained roughly 9%. The reaction reflected both the clinical result and what it implies for a modality that skeptics had written off as a COVID-era story.
How the vaccine is built
The therapy is manufactured one patient at a time. After surgery, the patient's tumor tissue undergoes whole exome and RNA sequencing. A computational pipeline then predicts which mutated peptides, called neoantigens, are most likely to be displayed by that patient's own immune-recognition molecules and seen by T cells. A bespoke synthetic mRNA encoding up to 34 of those neoantigens is produced, encapsulated in lipid nanoparticles, and injected. The body's cells translate the mRNA into the target proteins, which train the immune system to recognize and attack cells carrying those mutations. Keytruda releases a brake on the immune system so the newly primed T cells can act.
Each dose is tailored to a single person's cancer. Merck's Jane Healy, head of oncology early development, said the process takes about six weeks from tumor sample to finished vaccine, and patients receive Keytruda during that window.
The prediction step is where the computation matters. A tumor can carry hundreds of somatic mutations, and only a fraction will generate peptides the immune system can actually see. Choosing the wrong targets wastes doses on proteins that provoke no response. Moderna's neoantigen selection algorithm ranks candidates by predicted immunogenicity, and a 2020 study on colorectal tumors found the algorithm selected neoantigens that matched reactivity in patient tumor-infiltrating lymphocytes with high accuracy. Doing this by hand across hundreds of mutations, for every patient, and repeating it at commercial scale, is not practical. The machine learning is what makes per-patient personalization a manufacturing process rather than a research project.
The evidence trail behind the Phase 3 win
The result did not come from nowhere. The mid-stage KEYNOTE-942 trial, which paired the vaccine with Keytruda in resected stage III/IV melanoma, reported a 44% reduction in the risk of recurrence or death at its first readout, then held a 49% reduction at three-year and five-year follow-ups. A five-year update presented at the 2026 ASCO meeting also showed a 59% reduction in the risk of distant metastasis or death. The Phase 3 topline is consistent with that trajectory, though Moderna and Merck have not yet released the specific hazard ratios and plan to present full data at a medical meeting.
The companies are running nine Phase 2 and Phase 3 trials of the vaccine across tumor types. Late-stage studies are under way in non-small cell lung cancer, with mid-stage trials in bladder and kidney cancers and earlier work in pancreatic and gastric tumors. Robert Langer, a Moderna cofounder and MIT professor, called the melanoma data a potential paradigm shift and said he saw no reason the approach would not extend to other cancers. That optimism remains a hypothesis awaiting data. Melanoma carries a high mutation burden, which gives the neoantigen approach more targets to work with. Cancers with fewer mutations may prove harder, and the lung, bladder, and kidney readouts will test whether the melanoma outcome generalizes.
What "AI" does and does not mean here
The headlines framed this as AI curing cancer. The reality is narrower and more useful to understand. The computation Moderna uses is a specialized bioinformatics and machine learning pipeline for neoantigen prediction, sequence optimization, and manufacturing orchestration. Moderna built a digital system it calls Maestro to manage the per-patient manufacturing workflow, giving cross-functional views of production for each individual lot. The current biopsy-to-patient turnaround of roughly four to eight weeks is a scheduling and logistics challenge as much as a modeling one, and compressing it is the main obstacle between the therapy and delivery at commercial scale.
The vaccine did not shorten the clinical trials, avoid the biology of efficacy and safety, or replace the immunologists who designed the program. What the algorithm did was make individualized targeting feasible across thousands of patients. That is a real contribution, and it is different from the claim that a model discovered or designed the drug from scratch.
The wider pharma AI race
The Moderna result lands in the middle of a broader push to use machine learning in drug development, and it reframes how that push should be judged. The other side of the field is drug discovery, where AI-native companies design novel molecules rather than personalizing an existing modality.
By mid-2026 the discovery field had a measurable clinical pipeline and a mixed record. An analysis presented at ASCO counted 117 AI-enabled therapeutic assets across 63 companies in human trials, of which about 51% had completed Phase 1 and roughly 7% had completed Phase 2. Estimates of the broader universe run to more than 170 programs. No AI-discovered drug has yet received FDA approval.
Insilico Medicine has the field's clearest clinical proof point. Its rentosertib, a drug for idiopathic pulmonary fibrosis that the company says was both discovered and designed with AI, posted positive Phase 2a results published in Nature Medicine, showing dose-dependent improvement in lung function. Insilico reached Phase 1 from concept in about 30 months, well under the usual timeline, and registered a Phase 3 study. Recursion Pharmaceuticals, which absorbed Exscientia, reported that its familial adenomatous polyposis candidate reduced polyp burden in 75% of patients in an early trial, with durable responses at week 25, while discontinuing several other programs. Isomorphic Labs, the Google DeepMind spinout built around AlphaFold, has raised more capital than any peer and has said it aims to dose its first patient with an AI-designed candidate by the end of 2026, a milestone it has not yet reached.
The money has moved ahead of the clinical evidence. Eli Lilly signed a collaboration with Insilico worth up to $2.75 billion in March 2026, the largest AI-era pharma deal on record, and committed up to $1 billion with NVIDIA to build an AI research supercomputer. Isomorphic's deal book with Lilly and Novartis approaches $3 billion in upfront and milestone value, and Johnson & Johnson joined as a third pharma partner. More than $11 billion flowed into the space across roughly 348 rounds in 2025. Repeat partnerships from multiple top-ten drugmakers are a stronger signal than any single funding round, because they reflect independent diligence rather than investor enthusiasm.
The honest read, echoed by several people in the field, is that AI has demonstrably compressed early discovery timelines but has not yet been shown to improve the roughly 90% failure rate that defines clinical development. AI-discovered compounds have so far progressed through trials at rates similar to conventionally discovered molecules. The first positive Phase 3 readout from a current-generation AI discovery platform would test whether the early speed translates into better odds of success.
Why Moderna's result matters for the category
Moderna's vaccine is not a discovery-platform drug, so it sits in a different bucket from Insilico or Isomorphic. But it offers the category something it has lacked, which is a late-stage clinical win for a therapy that could not exist without heavy computation. The neoantigen selection and per-patient manufacturing are the load-bearing pieces, and they cleared Phase 3. That is a proof point for computation embedded in a therapeutic workflow, even if it is not the proof point the pure discovery companies still need.
The distinction is worth holding onto as the coverage settles. Moderna showed that machine learning can make an individualized therapy work at trial scale in melanoma. Whether that extends to lower-mutation cancers, and whether AI-designed molecules can clear late-stage efficacy trials, are separate open questions. Both will be answered by data over the next two years rather than by the size of the next deal.
Next steps
Moderna and Merck plan to present full INTerpath-001 data at an upcoming medical meeting and to engage regulators on filings. The lung, bladder, and kidney readouts will show whether the melanoma outcome generalizes. In parallel, the first Phase 3 result from an AI discovery platform, along with Isomorphic's first dosing, will indicate whether the broader field is moving from proof of concept toward approved products. For anyone tracking AI in medicine, those readouts matter more than the current wave of stock moves.
David Borish is the author of The Tony Hawk Paradox. Read more of his work at davidborish.com.
