An AI-designed cancer therapy passed the hardest test in medicine for the first time.

Moderna and Merck announced Wednesday that intismeran autogene, a personalized mRNA cancer vaccine, met its primary endpoints in a Phase 3 melanoma trial of more than 1,000 patients. Given alongside Merck's Keytruda, it produced statistically significant improvements in both recurrence-free survival and distant metastasis-free survival in people whose tumors had been surgically removed but who were at high risk of the cancer returning.

Moderna stock rose more than 126 percent, its best single session ever, and Merck rose over 9 percent. CEO Stéphane Bancel called it as big a day for humanity as November 16, 2020, when Moderna released its Covid vaccine results.

The reason this belongs in an AI newsletter rather than only a biotech one is how the vaccine gets made. Every dose is a different drug, and a model decides what goes in it.

Conventional cancer treatment is built on generalization. Chemotherapy exploits the fact that cancer cells replicate faster than healthy ones, killing both and hoping the patient outlasts the tumor. Immunotherapy like Keytruda broadly releases the brakes on the immune system.

Intismeran does the opposite. Doctors take a sample of the patient's own tumor and sequence it. Their cancer carries mutations that exist nowhere else, in that person or anyone else, and those mutations produce distinctive proteins on the tumor's surface called neoantigens. The vaccine is built to teach that specific patient's immune system to recognize those specific proteins.

The bottleneck is selection. A tumor may carry hundreds of mutations, and only some produce neoantigens the immune system can actually be trained to see. Choosing which ones to encode into the mRNA is a prediction problem across a large space, and it has to be solved fresh for every single patient, quickly enough that the person is still a candidate for treatment.

That is where AI does the work. Machine learning models rank candidate neoantigens by how likely each is to provoke a useful immune response, and the vaccine is manufactured from the top selections. Analysts covering the result singled out the AI-driven tailoring of neoantigens as the pivotal advance rather than the mRNA delivery itself, which has been well understood since Covid.

Without that computational step, personalized cancer vaccines are a research curiosity. With it, they are manufacturable at population scale.

The framing circulating online is running well past the evidence, so it is worth being precise about four things.

It does not cure existing cancer. The trial enrolled patients who had already had their melanoma surgically removed and were at high risk of recurrence. What the vaccine did was reduce the chance the cancer came back or spread. That is enormously valuable and it is a different claim.

It is melanoma only. The trial covered one cancer type, the deadliest form of skin cancer. Researchers hope the approach extends to other tumors, and that hope is reasonable given the mechanism, but it is untested.

It is not a standalone treatment. Intismeran was given in combination with Keytruda, an existing immunotherapy. The result validates the addition, not a replacement.

And most importantly, nobody outside the two companies has seen the data. Moderna and Merck issued a press release. They have not published in a peer-reviewed journal and have not released the underlying numbers, saying the results will be presented at an upcoming medical meeting and shared with regulators. Of the 24 brokerages covering Moderna, 22 still hold ratings of hold or worse, and at least one analyst explicitly urged caution pending full Phase 3 data.

None of that makes the announcement less significant. It makes the celebration premature by a few months.

If this holds up under peer review, the interesting part is not the melanoma result. It is that a workflow has been validated end to end.

Sequence a patient's tumor. Use a model to predict which of their unique mutations the immune system can be trained to attack. Manufacture a bespoke therapy from that prediction. Deliver it. Measure whether it worked in a controlled trial against a real clinical endpoint.

That loop is the thing every AI-in-medicine claim has been promising for a decade, and it has mostly produced papers rather than approvals. Reuters reporting notes doctors hope the approach works against many types of tumors, and the reason for that optimism is mechanistic rather than wishful, since the same neoantigen logic applies to any cancer that mutates, which is all of them.

Moderna says it aims to file for regulatory approval within months, without giving a specific timeline. The context makes that more fraught than usual: the US Department of Health and Human Services canceled a $590 million Moderna influenza contract last year and announced it would stop funding mRNA research through BARDA. A landmark oncology result arriving in that political environment is going to be a test of more than the science.

For anyone who works in or around AI, this is the most consequential validation of the year, and it did not come from a chatbot.

It also cuts against the framing that dominates most AI coverage, including this newsletter. We spend most weeks on benchmark deltas, pricing changes, and which lab shipped what. Meanwhile the highest-stakes application of machine learning ran a thousand-patient clinical trial and hit its endpoints, and it did so using models that predict protein behavior rather than generate text.

The honest position today is that AI-designed personalized cancer vaccines cleared Phase 3 for the first time, in one cancer, as an add-on therapy, with data nobody has independently reviewed yet. Every one of those qualifiers matters. And it is still the biggest thing that happened this week.

Watch for three things: the full data presentation at the medical meeting, whether the regulatory filing actually lands within months as Moderna claims, and whether any trial in a second tumor type reports out. That third one is what turns a melanoma result into a platform.

The stories moving fast beyond today's headlines:

The result. Moderna and Merck announced Wednesday that intismeran autogene, a personalized mRNA cancer vaccine, met its primary endpoints in a Phase 3 melanoma trial of over 1,000 patients. Added to Keytruda, it significantly improved recurrence-free and distant metastasis-free survival in patients whose tumors had been surgically removed but who remained at high risk. It is the first therapy of its kind to succeed at Phase 3. Moderna stock rose more than 126 percent, its best day ever, and Merck gained over 9 percent.

Where AI comes in. The vaccine is built from a sample of each patient's own tumor. Machine learning models rank which of that person's unique mutations, called neoantigens, the immune system can most likely be trained to attack, and the vaccine is manufactured from the top predictions. Every dose is a different drug. Analysts singled out the AI-driven neoantigen selection as the pivotal advance rather than the mRNA delivery, which has been well understood since Covid.

What it is not. It does not cure existing cancer, it reduces recurrence after surgery. It covers melanoma only. It works in combination with Keytruda rather than standalone. And the data has not been published or peer-reviewed, with results due at an upcoming medical meeting, which is why 22 of 24 brokerages covering Moderna still rate it hold or worse.

The through-line. If it holds up, the validated thing is a workflow rather than a drug: sequence a tumor, predict targets with a model, manufacture a bespoke therapy, prove it against a clinical endpoint. That loop applies to any cancer that mutates, which is all of them. Moderna says it will file for approval within months, in a political environment where HHS canceled a $590 million Moderna contract and stopped funding mRNA research through BARDA.

Stay building. 🤖