What's this about?
Moderna and Merck announced positive phase 3 results for mRNA-4157, a personalized cancer vaccine for melanoma, and Moderna's stock roughly doubled. The approach: take the patient's own tumour tissue, use AI to analyse it and select roughly 20 of the most representative antigens, encode those into an mRNA vaccine, and inject it to train the patient's T cells to recognize them — preventing recurrence after surgery. Melanoma was chosen first because it's among the solid tumours most responsive to immunotherapy; the reasoning is prove it there, then port the approach to other cancers. The obstacle is cost, since every dose is manufactured for one patient.
What exactly is an antigen, and what is the AI doing?
An antigen is a molecular marker the immune system can recognize and attack. A tumour carries mutations that produce antigens unique to that patient's cancer — which is what makes them a good target, and also why an off-the-shelf vaccine can't work. The hard problem is choosing: a tumour may present hundreds of candidate antigens, most useless because they're not distinctive enough or won't provoke a strong response. The AI's job is prediction — ranking which roughly twenty will actually train the immune system effectively. It's the same selection-under-expensive-evaluation problem today's Large Discovery Models paper describes, in a case where the evaluation is a patient.
Why is this worth noticing?
- Phase 3 is the real bar. Earlier-stage results in oncology frequently don't replicate; a positive phase 3 is the point where a mechanism becomes a medicine, and the market reaction reflects that.
- It's a concrete answer to this week's argument about whether AI will cure diseases. Several people, including a clinician, pointed out the enormous gap between designing a candidate and an approved therapy. This is a case that has crossed most of that gap — with AI in a specific, bounded role.
- The economics are the open question. Bespoke manufacture per patient is the opposite of how drugs achieve scale, and it's the constraint that will determine whether this generalizes.