On 19 August 2026, Merck and Moderna announced that their personalised cancer vaccine, intismeran autogene, had met both efficacy endpoints in a Phase 3 trial. The trial, INTerpath-001, randomised 1,137 people with completely resected stage IIB to IV melanoma, two to one, to the vaccine plus Keytruda or Keytruda alone. The combination showed statistically significant improvements in recurrence-free survival and in distant metastasis-free survival.
That is the first positive Phase 3 result for an individualised neoantigen therapy, and the first for any mRNA-based cancer treatment. It is a big deal.
The same day, Elon Musk posted on X: “Despite its obvious misuse during Covid, mRNA has tremendous promise for curing diseases. Artificial RNA essentially makes curing diseases a software problem.” He has said versions of this since 2020, when he wrote that synthetic RNA and DNA “basically makes the solution to many diseases a software problem.”
I build software. So I want to take the analogy seriously, work out which parts of it are true, and be honest about which parts are not.
What the vaccine actually is
Start with what happens to a patient. Surgeons remove the tumour. The tumour is sequenced alongside the patient’s healthy tissue, and the differences are the mutations unique to that person’s cancer. An algorithm predicts which of those mutations produce fragments, called neoantigens, that the patient’s immune system is most likely to recognise. Up to 34 of them are chosen. Their sequences are written into a single mRNA molecule, wrapped in a lipid nanoparticle, and injected. The patient’s own cells read the mRNA, produce the fragments, and the immune system learns to attack anything carrying them.
Every patient gets a different molecule. That is what “individualised” means here. The factory is not making one drug. It is making one drug per person, from a specification generated by a pipeline.
Look at that pipeline through an engineer’s eyes and Musk’s point becomes obvious. Sequence in, prediction step, sequence out, compile to a molecule. The hard, expensive, decades-long part of traditional drug discovery, finding a small molecule that binds the right target and nothing else, is gone. Instead you have a general-purpose delivery platform and a payload that is, literally, a string. The chemistry is fixed. The string changes.
That is a software problem. Or at least the part that used to be the bottleneck now looks like one.
The evidence so far
The Phase 3 result did not come from nowhere. The earlier Phase 2b trial, KEYNOTE-942, followed 157 melanoma patients, and the five-year data presented at ASCO in June 2026 showed the combination reduced the risk of recurrence or death by 49 percent compared with Keytruda alone, with a hazard ratio of 0.51. The risk of distant metastasis or death went down by 59 percent. The FDA gave the therapy breakthrough designation in February 2023 and the EMA put it on its PRIME scheme two months later.
The UK is already building infrastructure around this. The NHS Cancer Vaccine Launch Pad went live in spring 2024 with a target of recruiting 10,000 patients into personalised cancer vaccine trials by 2030.
So the claim is not hype. Something real has crossed a line that many people, including many oncologists, did not expect to see crossed this decade.
Where the analogy breaks
Nine days after the Moderna announcement, BioNTech stopped a Phase 2 trial of its own individualised mRNA vaccine, autogene cevumeran, in resected colorectal cancer. The independent monitoring board found a numerical imbalance in overall survival between the arms and concluded that continuing was unlikely to change the outcome. No new safety signal. It just did not work in that setting.
The same company’s pancreatic cancer trial with the same platform continues, and its earlier Phase 1 data in pancreatic cancer, published in Nature in 2023, showed that patients who mounted a strong T-cell response had no recurrence at 18 months.
Hold those two results side by side. Same technology, same idea, same “software.” One cancer says yes, another says no, and a third is still being decided. If this were a software problem in the ordinary sense, you would not expect the answer to change that much when you change the input.
Here is how I would put it. The pipeline that designs the molecule is software. The thing that runs the molecule is the immune system, and the immune system is a runtime nobody wrote, nobody has the source for, and nobody can unit test. You can compile a perfect program and still have it fail because the machine it runs on behaves differently for every patient and every tumour type. The colorectal result is not a bug in the code. It is the runtime saying no.
There are other parts that are not software either. Manufacturing one molecule per person takes weeks, and a patient with aggressive disease may not have weeks. The neoantigen prediction step is a model, with all the false positives and false negatives that implies, and we do not yet know which of the 34 chosen fragments are doing the work. And the trials are still trials: years long, thousands of patients, endpoints measured in survival rather than in a test suite.
What I take from it
I think Musk is right about the direction and wrong about the destination. The part of medicine that mRNA turns into software is the part where you decide what to make. That part used to consume most of the time and money, and it is now a pipeline. What is left is the part where you find out whether it works, and that part is stubbornly, expensively biological.
For an engineer that is actually the familiar situation. We have compilers that never fail and deployments that fail all the time. The interesting work moved from writing the program to understanding the machine. Medicine seems to be arriving at the same place. The string is easy now. The runtime is the whole problem.