AI as a Compiler for Biology

Yesterday I watched Moderna gain roughly $44 billion in market value in a few hours.

The number was so absurd that I first assumed I had misunderstood it. Moderna had spent years living in the shadow of its COVID success, trying to prove that mRNA was not just a technology that happened to be useful during one extraordinary pandemic. Then Merck and Moderna announced that their individualized mRNA therapy for melanoma had met its Phase 3 endpoints, and the market suddenly behaved as if an entirely different future had become possible.

The obvious story is that a personalized cancer vaccine worked.

The stranger story is that every patient receives a different drug.

And AI helps decide what that drug should contain.

Moderna's intismeran autogene, previously called mRNA-4157 or V940, begins with the patient's own cancer. Tumor and blood samples are sequenced. The system looks for mutations that distinguish the tumor from the person's normal cells. Those mutations can create neoantigens, abnormal protein fragments that the immune system might learn to recognize as belonging to the cancer.

There can be far more possible mutations than could sensibly be put into one treatment. Moderna says a series of integrated AI algorithms reviews the sequencing data and predicts up to 34 neoantigens most likely to produce an immune response. Those selected targets are then encoded into a single synthetic mRNA molecule made for that patient.

So the pipeline looks something like this:

tumor + blood → sequencing → mutations → computational selection → up to 34 targets → mRNA design → manufacturing → injection → immune response

I keep thinking about this as a compiler.

A software compiler takes one representation of something and translates it into another representation that a machine can execute. I write source code in a language I can reason about. The compiler transforms it into instructions a processor can run.

Here the source is not JavaScript or C. It is a particular human body and a particular tumor.

The output is not machine code. It is a molecular instruction.

That metaphor has limits, and they matter. A tumor is not clean source code. Biology is noisy, incomplete and contextual. An AI model is not a doctor. The algorithm does not independently decide that a person should receive treatment, and the treatment is not simply the output of AI. Sequencing, immunology, clinical judgment, mRNA engineering, manufacturing, Keytruda, regulation and the patient's own immune system are all part of the result.

But the compiler metaphor still reveals something important.

The drug is becoming the output of a computational system.

For most of pharmaceutical history, the industrial model has been to discover one medicine, prove that it works, manufacture enormous quantities of the same thing and distribute identical doses to many people.

This system starts somewhere else.

Understand one person, compute one treatment, manufacture one person's medicine.

That is a very different product architecture.

The medicine is no longer only the molecule in the vial. The product is the entire pipeline that can reliably move from biopsy to sequencing, from sequencing to prediction, from prediction to molecular design, from design to manufacturing, and from manufacturing back to the right person before the treatment arrives too late.

This is why the AI part interests me more than the usual headline about AI discovering drugs.

The AI is not sitting beside a scientist generating promising molecules and then disappearing from the story. It is inside the production loop. Patient data enters the system. A patient-specific molecular design comes out. If the pipeline works, the result is manufactured and put into a human body.

Generative AI suddenly feels less like a category of media technology and more like a general industrial pattern.

We have become used to:

prompt → generated text

or

description → generated image

or

specification → generated software

Now we can start to see another pattern:

biological state → generated intervention

That changes what the word personalized means.

Most software personalization is actually selection. Netflix selects a film. Spotify selects a song. An online shop selects a product. Even many personalized health products mostly reorder advice that already exists.

This is different. The object itself changes.

Patient A and patient B can enter the same medical system and receive different molecular sequences because their cancers are different.

The product is a function.

Not in the mathematical sense that biology becomes perfectly deterministic, but in the product-design sense that what is delivered depends on the input. The factory is no longer only reproducing an object. It is executing a process capable of producing variation safely.

That is both beautiful and terrifying.

Mass production gave us extraordinary reliability because sameness is easier to control. Make ten million identical tablets and the quality problem is difficult but comprehensible. Make a medicine once for one person and suddenly identity, provenance, timing and data integrity become part of the medicine itself.

Which tumor sample produced this sequence?

Which algorithm version ranked the neoantigens?

What evidence supported those selections?

Was the right sequence manufactured?

Did the right batch reach the right person?

How do we audit a treatment that may never be manufactured in exactly the same form again?

A personalized medicine system needs some of the same things we ask from trustworthy software: provenance, versioning, observability, validation, security, human review and the ability to understand how an output came to exist.

Except software can usually be patched.

A molecular intervention enters a body.

This makes trust a design requirement rather than a nice interface layer placed on top afterward.

Privacy matters for the same reason. A tumor genome is not generic product analytics. The deeply personal data needed to make the system useful is also the data that makes misuse, leakage and exclusion more consequential. If personalized medicine depends on people allowing institutions to sequence, store and compute over intimate biological information, then privacy and governance are part of the medical infrastructure.

So is representation.

A system that learns which neoantigens are likely to produce immune responses needs evidence across different human genetic backgrounds and HLA types. If some populations are poorly represented in the data used to build and evaluate those systems, personalization can become a strange kind of false promise: technically individualized, but built on knowledge that represents some people much better than others.

Then comes the largest systems problem of all.

Who gets this?

A treatment that requires tumor sequencing, computational analysis, individualized manufacturing, specialized logistics and expensive immunotherapy can be personalized at the molecular level while remaining extremely unequal at the social level.

We could build medicine that understands one person's cancer with astonishing precision and still fail to build a health system capable of delivering it to most people.

That contradiction feels like the real next problem.

The scientific achievement is learning to produce uniqueness. The social achievement would be learning to distribute it.

We learned how to mass-produce identical medicine in the twentieth century. Perhaps one of the infrastructure problems of the twenty-first is learning how to mass-produce uniqueness.

That means making sequencing cheaper and geographically available. It means manufacturing systems that can handle thousands or millions of tiny patient-specific batches without losing quality. It means standards that allow hospitals, laboratories, software and factories to exchange information safely. It means reimbursement models that can understand a drug manufactured once. It means strong privacy protections. It means explaining the process to patients without turning uncertainty into marketing. It means building the whole chain so that a breakthrough does not stop at the hospitals and countries already richest in technology.

I added this as a new prototype questline on yah.quest: Make personalized medicine available to everyone.

Not because yah.quest is going to solve cancer, and definitely not because strangers on the internet should be designing treatments. The useful work around a frontier like this is much larger than the wet lab. We can map access bottlenecks, compare public standards, audit patient journeys, understand manufacturing capacity, study reimbursement, improve explanations, document data-governance gaps and make the infrastructure around the science more legible.

This is the part of AI I find most hopeful.

Not the machine replacing the human who understands the world.

The machine becoming one translation layer inside a much larger human system, taking something too complicated for us to handle manually at scale and helping turn it into something specific enough to act on.

A tumor becomes data. Data becomes a ranked set of targets. Targets become an mRNA sequence. The sequence becomes a medicine. The medicine enters a person. The immune system decides what happens next.

There are humans, institutions, factories, laws and uncertainties at every step.

AI is not the whole system.

But it may be becoming one of the compilers through which biology becomes actionable.

AI authorship note

This essay began with Kris Haamer noticing Moderna's extraordinary market move after the Phase 3 melanoma announcement and asking how AI was actually used in the research. Kris identified the central framing: AI as a compiler that helps translate individual biological data into a patient-specific molecular intervention, and then connected that technical shift to the systems problem of making personalized medicine broadly accessible. AI helped research, structure, edit and implement the essay. Kris supplied the questions, framing, systems argument, yah.quest connection and publication decision.