In a drug made for one patient, the AI is the product

Investors are backing AI where it changes least about what a drug costs. In medicines made for one patient, design repeats with every sale, and AI could set the margin.

Investors are in talks to value Isomorphic Labs, Alphabet’s AI drug-discovery company, at $40 billion or more, Bloomberg reported this month. Isomorphic uses AI to find better drug candidates, and it has no clinical result yet. Aurora Therapeutics, the first startup built to make custom CRISPR treatments, raised a $16 million seed round in January. I believe more of that money should go to companies like Aurora.

In a drug made for one patient, the trials, the approval and the manufacturing process are shared across every patient on a platform. What has to be done again for each patient is the design: choosing the edit, predicting where else in the genome it might act, and running the lab work that checks both. KJ’s team needed months of experiments for that. AI suits this work. It is mostly prediction, working out from a patient’s DNA which edit will fix the mutation and what else it might hit, and it repeats with every patient. A model that predicts well replaces experiments for each new patient and gets better with every patient it learns from. This is the part of drug development where AI can make the difference. Here it sets the cost of each patient, and that cost decides whether a platform treats a handful of children or thousands. This year the FDA started writing the rules that make such a business possible.

“That’s not scalable”

Endpoints News opened its AI Day on 14 October with a panel on AI and custom genetic medicines. Ryan Cross, its senior science correspondent, asked Melissa Moore, Moderna’s former chief scientific officer, Le Cong of Stanford and John Finn of Basecamp Research how much faster AI could make the next treatment like Baby KJ’s.

KJ was born in August 2024 with CPS1 deficiency, a rare disorder that stops the liver from clearing ammonia. Doctors at the Children’s Hospital of Philadelphia (CHOP) and Penn Medicine built a gene editor that corrects his mutation. It took six months from diagnosis to infusion, and their paper in the New England Journal of Medicine describes no AI in the design. The team screened candidate editors in liver cells, then tested the best one in mice and monkeys. KJ received three doses between February and April 2025, and a year later he was walking and talking. NIH grants, a CHOP research programme and donated materials paid for the work.

The panel’s answer was that design had not been the slow part. Finn said he could “print out some mRNA” and “make it in a week”; the testing and the regulatory work took the time. Moore said the figure she had heard for KJ’s treatment was $3 million, and “that’s not scalable.” After KJ’s story broke, CHOP received thousands of requests from other families.

Better candidates, same bill

In the standard process, a company designs a drug once and then spends a decade proving it works. AI has gone into the design, and it helps there. Boston Consulting Group analysts (Jayatunga and colleagues, Drug Discovery Today, 2024) found that AI-discovered molecules pass phase 1 safety trials 80% to 90% of the time, well above the industry’s record. In phase 2, where a drug first has to work in patients, they succeed about 40% of the time, no better than other drugs. None has been approved.

Deloitte puts the average cost of bringing a drug to market at $2.67 billion for the 20 largest drugmakers, counting failures. Most of that is spent after design, on trials and on the programmes that fail them. Better candidates can trim it, but they do not remove the expensive part, which is proving in patients that a drug works. As I argued in “We now design more drugs than we can test”, the new candidates wait in the same queue. Investors pricing Isomorphic at $40 billion are paying for the step AI has improved, in a process where most of the money goes elsewhere.

Where design is the cost

In a drug made for one patient, design comes back with every sale. Moderna and Merck’s personalised cancer vaccine, intismeran, already works this way. Software reads the DNA of each patient’s tumour, picks up to 34 mutations unique to it, and encodes them in that patient’s mRNA vaccine; Merck calls the result “algorithmically derived.” In August the companies reported that the vaccine met its main goal in a phase 3 trial of 1,137 melanoma patients. Every patient got a different drug, and the trial tested the process that made them. How much of the result came from the software is not yet known, but the process passed phase 3.

Gene editing for rare disease is the next and bigger case. KJ’s team needed months of lab work for one child. A company treating a thousand children a year has to repeat that work a thousand times. For each child it has to design the guide, the short genetic sequence that steers the editor to the child’s mutation, predict where else in the genome the editor might act, and show afterwards that the edit worked, which the FDA’s draft rules require. It has to do this in weeks, without new animal studies for each child. Moore said on the panel that personalised medicine cannot afford to make 10 or 15 versions and test them, so “we need AI design to do single-shot design where we know that the mRNA that we design is going to be effective.”

Three kinds of reuse could make this cheap, and only one of them needs AI. The FDA’s framework lets one approval cover many patients. A fixed delivery particle, editor and manufacturing process let one production line serve them all, and ARPA-H is paying for equipment to run such lines at or near hospitals. Both will happen without AI, and both will be open to every company that enters. Neither makes a new guide safe or effective, and neither removes the testing each child’s medicine needs.

The third kind of reuse is the one a company can own, and it is why, in this business, the AI is the product. Every patient treated produces data on which designs worked. A company whose software learns from that data needs fewer experiments for the next patient, so each patient costs it less than the last, and a competitor working under the same rules with the same production line starts behind. That is the bet. It needs the right to learn from each treatment and a model that works on mutations it has never seen, which nobody has shown yet. How fast the cost per patient falls is what an investor is buying.

Why now

Until this year, nobody could sell a drug like KJ’s. He was treated under a one-off exception that bars anyone from selling the treatment. A standard approval would have cost more than US$25 million and taken four years, according to Fyodor Urnov of UC Berkeley and Sadik Kassim of Danaher, writing in Nature in April. Kiran Musunuru, who designed KJ’s editor, told MIT Technology Review he saw no scenario in which a for-profit company would find the rarest mutations commercially viable.

In February the FDA published draft guidance for what it calls a “plausible mechanism” framework. One approval could cover several mutations in the same gene, as long as each is corrected the same way, and a company could add new mutations later. CHOP plans a trial built on it. The delivery particle, the editing enzyme and the manufacturing stay the same for every child, and only the guide changes. Results in five to ten patients across seven related liver disorders could be enough to approve the whole platform, and the FDA agreed that liver cells engineered in the lab to carry a patient’s mutation can replace new animal studies.

Under the framework, Urnov and Kassim estimate, each new patient could be treated in about three months for under $250,000. Insurers already pay for drugs made one patient at a time, such as CAR-T cancer therapies built from each patient’s own cells, and one-time gene therapies sell at list prices of $2.2 million for Vertex’s Casgevy and $4.25 million for Orchard’s Lenmeldy.

Manufacturing is moving too. In September ARPA-H, a federal research agency, committed up to $125 million to equipment that would make RNA medicines at or near the hospital, which Moore said should take four days from sequence to medicine. On 14 October the FDA said it would scale manufacturing reviews to the risk for drugs meant for small groups of patients, and it holds a public meeting on the plan on 28 October.

The first platforms will use lipid nanoparticles, which deliver mainly to the liver. That reaches well beyond liver disease, because the liver makes many of the proteins the blood carries. CHOP’s ARPA-H programme covers clotting and bleeding disorders as well as metabolic ones.

What would prove this wrong

AI may not cut the lab work. Nobody has yet shown that software can design gene edits reliably enough to skip experiments. If it cannot, every child still needs a project the size of KJ’s, and the platform stays a hospital service.

The numbers may not close. If the first approval and the testing for each patient cost too much, cheaper later patients will not pay it back.

The rules may narrow. The guidance is still a draft, and Marty Makary, the FDA commissioner who backed it, resigned in May. A final rule narrower than the draft would turn each new mutation back into its own US$25 million application.

Something simpler may win. Finn made this case on the panel. For a gene with a thousand disease-causing mutations, “we can either make a thousand base editors, or put a healthy copy of a gene in the right place.” If inserting a whole healthy gene works safely, one product covers every mutation in that gene, and nobody needs a new design per patient. Rare disease would still be a large business, but the AI thesis would be wrong.

The opportunity

About 30 million Americans live with one of more than 10,000 rare diseases, by the NIH’s count, and only about 500 of those diseases have a treatment. Many are not caused by a single mutation an editor could correct. Even so, treating 1% of those patients at the $250,000 cost estimate would mean $75 billion of spending.

Charities and governments pay for most of the one-patient work today. The n-Lorem Foundation has treated 50 patients with custom drugs, free of charge, on donations. The Chan Zuckerberg Initiative gave $20 million to a new Center for Pediatric CRISPR Cures, and ARPA-H gave CHOP up to $38.9 million. Donations will not reach 30 million people, and venture capital has so far put in seed money.

The bet has the shape venture capital is built for. The first approval is a single milestone that either happens or does not, and nobody pays for anything before it. After it, each patient costs a fraction of the first, and the AI decides how small that fraction gets. The numbers to underwrite are the cost of the first approval and how fast the cost of each later patient falls. Founders should build the design software, the lab methods that validate it and the regulatory evidence as one asset, secure the right to learn from every treatment, and raise enough to reach the first approval.

The panel asked how much faster AI could make the next treatment like KJ’s. For an investor, cost matters more than speed. If the thousandth patient’s treatment costs a fraction of the first, because the software needs fewer experiments each time, the platform is a business. If it costs about the same, it stays a hospital service.

Jan Buza

Partner @ Zaka