Moderna's AI still waits on biology's clock

Dr. Stephen Hoge told Blackstone that AI will make Moderna faster, but not fast enough to skip biology.

I think that is the right warning for an industry drunk on acceleration stories, and Hoge is well placed to give it. Hoge, president of Moderna, said messenger RNA is nature's software system for turning genes into proteins, yet it took the ten years after he joined the then two year old startup in 2012 to make that software manufacturable, scalable and safe enough to use in humans.

The timeline matters more than the metaphors.

By his account, Moderna had tested nine different vaccines in humans before COVID, according to Hoge, and each experiment trained the company's algorithms on which mRNA sequence would best produce a given protein. The spike protein alone could be encoded by more than 10 to the 80th different mRNAs, Hoge said, a space larger than the number of molecules in the known universe. The system had to pick one.

When the SARS-CoV-2 sequence arrived in early 2020, the algorithm picked one code, mRNA-1273, in about an hour, Hoge said. Moderna never made a second candidate for that vaccine, he said, because a decade of prior design work had already answered the selection problem.

That history shapes his view of AI now. Hoge said AI already compresses work that once took teams of scientists, from sequence design to trial enrollment planning, into work one person can do quickly. He called that a productivity gain that has to matter. The next frontier, he argued, is discovery, where AI suggests what drug to make rather than how to make it faster.

He is optimistic there, but he is measured and I agree with the nuance. Hoge said individualized medicine, especially around the immune system, is where AI that can ingest far more personal data than humans can will unlock new ideas. For large population medicines, he said, humans are already good at picking targets and AI may add less.

The concession is fair, and it cuts against the hype that AI alone will produce cures in months. Hoge said even if you compress discovery to weeks, you still have to prove the outcome in people. If the endpoint is whether melanoma returns in three years, you wait three years, he said.

The melanoma story is his proof point for both promises and limits. Hoge pointed to a phase 2 trial run with Merck's Keytruda, an anti-PD-1 immunotherapy that helped transform melanoma from what he called a death sentence. The idea, he said, was to teach the immune system what a patient's specific cancer looks like instead of just making it generally angry. Each cancer carries private mutations called neoantigens that never appear in healthy tissue, he said, and Moderna's approach from about 2015 to 2019 was to build the engineering to target them.

The phase 2 result announced earlier this summer showed a clinically significant improvement over immunotherapy alone, Hoge said, describing roughly a 50 percent additional benefit over what had been the gold standard. He said phase 3 results are still to come.

The logistics show why scaling here does not look like COVID. The phase 3 trial discussed involved about 1,100 to 1,200 people across 42 countries, Hoge said, with about 700 or 800 needing an individualized vaccine and the rest receiving placebo plus standard of care. The workflow is personal. Tumor tissue is sequenced to find the signature mutations, about 34 neoantigens per melanoma patient in this program according to Hoge, then the algorithm designs a bespoke mRNA and the company manufactures nine doses just for that person. The average turnaround was about 42 days, or about six weeks, Hoge said, timed to the start of cancer treatment.

Hoge said scaling for personalized cancer is not scaling up. It is scaling down and then scaling out. One operator can oversee four more machines once they are running, he said, and Moderna has already had to build that capacity to run trials that will eventually enroll 4,000 or 5,000 patients with individual cancers, according to Hoge. That level, he said, is already near commercial scale because thousands a year matters in this disease and tens of thousands would be very meaningful.

He was blunt about what will not scale at AI speed. Hoge said the FDA still requires making the medicine and running the trials before you can talk to patients about it, and physical limits cap how many trials you can run or how many patients with stage three melanoma exist in a given year. He said if capital for biotech R&D and manufacturing does not grow alongside AI investment, the system will hit those limits quickly. The company built much of its machine learning foundation on AWS during the pandemic years to support that mRNA design and manufacturing, a platform history detailed by AWS.

If you are a founder asking what this means, plan for AI to give you more shots on goal, not a shorter game.