Moderna CEO Stéphane Bancel points out his company was using AI years before the ChatGPT hype to create new drugs, like its COVID-19 vaccine.
It’s also using the technology to review contracts and help answer questions from regulators. Two years ago, when what was a little-known biotech company topped $100 billion in market valuation based on its COVID vaccine success, it launched an AI academy to train its employees, including Bancel.
That’s boosting productivity at Moderna, which announced Wednesday that its experimental mRNA flu vaccine did well enough in a Phase 3 study that it could potentially advance through the U.S. regulatory approval process. And along with Merck, it’s planning a second Phase 3 trial of its mRNA-based therapy for a certain type of lung cancer.
“It’s a big transformation for Moderna, from what a lot of people perceive as a one-product vaccine company to a true platform,” Bancel told Semafor. Read below for our edited conversation on AI, management challenges the tech poses, and how artificial intelligence can help find a cure for cancer.
The View From Stéphane Bancel
Q: We have this intersection of AI, software, and biotech, which seems to be really heating up. Do these new techniques play into today’s news?
A: We have been doing machine learning for at least six years. So it’s a very pre-ChatGPT world for us. We believe, with biological knowledge exploding exponentially as a field, that complementing humans with machine learning was the right way to develop drugs over time. We developed the COVID-19 vaccine on the computer during a weekend. We never physically touched the virus; it was all information based. We are inventing new enzymes for how we make the mRNA in the factory. That did not come from humans; it came from machine learning. The new enzyme evolution bank. They tried it. It worked.
We convinced the FDA to change the manufacturing process, based on a new AI-invented enzyme. We’re using machine learning to invent new lipids to get mRNA in different cell types, opening big new verticals for Moderna. We’re using it as we interact with regulators, the Swiss, the French or whoever. They ask us a lot of questions about a new drug. They are slightly different but they are basically the same questions.
So we’ll use a large language model to basically prepare the answers for a regulator because the database is all there. And the system can read the questions from the regulator that we get in a document or an email or whatever, and write the answer that one of our experts will read and validate before we send it to the regulators. Sometimes you get 100 questions in one email on the manufacturing process, the safety. That’s just an example of how we want to really have it all around the company.
It’s going to change the way we all work, including myself. So what we’re trying to do is to help our teams, because you have people in IT that are really well-suited to data science and large language models. Mostly younger-generation scientists are all over this. But you have this layer of people who never used machine learning in their jobs. It’s a big management issue.
That’s why we started an AI Academy two years ago. We tried to get all the employees from a frontline operator or frontline scientist, up to the CEO, to go to these trainings, to try to get people to understand what data science is, what different algorithms exist for different tasks. And then to just play with it. Now we’re developing a lot of classes for prompt writing. It’s a new skill. It’s not like typing a Google search. It’s a very different way to think about how you interact with the machine.
We developed a ChatGPT for Moderna called M chat (because we don’t want to teach the rest of the planet the things we are learning with our data). We’re using it for pattern writing, contract writing. We are loading up all of our sensitive data.
My CIO was showing me a case where we received a document, we uploaded it into the system. And he showed me how to ask questions, and summaries and suggestions to M chat, versus reading the Word document myself. For contracts, in the old world, I would go ask my lawyer, ‘what is that clause for...’ Now I can do it myself. So think about the productivity gains and how everybody, including me, has to reinvent how to do their job.
Q: When you talk about those old timers who have never used machine learning, I wonder if they can teach the younger generation something as well. If they make good teams because they have a different kind of knowledge.
A: Exactly. During the dot-com era, I was in the young generation and had only been in the workplace a couple of years. I remember getting my Yahoo account. I’m aging myself. I remember coaching my boss about how to use the internet. He was trying to understand how to use it to solve business problems that, because I was junior at that time, I wasn’t thinking about.
We are seeing it happening now where people with more gray hair, and a lot of biology/science experience, are working collaboratively with the new kids that are coming out of grad school and college that are very versatile in this type of technology. Together, they can teach each other and become stronger together.
Q: You mentioned regulators. The FDA here laid out how they want to study, how they can change and adapt in this new world where we’re potentially going to see a lot more drugs come onto the market that are designed by AI. Do you believe that will happen and what do regulators need to do?
A: First, it’s kind of already happening. If you look at individualized new antigen therapy for cancer, we will select from a patient which mutation of DNA of a cancer cell we pick to code in the mRNA (that’s going to be the drug). It’s easily done by a machine learning system. It’s also learning patient by patient. What the regulator wants to see is that if you do the same input, you get the same output, because we need to control for safety and efficacy.
So in the case of a cancer vaccine, what we’re going to do is a bit like the OS of an iPhone. We’ll do version control, where, for example, everybody in the study has the same algorithm. We look at the data, we’ll apply machine learning. And then you go to 2.0 for the next data set of people, so that it’s controlled who gets what. And we can monitor efficacy and safety in a controlled way, by batches, if you will, of patients.
So I think they’re trying to think creatively, ‘How do we use the technology?’ Because fighting a technology revolution is always crazy. But what they want to do is evolve and be thoughtful, so as not to create things that will endanger people. It’s more of a journey. The good news is all the work happening across the industry around drug discovery, we still need Phase 1, Phase 2, Phase 3.
Q: You alluded to the individualized nature of your cancer products. Can you explain that further?
A: Let’s step back one minute to explain the basic biology of cancer. What we know today, which we didn’t know 20 years ago, is cancer is always a disease of DNA. And saying I have lung cancer or prostate cancer is interesting, because that’s one of the first cancers that happens. But what’s important to understand to treat cancer is what mutation of your DNA do you have. Like COVID, and flu, mutate.
Cancer is your DNA from what you got from mom and dad, in your first egg that was you, and how that DNA has evolved. Because of UV exposure with the sun, skin cancer happens, it creates mutation in your DNA. And when your DNA is mutated, it’s going to make proteins that are different from your healthy proteins.
The second piece we learned is that your immune system is built to eat your cancer cells. You and I, most probably, have had skin cancer many times in our lives. What we realize now is we all have cancer cells, thousands of them every day. What we’ve basically done through this technology is to weaponize the mRNA. So we basically do a little biopsy of your cancer cell. We will read all three gigabytes of blood cells of your cancer DNA. We do the same thing from a healthy set of you. And we send both to AWS and we compare letter by letter. Where the letters are different is a mutation.
Based on that information, we basically wrote this AI algorithm to pick which are the 34 mutations of your hundreds of cancer mutations that are more relevant to really get the biggest reaction from your immune system. That’s why there’s a lot of very complex AI in our system. We need to code in the mRNA that then is introduced as a vaccine. Then my immune system sees that and is able to recognize my cancer cell and go eat it.
Q: So that’s what you meant by targeted cancer?
A: Moderna, I think, is going to transform cancer care in a very profound way. It is the ability to teach your immune system what it missed. If you have cancer becoming one, then two cells, and you’re going to get a tumor, your immune system, because of stress, because you’re unhealthy at that time, it’s allowed to grow and your immune system becomes blind to it. You don’t see it, and you become really sick. We think within ourselves, we have the ability to cure cancer. We just need to re-educate your immune system how to do it. And that’s what we’re doing with mRNA.
Q: With these new advances in AI, it’s lowering the barriers to entry for this industry. That raises questions around how do you patent things when they’re created by AI? Or maybe you can create slight variations of them pretty easily.
A: When we have a patent issued by the U.S. Patent Office, we need to prove the molecule you want to patent does what you say it will do. So the idea that you have AI and try a lot of things and print tens of thousands of pages of different molecule alternatives and get a patent is naivete. You have to reduce it to practice. You need to go make it and that’s a physical thing, not a digital thing. And then you file your patent that molecule XYZ will do X and that’s going to be your claim for your patent.
So do I believe AI is going to make patents irrelevant? I don’t because of that last piece of the puzzle, which is you have to make the molecule and you have to give it to an animal.
Is it going to level the playing field for small companies versus big pharma like the Pfizers of this world? I agree because of the ability to get access to a lot of CPU power right away, without having a lot of experience yourself and so on. So is it going to take all the barriers out? No. Is it going to lower some barriers? Yes.