How AI can unlock life-saving cures, with Siddhartha Mukherjee

0:00If you've spent any time building AI products or leading technical teams, you know this. Transformation doesn't fail because of ideas. It fails because teams can't move together. Enter Atlassian's Teamwork Collection. It has planning in Jira, documentation in Confluence, video updates in Loom, and now AI agents in Rovo, which connects the dots across your work so nothing gets lost. It's one AI-powered teamwork platform designed for how modern teams actually build. Learn more at Atlassian.com slash TeamChanger. That's A-T-L-A-S-S-I-A-N dot com slash teamchanger Starting a business comes with its share of ups and downs, which is why staying true to your vision is essential.

0:53A non-negotiable for Romeo and Milka Bregali, Capital One business customers and co-owners of Ra's plant-based restaurant in New York. Romeo and Milka took a leap of faith when starting their own restaurant, gutting an empty space and building it from the ground up. every pipe, every wall, every detail. But building from scratch came with a heavy financial burden, which is when they turned to their Capital One business card. With the flexibility of the card's no-pre-set spending limit, they were able to spend more and earn more rewards while bringing their vision to life. Today, Raz's success is proof that with passion and the right support, it's possible to make your dreams a reality.

1:30Learn more at CapitalOne.com slash business cards.

1:42If you know me at all, you know that I love tracking my own health data. I am obsessed with my WHOOP, and I use it religiously to track my sleep, my movement, and now my biological age. And I truly believe that this is just the beginning. AI can help us unlock a world where we heal faster and live better, longer lives. My guest today, Siddhartha Mukherjee, is at the cutting edge of a very specific application of AI in health, drug discovery. His Pulitzer Prize-winning book, The Emperor of All Maladies, redefined our understanding of cancer. Sid has now co-founded Manus AI with Reid Hoffman to accelerate drug discovery through AI.

2:29In a world where traditional drug development is too inefficient, too expensive and slow, Manus is aiming for faster, cheaper, life-saving cures.

2:44I'm Rana El-Khalyubi, and this is Pioneers of AI, a podcast taking you behind the scenes of the AI revolution.

3:01Welcome to Pioneers of AI. I'm so excited for our conversation. My pleasure. So you are a true polymath. You're a cancer researcher. You're an author. You now have Manus AI. I am curious, take us back to the early days of your journey. How did you get interested in cancer and medicine? So when I was 18 years old, I took probably the most consequential and longest trip of my life, which is when I came to Stanford University as an undergraduate. Those days were very different days. Remember, no cell phones, no email from India. Grew up in Delhi. And there were four students from India, four undergraduates in total at Stanford at those years.

3:47I imagine the number has become at least 10 times, if not 20 times that today. And so I landed up at Stanford. my second year in became more and more interested in genetics in particular cancer genetics and I knocked and this was what you could do at a place like Stanford I imagine you can do it in other places too but I knocked on Paul Berg's door Paul Berg had won the Nobel Prize for inventing gene cloning or recombinant DNA and not only that this is very important Paul had then organized what's now known as the Asilomar Conferences, to consider placing a moratorium on recombinant DNA and genetic technology until the ethical, legal, and biosafety issues had been figured out.

4:38And that combination of someone who could span these two worlds, a world of what I would call the humanistic sciences and the world of pure science, the world of how science impacts humanity or human beings, either directly or at a population level, was very attractive to me. And so I knocked on Paul's door and I said, Paul, I'd like to come and work with you. And I was a second year undergraduate. Maybe Paul was in a very particularly good mood that day. He said, sure, come work with me. And so that began a 40-year relationship in which we had been friends and so forth. The point of this, of course, is that, and we would meet often, and every time I was in the Bay Area, he and I would have lunch together.

5:31But the important thing was that it was with Paul that I began to learn that science has a human aspect and everything that we do in human interaction and human lives can be taken back to science. And that was a very important formative moment for me. Yeah, that's amazing. And obviously, this is an AI podcast and it's still extremely relevant. Extremely relevant, this whole idea. And it's a recurrent theme in my life. You know, I take scientific work, try to distill it into readable forms, try to explain it to myself, to the world. And then I take it back from there into either the AI world, et cetera, and then try to learn something from the history and from what I wrote and from what I thought about while I was writing.

6:20I am curious about how you started writing. What got you curious about writing and also documenting, I guess, when you were putting the book together, you started documenting? Yeah, I mean, the book, which eventually, so I should say that the Emperor of All Maladies is coming, it's his 15th year anniversary, is coming out in a new edition with four additional chapters and a massive new edition because so much that's happened in the 15 years. In any case, the book began as a journal, and then I soon realized that it was much bigger than a journal, that there was a lot more to say and a lot more to think about.

6:59It became bigger and bigger and bigger, and at one point in time, I thought to myself, even as an oncologist, as a cancer doctor, as a cancer scientist, I didn't know so much about why or how we, researchers, scientists, doctors, patients, had arrived here now. And that's how this whole effort began. You know, I'm Egyptian. So I was struck that in your book, the earliest references to cancer were actually in ancient Egyptian texts. So we've been grappling with cancer for a long time. What have we learned from the evolution of this disease over the years? Well, we've learned a lot. I mean, you know, it would be hard to distill it in a few sentences.

7:39But the big picture is that, first of all, cancer is a genetic disease. It occurs because of changes in genes that either activate or inactivate cellular growth. And in doing so, the original cancer cell takes in nutrients, performs metabolism, alters the way it uses energy. But it's a cell where the genes have mutated. So that's the first thing. The second thing we've learned is that the mutations can come from many sources. They can come from random chance, which is when cells divide, they could make a mistake in copying. And so the daughter cell, the cell that's born, could have a mutation or a change in the genes.

8:31They can come from inherited changes. So they can come from your parents' DNA. they can come from viruses in particular some viruses that can either cause mutations or introduce new genes into cells and number four they can come from carcinogens from chemical or environmental insults that alter the genetic makeup of your cells that's the second thing we've learned and maybe the third thing that i'll say that we've learned that's sort of very globally important is that we now know that cancer doesn't occur in isolation, but rather it utilizes its environment, its micro environment, its tissue environment to essentially sustain itself.

9:21It borrows nutrients and signals, and that's why only certain cancers grow in only certain kinds of places. So of course, there are many, many, many, many details we've learned, but those would be sort of very high level, three sort of very big lessons in cancer. And I think what a lot of your writing has done, too, is that it's opened this kind of science and these findings so much wider audience and humanized it and made it accessible, right? Well, absolutely. I hope it has. The book is written for everyone. Cancer biologists read it. Young doctors read it. Patients read it. A very interesting group of people who read it are people who might be completely outside the field, but have an idea about it.

10:05And might come and write me a message saying, you know, has anyone thought about trying this, that, or the other? And that's why the field is constantly mobile. And that mobility is very important to me because it allows me to explore new spaces for discovering new drugs or making new medicines and changing the way we treat cancer. Before we dig into AI and Manus, I want to ask you about the state of cancer diagnosis that's making the news. we're seeing a lot of headlines about the rise of cancer in young people. What do you make of that? Is that just, you know, is it more cancer for real or is it just better diagnosis?

10:40Both. I recently wrote a long piece for The New Yorker about this. They, you know, they asked me specifically to tackle this kind of question. The problem with diagnosis is that, of course, that if you invent a new diagnostic test, the minute you invent the new diagnostic test, the number of diagnosis was going to go up. The most striking example is years ago in South Korea, they started giving out ultrasounds to primary care doctors and they would use that ultrasound to diagnose thyroid cancer. And so the incidence of thyroid cancer skyrocketed. The number of surgeries for thyroid cancer skyrocketed.

11:18Guess what difference there was to the deaths from thyroid cancer? I don't know. Zero. Zero? Zero. So that's an example where increased diagnosis does not mean increased mortality. It means increased diagnosis. That's what it means. So we are yet to find out. I mean, I would say that there are certain cancers where we know for sure there has been an increase in mortality, modest increase in mortality. That's a real statistic. For virtually all cancers, there's been a decrease in the United States of mortality. I can give you real numbers. So in 2000, the mortality from cancer, all cancers in the United States was 200 per 100 ,000 people.

12:04In the year 2024, 2025, it's going to be on the order of 140 per 100 ,000 odd people. So there's been a absolute reduction in cancer mortality. Now, that's true for most cancers, but it's not true for some. So we'll have to see whether these new diagnoses of cancer in younger people really translates into more increasing mortality. I'm afraid it might. And in which case, we'll have to really find out why. Our understanding of cancer has evolved, allowing us to fight the disease in innovative ways. But how can leveraging AI help? For Sid, diving into this question was a leap of faith. We'll find out why after a short break.

12:57Are you curious about the hidden side of everything? Then I have a podcast for you. I'm Stephen Dubner, host of Freakonomics Radio. Each week we hear from some of the most fascinating scholars and thinkers as we tackle big topics like how whales became the face of environmental activism, how to succeed at failing, and whether public transportation should be free. Go ahead, listen to Freakonomics Radio wherever you get your podcasts.

13:29Let's talk about Manus. So you and Reid recently founded Manus AI together, and you've said Manus AI is your leap of faith. Manus AI is a leap of faith because when I started thinking about AI, I thought that every aspect of drug discovery was inefficient, was not working, and could be made AI native and much, much more efficient. And by that, I mean, let's just walk through them. So you start with a target, usually. By target, I mean, it's usually a dysfunctional protein, could be a dysfunctional, some kind of molecule in a cell. So that's your target. From there on, you validate that target using experimental methods in laboratory, and then you try to use any mechanism.

14:15Either you try to jam it with a small molecule, like a lock jamming a key, or you try to prevent that target from being active using a variety of strategies. Manus AI is an attempt to take all of these steps, all the way to human clinical trials, and make them AI native. Right now, our focus is mainly on making molecular generators. So we have partnerships that do these other aspects of drug development. But right now, the centerpiece of NSAI is to build probably, we hope, one of the world's most powerful molecule generators. So a true generative foundation model for making new molecules that will become medicine.

15:02And then would the business model then become that you partner with other pharma and biotech companies and kind of license the molecule generator to them? Yeah, we license them, but we license them in partnership. So we don't, what we don't do, just to be very clear, we don't create, for instance, enterprise software for drug discovery. We don't, and, you know, we use LLMs a little bit, but really we make foundation models and then use those foundation models to make new molecules and then get those molecules out through partnerships with various companies until you reach, you know, an inflection point medically.

15:39And that's the purpose of Manus AI. I imagine that biotech and pharma have been using AI for years. So what's different now? So the kind of AI that biotech and pharma have been using has, I would say, very first generation AI. And in fact, it's not even clear that they have been using AI for true molecular generation. But there's a lot of machine learning baked into the discovery process already, right? There is some machine learning baked into some parts of the discovery process. So there is certainly machine learning baked into target discovery, what you want to send your target to. And you can imagine that's very amenable to machine learning.

16:20You can imagine that's something that LLMs would be very good at. You could search through a vast data space to find targets. And that has become relatively AI native. You can imagine that as a protein LLM. So those are now being incorporated into drug discovery, but they weren't before. Where there's not been a lot of AI is in the actual molecular generation. Like a generative chemistry. A generative chemistry, exactly. LLMs won't work here. You can borrow some important principles from LLMs and from what's been learned before. But these are, I'm going to use a fancy word, and I'm happy to explain it, but these are true neurosymbolic systems.

17:03A neurosymbolic system is a system that incorporates, on one hand, neural networks and AI, which of course is important because it's the learning aspect of it. But it's also symbolic in the sense that it is constrained by the rules of physics and chemistry. Just to give you an example, you could say that a sentence generator is largely dictated by tokenizing language and then putting those tokens together in a meaningful way. And it has a little bit of symbolic gestures like grammar, but it's largely governed by that tokenization. When now you enter the world of drug discovery, the symbolic aspects take on a much, much larger role, because physics is physics.

17:54Chemistry is chemistry. You can't violate those rules. Those rules are inviolable rules. Think of it this way. An LLM for images can make up a horse with five legs or a camel with wings. It can bend reality because, in this case, it's just generating a picture. But to make medicine, you need a model based on the immutable laws of physics and chemistry. You actually don't want it to make up something that cannot exist in the natural world. So off-the-shelf LLMs won't work here. And so you have to build a system which on one hand learns, but also is cognizant of the rules, has to keep note of the rules.

18:35And then not only that, it has to pay attention to those rules. Some rules are more important than other rules. To build a foundation model is much more complicated. What is the data you feed into these models if it's going to be generative? So there's a lot of data out there. It's mostly structural data. You know, since the 1960s, scientists have been depositing structures, crystal structures of proteins into a database called the PDB protein database it's a vast database it's a huge public service and so we use some of that we use laws of physics and chemistry we also use programs like alpha fold and finally we use molecular dynamics in other words how does the protein jiggle in space in real time and if that jiggling makes a difference to the drug or not Given that much of the data Manus AI is using comes from public datasets that any drug discovery has access to, what gives Manus AI an edge?

19:38More on that in a minute. Stay with us.

19:47Meet Nicole Nicholas, Capital One business customer and co-owner of Ansett Uncles, a plant-based restaurant and community space in Brooklyn, New York, that got its start from a need for unity. The inspiration, it was born from the desire to create a space that felt like home, where we can connect community culture, good food, and come together with family and friends. That's how we birthed aunts and uncles. Nicole and her husband, Mike, were fulfilling their dream of bringing people together out of their home kitchen. But they soon learned that the demand for community was greater than they knew.

20:18It became overwhelming and we were like, we need home, but not in our actual home. We realized that there was also a need in our community for something bigger in our neighborhood. So we had to find a place. Moving from a home operation into a storefront was a huge next step. But Nicole and Mike were able to take it on with the help of Capital One Business. It's not for the weak as a small business. Finding resources is super important because that's the way you'll be able to manage and scale. We would have never done that without having Capital One to be able to help us along the way. The cashback rewards are very helpful.

20:53So, you know, it just gave us that runway to be able to breathe a little bit. Then you get to focus on the cooking of the food and making the experience great. To learn more, go to CapitalOne.com slash business cards.

21:11With my investor hat on, I invest in early stage AI companies. I am wondering, what is your competitive moat? If all of your data sources, it sounds like are these public data sets, what makes Manus different? Well, I think two or three big differentiators, I would say. Number one is we are not an enterprise software company. So we actually want to make medicines and we want to make them end to end. So that's a big difference. That's a strategic difference. The second is that even though there's a lot of data out there, we are building what I would think is one of the world's most powerful foundation models.

21:52It all depends on the model. And also, these are very disparate data sources, it sounds like, how you synthesize the multimodal approach for that. So I think the way we harmonize the data, all these data sets, is very different. It's proprietary. we have a very, very proprietary approach to making generative chemistry. So we do both. We do search and fit, which is non-generative. So you can search molecular space and try to fit that. And I guess AI helps make that faster, more efficient. It can make it faster, efficient, but it's fundamentally not generative because you're searching and fitting.

22:31But one of the jewels in our crown is that we have a generative foundation model. So in which we basically look at a molecular pocket and try to actually make a chemical that fits. And that's completely new chemical matter. The third big difference is that we are multimodal in a way companies, most companies are not. We've already started making drugs that are moving towards clinic right now. Even as we speak, they are going through pre-IND, pre-clinical. They've been partnered. They're sort of moving forward. Very exciting space. So that's one. we have a sister company that we helped launch called Athea, which works on siRNA, so RNA as a drug.

23:13Which is the same kind of technology behind the COVID vaccines. For instance, yeah, but it's a different kind of RNA in this case. And we work across molecular disciplines. In fact, we have a sister company or a partner company which does cell and gene therapy. So again, that makes us very different because we are multimodal. And being multimodal, we can combine therapies. We can take an antibody and add to it a small molecule and make a combined therapy that's a combination of both. Last point is that we have a global infrastructure. I think it's important to note that I think the United States has basically priced itself out of many clinical trials.

23:55Just to give you one example, China in 2010, the United States in licensed about$5 billion of drugs from China. In 2025, that number is going to be about$60 billion. Wow. So you can imagine how much work is going on outside the US. So we have a global infrastructure in India spreading across now into Australia and other places. And that's another big differentiator for us. So let's say MANA succeeds, and you will. What will drug development look like in the next, I don't know, decade? And then also if the whole thing becomes AI native, what will that mean for drug development? The ultimate aim is cheaper, faster, better.

24:41So drug pharmaceutical companies claim correctly in some cases and incorrectly in some other cases that the reason that medicines, novel medicines, cost hundreds of thousands of dollars is because 90 % of them fail in early trials. If you could change that number from 90 % to even, let's say, 60%, you would dramatically change the economics of medicines. So that's the cheaper. The better is if you start with garbage in, you'll end with garbage out. In a lot of cases, there's this problem, and they're not better because they're not using the enormous power that we could have of machine learning.

25:20and faster because obviously machine learning enabled AI native system could take hours to write and survey the literature. Take an initial stab at the giraffe. Exactly. And again, I'm really focusing on things I don't think are sort of pipe dreams in some kind of distant vacuous future. I'm talking about things that are very proximately available and their companies that are already working on this right now. Very cool. Where is the human in the loop in all of this? So we've tried, we know we have human in the loop at every step. So just to give you one example, we recently encountered a medicine which was moving into quite advanced trials.

26:08Our computational chemists took one look at it and they said that'll never make a medicine because it's going to be so insoluble because it was a very aromatic planar structure. Now, you don't know, I don't know. Well, I know, but these flat aromatic planar structures, which look like Frisbees, are very insoluble and they tend to make very bad drugs. And so the trick is now to find out how to tweak that system. And now that knowledge needs to be funneled back so that if you look, if you do a search and next time around, don't make a structure that looks like that. One of the things I'm most excited about in the kind of the intersection of AI and health is this idea that we're at the cusp of a health span revolution that's more, it's personalized medicine, but it's also like really focused on wellness.